**Application of Integrated Optical Electric-Field Sensor on the Measurements of Transient Voltages in AC High-Voltage Power Grids**

**Shijun Xie 1,\*, Yu Zhang 1, Huaiyuan Yang 2, Hao Yu 2, Zhou Mu 1, Chenmeng Zhang 1, Shupin Cao 1, Xiaoqing Chang <sup>1</sup> and Ruorong Hua <sup>1</sup>**


Received: 28 March 2019; Accepted: 7 May 2019; Published: 13 May 2019

**Featured Application: Comparing with traditional measuring techniques based on electrical engineering, the novel measuring techniques for transient voltage in AC high-voltage power grid based on optical electric-field sensor is of fast response speed, wide bandwidth, small size, light weight and safety.**

**Abstract:** Transient voltages in the power grid are the key for the fault analysis of a power grid, optimized insulation design, and the standardization of the high-voltage testing method. The traditional measuring equipment, based on electrical engineering, normally has a limited bandwidth and response speed, which are also featured by a huge size and heavy weight. In this paper, an integrated optical electric-field sensor based on the Pockels effect was developed and applied to measure the transient voltages on the high-voltage conductors in a non-contact measuring mode. The measuring system has a response speed faster than 6 ns and a wide bandwidth ranging from 5 Hz to 100 MHz. Moreover, the sensors have the dimensions of 18 mm by 18 mm by 48 mm and a light weight of dozens of grams. The measuring systems were employed to monitor the lightning transient voltages on a 220 kV overhead transmission line. The switching transient voltages were also measured by the measuring system during the commissioning of the 500 kV middle Tibet power grid. In 2017, 307 lightning transient voltages caused by induction stroke were recorded. The characteristics of these voltage waveforms are different from the standard lightning impulse voltage proposed by IEC standards. Three types of typical switching transient voltage in 500 kV AC power grid were measured, and the peak values of these overvoltages can reach 1.73 times rated voltage.

**Keywords:** optical; electric-field; sensor; measurement; transient voltage; AC power grid; Pockels effect

#### **1. Introduction**

The wide existence of the transient voltage in the power grid threatens the insulation system of high-voltage plants, which often causes the grid faults [1]. The accurate measurement of transient voltage is the key for the fault analysis, the optimized insulation design and the standardization of high-voltage testing techniques.

Various types of transient voltages can be found in a power grid. The fastest transient voltages (i.e., very fast front overvoltage, VFFO) usually appear at a gas-insulated switchgear (GIS), which is caused by the operation of the GIS disconnectors [2–4]. A VFFO commonly has high frequency components of approximately 100 MHz and a rising time of several nanoseconds [5]. For the fast-front transient voltages, the rising time is normally in the range of 0.1 μs to 20 μs [5]. In terms of a transient voltage caused by normal operation, the rising time is in the range of microseconds. It is noted that these transient voltages are usually superimposed on the power-frequency voltage, which requires a wide bandwidth and fast response speed for the voltage measuring system to be well measured. Moreover, portability and safety are also advised for developing a voltage measuring system.

Subject to the requirements of insulation and capacity, traditional electrotechnical voltage measuring techniques usually bring huge sized and heavy weighted equipment, especially for those applied in an extra-high voltage (EHV) or ultra-high voltage (UHV) power grids. It is also noted that their bandwidth and response speed are unfavorably limited. For example, the capacitance-voltage transformer (CVT) which is a commonly used voltage measuring device in the high-voltage AC power grid typically has a weight of hundreds of kilograms (for 500 kV) and a bandwidth lower than 1 kHz [6–8]. Besides, the bushing tap is applied mainly to measure the transient voltage during a substation commissioning, of which the performance is mainly affected by the manufacturing process and the bushing structure. As a consequence, the upper bandwidth is generally not higher than 5 MHz for a bushing tap technique [9–11]. In addition, the portability cannot be achieved on these techniques.

Thus, subject to the future demand of the power system, it is highly expected to develop a safe and easy-to-apply voltage measuring technique. For a quasi-static problem, since the electric field is considered to be irrotational, the space electric field around a conductor is linear to the potential of the conductor. Based on the measuring technique for the electric field, a non-contacted measuring method for transient voltage could be possible. On the other hand, as the development of optical sensing technology, several electro-optical effects have been used for measuring electric fields. The electro-chromatic effect has a response time of ~100 s, and it is suitable only for extremely low-frequency electric field measurement [12,13]. The Kerr effect is a quadratic electro-optic effect and normally has a small electro-optic coefficient. Thus, the sensors based on the Kerr effect have a low sensitivity [14,15]. The Pockels effect is known as a linear electro-optic effect, which means the output voltage is linear to the measured electric field. The sensor based on the Pockels effect normally has a fast response speed, small size, and are passive. It is particularly used for the measurement of high-amplitude transient electric field [16].

In this paper, based on the linear correspondence between the transient voltage and the transient electric-field caused, a non-contact transient voltage measuring technique was proposed. An integrated optical electric-field sensor based on the Pockels effect was introduced to implement the field measurement. The measuring system was developed and tested for its bandwidth and response speed. A series of measured lightning transient voltages and typical switching transient voltages were presented. Finally, the characteristics of these measured transient voltages and the decoupling issues were discussed.

#### **2. Measuring Method and Measuring System**

#### *2.1. Method*

For a quasi-static problem, the voltage *U*(*t*) of the conductor is linear to the electric field *E*(*t*) caused by it, which can be expressed as:

$$dU(t) = kE(t). \tag{1}$$

Normally, the voltage waveform of a transient process includes the power-frequency component and the transient component. Since the amplitude of the power-frequency voltage is known for a power grid, by comparing the amplitudes of the power-frequency voltage and the measured power-frequency electric field, the coefficient *k* in (1) can be given.

In an actual AC power system, the transmission line contains three phase conductors, which unavoidably leads to the coupling effect for the non-contact measuring technique as illustrated in Figure 1. Beneath each phase conductor, an electric-field sensor is installed denoted as Sensor A, Sensor B, and Sensor C for the corresponding phase.

**Figure 1.** Schematic diagram of the measuring principle. *U*a, *U*b, *U*c—potentials on phase conductors; *P*SA, *P*SB, *P*SC—positions where electric field sensors are set; *E*aA, *E*bA, *E*cA—electric field produced by *U*a, *U*b, and *U*c, respectively at position *P*SA.

Sensors are arranged at the height level with enough air insulation clearance against corresponding conductors. As conductors are energized, the potentials of three-phase conductors produce electric fields at sensor's locations denoted by *P*SA, *P*SB, and *P*SC in Figure 1. Taking *P*SA as an example, the electric field at *P*SA is a compound electric field contributed by *E*aA, *E*bA, and *E*cA which can be decomposed into the components in the x-axis (*E*aA-x, *E*bA-x, *E*cA-x) and z-axis (*E*aA-z, *E*bA-z, *E*cA-z) respectively. The total electric field at *P*SA (*E*A) can be given by:

$$\begin{aligned} \mathbf{E}\_{A}(t) &= \mathbf{E}\_{A-\mathbf{x}}(t) + \mathbf{E}\_{A-\mathbf{z}}(t)i \\ &= \begin{bmatrix} k\_{\mathrm{d}A-\mathbf{x}} & k\_{\mathrm{b}A-\mathbf{x}} & k\_{\mathrm{c}A-\mathbf{x}} \end{bmatrix} \begin{bmatrix} \mathcal{U}\_{\mathrm{d}}(t) \\ \mathcal{U}\_{\mathrm{b}}(t) \\ \mathcal{U}\_{\mathrm{c}}(t) \end{bmatrix} + \begin{bmatrix} k\_{\mathrm{d}A-\mathbf{z}} & k\_{\mathrm{b}A-\mathbf{z}} & k\_{\mathrm{c}A-\mathbf{z}} \end{bmatrix} \begin{bmatrix} \mathcal{U}\_{\mathrm{d}}(t) \\ \mathcal{U}\_{\mathrm{b}}(t) \\ \mathcal{U}\_{\mathrm{c}}(t) \end{bmatrix} \end{aligned} \tag{2}$$

where *k* is a coefficient equalling to the ratio of electric field detected to the origin potential. The first letter in the subscript of *k* represents the phase conductor producing the aimed electric field, and the second letter represents the phase under investigation. To integrate three phases, the electric field matrix can be written as follows:

$$\begin{bmatrix} \mathbf{E}\_{A} \\ \mathbf{E}\_{B} \\ \mathbf{E}\_{C} \end{bmatrix} = \begin{bmatrix} E\_{A-\mathbf{x}} \\ E\_{B-\mathbf{x}} \\ E\_{C-\mathbf{x}} \end{bmatrix} + \begin{bmatrix} E\_{A-\mathbf{z}}(t) \\ E\_{B-\mathbf{z}}(t) \\ E\_{C-\mathbf{z}}(t) \end{bmatrix} \mathbf{i} = \begin{bmatrix} k\_{\mathrm{d}A-\mathbf{x}} & k\_{\mathrm{b}A-\mathbf{x}} & k\_{\mathrm{c}A-\mathbf{x}} \\ k\_{\mathrm{d}B-\mathbf{x}} & k\_{\mathrm{b}B-\mathbf{x}} & k\_{\mathrm{c}B-\mathbf{x}} \\ k\_{\mathrm{d}C-\mathbf{x}} & k\_{\mathrm{b}C-\mathbf{x}} & k\_{\mathrm{c}C-\mathbf{x}} \end{bmatrix} \begin{bmatrix} \mathbf{L}\_{\mathrm{d}}(t) \\ \mathbf{L}\_{\mathrm{d}}(t) \\ k\_{\mathrm{d}\mathbf{C}} - z \end{bmatrix} + \begin{bmatrix} k\_{\mathrm{d}A-\mathbf{z}} & k\_{\mathrm{b}A-\mathbf{z}} & k\_{\mathrm{c}A-\mathbf{z}} \\ k\_{\mathrm{d}B-\mathbf{z}} & k\_{\mathrm{b}B-\mathbf{z}} & k\_{\mathrm{c}C-\mathbf{z}} \\ k\_{\mathrm{d}C-\mathbf{z}} & k\_{\mathrm{b}C-\mathbf{z}} & k\_{\mathrm{c}C-\mathbf{z}} \end{bmatrix} \begin{bmatrix} \mathbf{L}\_{\mathrm{d}}(t) \\ \mathbf{L}\_{\mathrm{b}}(t) \\ \mathbf{L}\_{\mathrm{c}}(t) \end{bmatrix} \text{i.} \tag{3}$$

Equation (3) can be divided into

⎡ ⎢⎢⎢⎢⎢⎢⎢⎢⎣

$$
\begin{bmatrix} E\_{A-\chi}(t) \\ E\_{B-\chi}(t) \\ E\_{C-\chi}(t) \end{bmatrix} = \begin{bmatrix} k\_{aA-\chi} & k\_{bA-\chi} & k\_{cA-\chi} \\ k\_{aB-\chi} & k\_{bB-\chi} & k\_{cB-\chi} \\ k\_{aC-\chi} & k\_{bC-\chi} & k\_{cC-\chi} \end{bmatrix} \begin{bmatrix} \mathcal{U}\_{d}(t) \\ \mathcal{U}\_{b}(t) \\ \mathcal{U}\_{c}(t) \end{bmatrix} \tag{4}$$

and

$$
\begin{bmatrix} E\_{A-z}(t) \\ E\_{B-z}(t) \\ E\_{C-z}(t) \end{bmatrix} = \begin{bmatrix} k\_{aA-z} & k\_{bA-z} & k\_{cA-z} \\ k\_{aB-z} & k\_{bB-z} & k\_{cB-z} \\ k\_{aC-z} & k\_{bC-z} & k\_{cC-z} \end{bmatrix} \begin{bmatrix} \mathcal{U}\_{a}(t) \\ \mathcal{U}\_{b}(t) \\ \mathcal{U}\_{c}(t) \end{bmatrix} \tag{5}
$$

Based on either (4) or (5), the voltage waveforms on a conductor can be given by measuring the electric fields in the case that the *k* matrix is known.

#### *2.2. Sensor*

An integrated optical electric-field sensor is developed based on the Pockels effect [17–19]. In this optical sensor, an optical waveguide is fabricated on a LiNbO3 substrate by titanium diffusion. The refractive index of the optical waveguide would be changed under an external electric field (*E*). Thus, while a polarized light passes through the optical waveguide, there would be a phase difference between its vertical and horizontal components. The modulated polarized light was transmitted to a common path interferometer via a polarization maintaining optical fiber, while the phase difference caused by external electric-field is demodulated into light intensity. Eventually, the light intensity is transferred into an electrical signal (*U*out) by a laser receiver. The relationship between *E* and *U*out is given as [20,21]:

$$\mathcal{U}\_{\rm out} = A \cdot \left[ 1 + b \cdot \cos \left( \phi\_0 + \frac{E}{E\_\pi} \pi \right) \right] \tag{6}$$

where *A* represents the photoelectric conversion coefficient and the transmission loss. *b* denotes the extinction ratio. φ<sup>0</sup> can be controlled near π/2 through the production of the sensor. In the case that the external electric field is much smaller than the half-wave electric field (*E*π), by Taylor's expansion, (6) can be rewritten as:

$$
\Delta L\_{out} = A \cdot \left[1 + b \cdot \frac{E}{E\_{\pi}} \pi\right]. \tag{7}
$$

Thus, the output voltage of the optical electric-field measuring system is linear to the external electric field.

Figure 2 shows the schematic diagram of the developed optical sensor which has the dimensions of 18 mm by 18 mm by 48 mm and a weight of dozens of grams. The sensors, as well as the input and output fibres, are made of dielectric materials which could guarantee the safety under high-voltage conditions. Moreover, the sensor indicates a good directivity as it is sensitive to the electric-field component perpendicular to the front of the sensor. On the field measurement, the sensor is usually facing the measured conductor. Thus, instead of (4), (5) is normally used to compute the transient voltage.

**Figure 2.** Integrated optical electric-field sensor. (**a**) schematic diagram; (**b**) product photo.

#### *2.3. Measuring System*

The measuring system contains an optical electric-field measuring unit and a data processing unit. The optical electric-field measuring unit includes optical sensors, a laser source, a laser receiver, and transmission fibers. The output analog signals of the laser receiver are converted into digital signals by a high-speed data acquisition card (HS-DAQ) with 4 acquisition channels and a sampling ratio of 40 MS/s. The digital signals are temporarily stored in the internal storage and further analyzed by the central processor to determine whether a transient process occurs. While the system is triggered by the transient voltage, the transient voltage waveform from 20 ms before to 80 ms after the triggering will be stored into the local storage hard disk. Simultaneously, 400 milliseconds voltage waveform after the triggering time would also be recorded with a sampling ratio of 20 kS/s. All these signals are timed by

a GPS module with the error of less than 100 ns. The field measured signal will be transmitted to the central station by a 4G wireless communication module. Thus, the HS-DAQ, the central processor, the internal storage, the hard disk, the GPS module, and the communication module constitute the data processing system. Except for the sensors and optical fibers, all the other devices are integrated into a cabinet or a portable suitcase and powered by an isolated power supply. The general connection diagram of the measuring system is shown in Figure 3.

**Figure 3.** General connection diagram of the measuring system.

Due to the nature of randomness for the lightning strike, the sensors used for measuring lightning transient voltages are installed outdoor for a long term. To protect them from environmental effects, sensor containers are used, which are supported by a grounded metal tube to the same level of the base of adjacent insulators, as shown in Figure 4a. For the measurement of switching transient voltage, the sensors are just supported by an insulating rod temporarily, as shown in Figure 4b. On both arrangements, the clearances between the sensor and the high-voltage conductor are no shorter than the necessary distance specified in [5].

**Figure 4.** Field arrangement of sensors. (**a**) Field arrangement of sensors for monitoring the lightning transient voltage; (**b**) Field arrangement of sensors for measuring the switching transient voltage.

#### *2.4. Performance Test*

To ensure that the developed measuring system meets the performance requirements for the transient voltage measurement, a series of tests including the response speed and frequency response have been carried out in the laboratory.

A 5 kV 1.2/30 μs lighting impulse was imposed on a circular-shape plane-plane gap with the gap distance of 10 cm and the radius of 50 cm. A 996:1 standard impulse voltage divider was employed to offer the standard measuring result. The normalized measuring results of the developed measuring system and the standard divider are shown in Figure 5a. To determine the response time of the measuring system which is in the range of nanoseconds, a bounded wave electromagnetic pulse simulator, composed of charging transformer, impulse capacitor, sharping switch, parallel-plate transmission line, and matching resistor was employed, to generate a pulse with 6 ns for the rise time and 38 ns for the half time. A 104.2:1 standard resistive divider was used to offer standard measuring results. The normalized measuring results are shown in Figure 5b, which indicates that the response time of the measuring system is less than 6 ns.

**Figure 5.** Test results of response speed. (**a**) Lightning impulse; (**b**) electro-magnetic pulse.

To conduct the frequency response test, a signal generator was employed to generate a standard sinusoidal voltage from 5 Hz to 250 kHz. The frequency response of a higher frequency range is determined by a sharp voltage impulse which was imposed on a transverse electric and magnetic field (TEM) cell. The amplitude-frequency response is shown in Figure 6. Except for individual resonant frequencies, the curve of the amplitude-frequency response is flat from 5 Hz to 100 MHz.

**Figure 6.** Amplitude-frequency response.

Based on the characteristics of transient voltages defined in [5], performances of the measuring system developed in this paper have satisfied the requirements for measuring field transient voltages in AC high-voltage power grid.

#### **3. Results**

#### *3.1. Lightning Transient Voltages*

#### 3.1.1. Field Test Overview

Considering the randomness of lightning strikes, the measuring system was installed for a certain duration in a 220 kV Yan-Zhuang overhead transmission line connecting the Yanshenshang substation (Y-substation) and the Zhuangshang substation (Z-substation) with 78 towers and the length of 39.8 km. The area of the transmission lines has an average ground flash density about 1.012 times/km2 in 2015 and 2.155 times/km2 in 2016. Measuring systems were arranged on the entrances of Y-substation and Z-substation respectively.

The causes for the transient voltage in a substation are various, including the operations, lightning strikes, electro-magnetic interferences, etc. To distinguish the actual lightning transient voltage, the measuring systems are synchronized with the lightning location system of China (LLS). The LLS includes over 940 lightning detection terminals covering most areas of China. Its average detection efficiency is higher than 90% and the location errors are smaller than 500 m. The errors of GPS clock used in the measuring system and LLS are both within one hundred nanoseconds.

A ground flash event usually contains the first stroke and several subsequent strokes. The period between two continuous strokes is usually longer than 30 ms [22]. Since the length of Yan-Zhuang line is 39.8 km, the difference of the triggering time of the measuring systems and the LLS will not be longer than 150 μs. Thus, the recorded data is considered to be a lightning transient voltage waveform caused by an involved ground flash, while the measuring system is triggered and the LLS has recorded a ground flash near the Yan-Zhuang line in 1 ms. Moreover, both the event that the measuring equipment is triggered and a ground flash strike to the point near the monitored transmission line are of small probabilities. When two small-probability events happen almost at the same time, it is believed that these two observed processes are the same event.

#### 3.1.2. Typical Measurement Results

In 2017, a total of 832 pieces of data were recorded by the two measuring systems. 650 pieces of data were recorded by the system in Z-substation and 184 pieces of data had been recorded by the system in the Y-substation. 307 pieces of data can be synchronized with a ground flash recorded by the LLS.

Once a lightning stroke was recorded by two measuring systems, the double-end traveling wave location method (DTLM) is activated to determine the position along the Yan-Zhuang line and nearest to the lightning stroke point.

On 13:58:27, May 14th, 2017, the two measuring systems were triggered in succession and the original recorded voltage waveforms from the sensors of phase C are shown in Figure 7. Obvious voltage fluctuations can be seen in the dotted box. After filtering out the power frequency signal, the lightning transient voltage waveforms are shown in Figure 8. As shown in the figures, the arriving time of lightning transient voltage for the Y-substation and Z-substation are estimated to be 341,862.1 μs and 341,911.1 μs, respectively. The time difference is 49.0 μs. According to DTLM, the initial position (which is usually the position nearest to the lightning stroke point on the transmission line) of the lightning transient traveling wave is about 12.41 km away from the Y-substation, which is between No. 25 tower and No. 26 tower. Meanwhile, LLS indicated that there was a lightning stroke with a return stroke current of −49.5 kA near No. 23 tower. The location error could be caused by the sag of transmission lines, the distortion of transient voltage during transmitting, the nonideal traveling velocity, and the randomness of the lightning stroke.

**Figure 7.** Voltage waveforms recorded on 13:58:27, May 14th, 2017. (**a**) Field measured results in Y-substation; (**b**) field measured results in Z-substation.

**Figure 8.** Lightning transient voltage waveforms. (**a**) Field measured results in Y-substation; (**b**) field measured results in Z-substation.

#### *3.2. Switching Transient Voltages*

#### 3.2.1. Field Test Overview

Series of switching transient voltages were measured by a portable measuring system during the commissioning of 500 kV middle Tibet power grid. In this power grid, the transmission lines are mostly longer than 150 km with an average altitude beyond 3000 m. The operations for the transient voltage measurement mainly include:


For the operation (1), the sensors were arranged under the end terminal of the transmission lines and the sensors were arranged under the entrance conductors of the transformers for the operation (2) and (3).

#### 3.2.2. Typical Measuring Results

(1) Transient voltages during switching on the Tang-Xiang transmission line

Tang-Xiang transmission line connects the 500 kV Batang substation and the 500 kV Xiangcheng substation. The length of this line is about 190 km with the common-tower double-transmission scheme. Before the energization, the line is disconnected at both sides with a suspended potential. The sensors are arranged under the conductors on the Batang substation. The energization process is

accomplished by switching on the breakers on the Xiangcheng substation to connect the line to the 500 kV bus bar.

The typical measuring result after the decoupling process for this type of switching transient voltage is shown in Figure 9. The maximum overvoltage (*U*max) of 679.6 kV appeared on the Phase C which was 1.66 times as the rated value. While the conductor of Phase C was energized, a negative edge of the voltage formed on Phase B at the Xiangcheng side which then reached the Batang side (sensor installed) at *t*<sup>1</sup> (18.226 ms). Then this edge further transmitted back to the Xiangcheng substation and then reflected back to the Batang side at *t*<sup>2</sup> (19.596 ms). During the period between *t*<sup>1</sup> and *t*2, the traveling wave had passed through the transmission line twice. Thus, the actual average velocity of the traveling wave on this line was estimated to be about 2.77 <sup>×</sup> 108 m/s. Moreover, this estimated velocity from field measured data can be used in the fault determination in the future.

**Figure 9.** Transient voltage during switching on the Tang-Xiang transmission line.

(2) Transient voltages during switching on a power transformer to the bus bar

The transient voltages on switching the #3 power transformer onto the 500 kV bus bar on Batang substation were measured and the typical measured results are shown in Figure 10. To avoid involving overvoltage caused by magnetizing rush current during switching on a no-load power transformer, the closing resistors, and phase selection closing strategy were used during the operations. According to Figure 10, the breaker of Phase A were closed firstly and the voltage on the Phase A conductor reached the rated value immediately. Though the breaker for Phase B and Phase C were still open, the voltage on these two phases rose to about half of the rated value. This is because the flux linkage in the power transformer had been established by the winding of Phase A and the winding for the three phases were in the same flux linkage. About four and a quarter periods later (85.84 ms), the breakers of Phase B and Phase C were closed, and a transient overvoltage of approximately 554.2 kV (1.36 p.u.) can be found on Phase C.

**Figure 10.** Transient voltage during switching on the #3 power transformer on Batang substation.

(3) Transient voltages during switching off a power transformer from the bus bar

While switching off the no-load power transformer under non-zero running current, since the current flowing through an inductance never suddenly changes, the residual running current will charge the capacitance connecting to the transformer. This energization process would produce transient overvoltage. The amplitude of the overvoltage mainly depends on the current upon switching off the power transformer. Theoretically, a larger current leads to a higher overvoltage.

The transient voltage while switching off the #3 power transformer from the 500 kV bus bar on Batang substation is shown in Figure 11. The breaker of Phase C was opened while the voltage just crossed the zero level with a low value. For a no-load power transformer, it can be treated as an inductance of which the current lags the voltage by 90◦, indicating the current flowing through Phase C was near its peak value while the breaker opened. In this case, there is an overvoltage with an amplitude of −703.9 kV (1.73 p.u.) on Phase C, while the voltages on Phase A and Phase B are relatively lower.

**Figure 11.** Transient voltage during switching off the #3 power transformer on Batang substation.

#### **4. Discussion**

#### *4.1. Data Decoupling*

According to the measuring principle, a decoupling process is necessary to recover the actual voltage waveforms on a high-voltage conductor from the original measured data, for which the coefficient matrix is the key. Under an ideal condition, the coefficient matrix can be found by static electric field analysis. However, the actual on-site situation brings about unpredictable variations to the matrix. Therefore, determining the matrix from field measured data is more effective and reasonable. In [23], an example for acquiring the coefficient matrix according to the field measured data is presented in which three capacitor voltage dividers for three phases are switched onto the bus bar by disconnectors. Since the switching processes for the three phases are not absolutely synchronized, the changes of transient voltages would appear asynchronously on the conductors. Taking Phase A as an example, if a sharp voltage change happens in Phase A, all the sensors for the three phases will record an electric field change. According to the definition of *k*, the first column of the coefficient matrix could be determined. Using the same method, the second and third columns could also be found by recording a sharp voltage change on Phase B and Phase C. Moreover, based on the phase difference and amplitude among three-phase power frequency voltage, several equations could also be established to solve the *k* coefficient in (5). Above three switching transient measurement results are all decoupled.

For the 220 kV Yan-Zhuang overhead transmission line, no asynchronous transient process has been recorded by far. Thus, it is unfortunate that no opportunity has been given to solve the decoupling coefficients. Instead, a calibration work has been planned on the next power outage of this line. In the calibration, the impulse voltage will be applied to each phase conductor by turns to obtain the decoupling coefficients. However, since the data recorded is all triggered by induction strokes, the three-phase waveforms of transient voltages are synchronized with the same variation trend. Moreover, the coefficients in the matrix of (5) have the same polarity. With linear superposition, in (5), the variation trends of the measured electric field waveform (*E*) and the transient voltage waveform (*U*) would be the same in the time domain.

Setting the waveform on Phase A, Phase B, and Phase C are *U*A(*t*), *U*B(*t*), and *U*C(*t*), respectively. Since they have the same variation trend, they can express as:

$$\mathcal{U}L\_{A}(\mathbf{t}) = m\_{a}\mathcal{U}\_{n}(\mathbf{t})\tag{8}$$

$$\mathcal{U}\_{B}(\mathbf{t}) = m\_{b} \mathcal{U}\_{n}(\mathbf{t}) \tag{9}$$

$$\mathcal{U}\_{\mathbb{C}}(\mathbf{t}) = m\_{\mathbb{C}} \mathcal{U}\_{\mathbb{H}}(\mathbf{t}) \tag{10}$$

where *U*n(*t*) is a normalized waveform, and *m*a, *m*b, and *m*<sup>c</sup> are the amplitude coefficients for *U*A(*t*), *U*B(*t*), and *U*C(*t*), respectively. Taking *E*a-z(t) in (5) as an example, it can be express as:

$$E\_{\Lambda-z}(\mathbf{t}) = (k\_{a\Lambda-z}m\_a + k\_{b\Lambda-z}m\_b + k\_{c\Lambda-z}m\_c)\mathcal{U}\_n(\mathbf{t})\tag{11}$$

It is obvious that the composite electric field has the same variation trend with the original transient voltages. In other words, they will have the same value of time parameters, such as the rising time and half wave time. However, the amplitude value needs to implement the decoupling process.

#### *4.2. Characteristics of Lightning Transient Voltages*

The general waveform of the measured lightning transient voltage is a damped oscillation curve. The duration of oscillation can be several milliseconds. The distribution of rising time (*t*r) and half wave time (*t*h) are shown in Figure 12. The rising time (*t*r) of measured lightning transient voltage varies from several microseconds to about 200 microseconds. The distribution of *t*<sup>r</sup> is mainly in the region from 10 μs to 50 μs with a middle value of approximately 34.3 μs. The half wave time (*t*h) varies from several tens of microseconds to about 350 microseconds. The distribution of *t*<sup>h</sup> is mainly in the region from 40 μs to 160 μs with a middle value of about 105.25 μs. These features are much different from those of the standard lightning impulse voltage [24]. Thus, it is necessary to conduct research on the insulation performances under the standard lightning impulse voltage and the actual lightning voltage they suffered.

**Figure 12.** Distribution of rising time and half wave time. (**a**) Rising time; (**b**) half wave time.

#### **5. Conclusions**

To measure the transient voltages in an AC power grid safely, a non-contact measuring method was developed. An integrated optical electric-field sensor based on the Pockels effect was developed, as well as the measuring system. These measuring systems were used to monitor the lightning transient voltages and measure the typical switching transient voltages.

(1) The measuring system has a response time less than 6 nanoseconds and a relatively flat amplitude-frequency characteristic from 5 Hz to 100 MHz. Its performances meet the requirements for measuring transient voltages in an AC power grid.

(2) In 2017, 307 lightning transient voltage waveforms were recorded on a 220 kV overhead transmission line. All of these transients were caused by induction strokes with damped oscillation waveforms illustrated. Their rising time varies from several microseconds to about 200 microseconds, of which the middle value is 34.3 μs. The distribution of rising time mainly ranges from 10 μs to 50 μs. The half wave time varies from several tens of microseconds to about 350 microseconds, of which the middle value is 105.25 μs. The distribution of half wave time mainly ranges from 40 μs to 160 μs. These features are much different from those of the standard lightning impulse voltage.

(3) Three types of typical switching transient voltages, including switching on the no-load long transmission line, switching on the no-load power transformer, and switching off the no-load power transformer, were measured during the commissioning of 500 kV middle Tibet power grid. A 1.73 p.u. overvoltage was recorded during switching off the no-load power transformer because the breaker cut off the current near its peak value.

(4) Based on the wave transmission theory, the lightning stroke position, the actual transmission velocity of the traveling wave can be estimated with an accurate time of the transient voltage.

The new measuring technology based on optical electric-field sensor supply a more convenient field measuring method for transient voltages.

**Author Contributions:** Methodology, S.X.; equipment development, S.X., H.Y. (Huaiyuan Yang); field experiment, Y.Z., H.Y. (Huaiyuan Yang), H.Y. (Hao Yu), X.C., and S.X.; project administration, S.C.; writing—original draft preparation, S.X., C.Z.; writing—review and editing, R.H., Z.M.

**Funding:** This research was funded by the science and technology project of SGCC, grant number 52199918000G.

**Acknowledgments:** The authors wish to thank the anonymous reviewers, whose comments assisted in improving the clarity of the paper.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **A Traceable High-Accuracy Velocity Measurement by Electro-Optic Dual-Comb Interferometry**

**Bin Xue 1,\*, Haoyun Zhang 1, Tuo Zhao <sup>2</sup> and Haoming Jing <sup>3</sup>**


Received: 20 August 2019; Accepted: 19 September 2019; Published: 2 October 2019

**Abstract:** An electro-optic dual-comb Doppler velocimeter for high-accuracy velocity measurement is presented in this paper. The velocity information of the object can be accurately extracted from the change of repetition frequency, which is in the microwave frequency domain and can be locked to an atomic clock. We generate two optical combs by electro-optic phase modulators and trace their repetition frequencies to the rubidium clock. One functions as the measurement laser and the other the reference. Experimentally, we verify its high accuracy in the range of 100–300 mm/s with a maximum deviation of 0.44 mm/s. The proposed velocimeter combines the merits of high accuracy and wide range. In addition, since the repetition frequency used for the measurement is traceable to the rubidium clock, its potential superiority in traceability can be utilized in velocity metrology.

**Keywords:** electro-optic dual-comb interferometry; laser Doppler velocimetry; Traceability

#### **1. Introduction**

The laser Doppler velocimeter (LDV) is a well-known equipment used to measure motions, fluid, and airflow [1–6]. By illuminating the flow or object with a laser and measuring the scattering caused by movement, it is possible to calculate its speed [7,8]. LDV is capable of providing high spatial resolution and high response speed and is characterized by non-contact measurement. In recent years, the development of the laser Doppler velocimeter has been more and more mature, and various techniques for the LDV have been explored in practical applications [9–12]. Zeyuan Kuang et al. have designed a dual-polarization fiber grating laser-based laser Doppler velocimeter and achieved a velocity measurement range of 2–37 m/s [13]. Recently, the mode-locked laser is applied in absolute distance measurement and its precision is in the nanometer magnitude. Mohammad U. Piracha et al. utilize a train of oppositely chirped pulses to probe a fast-moving target at >91 m/s [1]. Yan Bai et al. utilize a mode-locked laser to measure a target, whose speed is dozens of m/s, through the method of heterodyne Doppler velocimetry, with the measurement error being only 0.4 m/s [14]. However, these technologies are all used for the field of high-speed measurement. Hongbin Zhu et al. have exploited a birefringent dual-frequency laser Doppler velocimeter in the low-velocity area [15]. In their work, the performance of the developed LDV is evaluated through velocity measurements with a range of 0.159 mm/s to 32.273 mm/s. They all achieved a high resolution and high accuracy but had no thought of metrology. In future, measuring instruments that can be traced directly to the natural standard is a tendency [16].

Recently, optical comb technology has been widely exploited in metrology laboratories and physics research and is starting to become commercially available [17,18]. The optical comb performs as a periodic interval comb in the frequency domain. The stabilized frequency comb is capable of functioning as a high precision wavelength ruler, which offers the unique advantage that the measurement uncertainty can be well traced directly to the atomic clock. Consequently, the measurement accuracy can be improved by several orders of magnitude in comparison to other methods [19–21]. Chih-Hao Li et al. have used a laser frequency comb to calibrate a spectrograph and in their work realize velocity measurement of astronomical objects to a precision of 1 cm/s [22]. Recently, electro-optic (EO) frequency combs have attracted the increasing interest of researchers. In 1994 [23], M. Kourogi generated frequency combs by electro-optical phase modulation. After this, he established the OptoComb Company and gradually applied electro-optic frequency combs on scanners, distance meters, and vibrometers. Distance measurement and vibration measurement technology based on the electro-optical comb has been rapidly developed. However, most applications of the optical comb all make measurements through detecting its phase change. Considering that the repetition frequency of the optical comb is located in the microwave field and can be locked to the atomic clock, by using the Doppler effect of the repetition frequency as a basic principle to measure the target velocity it is possible to trace the velocity directly to the atomic clock. This will also utilize the advantage of high time resolution and frequency resolution of the optical comb.

We are thus motivated to develop a method of an electro-optic dual-comb Doppler velocimeter, making the velocity measurement results able to be traced directly back to a rubidium atomic clock, which directly embodies the metrological thought of shortening the tracing links. In this paper, a narrow line width laser of 1543.5 nm is used as the seed laser. Two optical combs are modulated by the electro-optic modulators and their repetition frequencies are locked to the Rubidium atomic clock. The Doppler frequency shift, which occurs in the repetition frequency of the backward reflected light, is detected after the laser is illuminated on the moving target. By utilizing the Doppler effect formula, the target's velocity can be calculated. Inspired by this idea, we deduce some formulas that prove the correctness of the measurement principle. Then, we propose a measurement method based on it. A verified experiment is set up, and the experimental results show that our idea is feasible and that the proposed measurement method makes full use of the high-accuracy characteristic of the optical comb.

The remainder of this paper is organized as follows. In Section 2 we give original rational deduction towards the Doppler effect of optical comb repetition frequency. In Sections 3 and 4 we introduce the measurement method and our experimental system, followed by a detailed analysis of the experimental data. In Section 5 we conclude with a discussion and future outlook.

#### **2. Measurement Principle**

To generate an optical comb, an electro-optic phase modulator is used to modulate the seed laser to produce different frequency sidebands. The modulated laser can be expressed as

$$E = \sum\_{m=-16}^{16} A\_m \cos[2\pi (f\_0 + mf\_{\rm rep}) + q\_m] \tag{1}$$

where *m* is an integer (the range of *m* is [−16, 16], meaning that our experimental device produced 32 sidebands), *Am* is the amplitude of the *mth* sideband, *f0* is the optical frequency of the seed laser, *frep* is the repetition frequency of the modulated laser, and ϕ*<sup>m</sup>* is the initial phase. When this laser beam is incident on the moving target, the corresponding Doppler signal can be written as

$$E\_{dop} = \sum\_{m=-16}^{16} A\_m \cos\left[2\pi (f\_0 + mf\_{rep})(1 + \frac{2v}{c}) + \varphi\_m\right] \tag{2}$$

where *v* is the velocity of the moving target and *c* is the velocity of light. The repetition frequency of the optical comb changes from *frep* to *(1*+*2v*/*c)frep*. Thus, the Doppler shift occurring at the repetition frequency reflects the information of velocity. When the target is close to the laser source, *v* is positive and the repetition frequency is increased. When the target is moving towards the laser source, *v* is negative and the repetition frequency is reduced. From Formula (1) and Formula (2), the velocity of the moving target can be expressed as *v*=*frep*/*2frep*·*c*, where *frep* is the Doppler shift of the repetition frequency.

As the velocity of the moving target is much smaller than that of light, in order to improve the resolution, it is better to increase the modulation repetition frequency. For the same moving velocity, the higher the repetition frequency, the greater the Doppler shift and the higher the resolution. However, detecting high repetition frequency requires very expensive detection equipment. Hence, we developed a method using a dual optical comb to detect the Doppler shift. One optical comb modulates the repetition frequency to *frep.sig* as a signal source and the other optical comb modulates the repetition frequency to *frep.loc* as a local oscillator.

The signal source can be expressed as

$$E\_{\text{sig\\_}i} = \sum\_{m=-16}^{16} A\_m \cos[2\pi (f\_0 + f\_{\text{AOM}} + mf\_{\text{rep.sig}}) + \varphi\_{m1}] \tag{3}$$

The local oscillator can be expressed as

$$E\_{\rm loc} = \sum\_{m=-16}^{16} A\_m \cos \left[ 2\pi (f\_0 + mf\_{\rm rep.loc}) + \varphi\_{m2} \right] \tag{4}$$

Then, the beat signal manifests as a new comb in the frequency domain and can be detected by an oscilloscope. The detected beat signal can be expressed as

$$I \propto \sum\_{m=-16}^{16} \cos[2\pi (f\_{AOM} + m(f\_{rep.sig} - f\_{rep.loc}) + (\varphi\_{m1} - \varphi\_{m2}))] \tag{5}$$

After being reflected by the moving target, the signal source changes to

$$E\_{\rm sig\\_f} = \sum\_{m=-16}^{16} A\_m \cos[2\pi (f\_0 + f\_{\rm COM} + mf\_{\rm rep\,sig})(1 + \frac{2v}{c}) + \varphi\_{m3}] \tag{6}$$

Hence, the detected beat signal can be expressed as

$$I' \propto \sum\_{m=-16}^{16} \cos[2\pi((1+\frac{2v}{c})f\_{\rm{AOM}}+\frac{2v}{c}f\_{\rm{0}}+m(1+\frac{2v}{c})f\_{\rm{rep.sig}}-mf\_{\rm{rep.loc}})+(\varphi\_{m3}-\varphi\_{m2})] \tag{7}$$

When comparing Formula (5) and Formula (7), it can be seen that the repetition frequency changes from *frep.sig-frep.loc* to *(1*+*2v*/*c)frep.sig-frep.loc.* In this way, the velocity of the moving target can be calculated.

#### **3. Experiment Setup**

Figure 1 shows the experimental system. The output of the seed laser (RIO ORION Laser Module, 1543.5 nm, 5.1 kHz line width, 18 mW) is spilt into two parts. One part is modulated by an electro-optic phase modulator (EOM, EOSpace, <5 V half-wave voltage Vπ, 2 W RF power) with 10 GHz repetition frequency, which works as the seed of the signal source. The output of the EOM is amplified to 500 mW by an Er-doped fiber amplifier. After the optical circular, the beam size of the signal source is expanded to 20 mm by a large collimator (micro laser system, FC40), which can guarantee that the received reflected light is strong enough. Then, the laser is incident on a moving corner prism, which is mounted on an electrical rail (Zolix, LMA-TR-200, travel 200 mm, resolution 1 μm/s) whose velocity can be precisely controlled. The other part serves as the seed of the local oscillator. An acousto-optic modulator (AOM; AA opto-electronic MT110, 110 MHz frequency shift) is used to avoid the frequency ambiguity. Then, the output of the AOM is modulated by another EOM (EOSpace, <5VVπ, 2W RF power) with 10.001 GHz repetition frequency. The RF driver1 (Agilent N5173B) and RF driver2 (Agilent E8267D) are locked to a rubidium atomic clock (Microsemi 8040, 2×10−<sup>11</sup> stability with 1 s averaging

time). Finally, the two parts are combined by a beam splitter and detected by a fast photodetector (Menlosystems FD310). The stored waveforms by an oscilloscope (LeCroy Waverunner 610Zi) can be processed to obtain the Doppler shift information.

**Figure 1.** Experimental system for the dual optical comb laser Doppler velocimeter (LDV). Legend: RF1, RF driver1; RF2, RF driver2; EOM, electro-optic modulator; AOM, acousto-optic modulator; EDFA, Er-doped fiber amplifier; PD, photodetector.

#### **4. Results and Discussion**

Before moving the target, beat notes between the signal source and the local oscillator were measured by a spectrum analyzer (ROHDE&SCHWARZ FSH8, 3 kHz RBW). Then, the waveform obtained by the oscilloscope was Fourier transformed to obtain the new comb due to beat. As shown in Figure 2, we can see that the Fourier transform can basically reproduce the beat frequency spectrum. Additionally, the center frequency of the beat notes is 110 MHz, which is the driving frequency of the AOM. The frequency interval of the comb teeth is 1 MHz, which is the frequency difference between the driving frequencies of the two RF drivers. Actually, the frequency of RF driver1 is 10 GHz and the frequency of RF driver2 is 10.001 GHz. This means that *frep.sig* = 10 GHz and *frep.loc* = 10.001 GHz.

**Figure 2.** Beat notes between the signal source and the local oscillator. The blue line is measured by a spectrum analyzer with the left vertical axis; the red line is the result of Fourier transforming the waveform obtained by the oscilloscope with the right vertical axis.

To show the feasibility of the electro-optic dual-comb Doppler velocimeter, Figure 3 shows the difference between Fourier transforming the waveforms when the target is stationary and when it is moving. In Figure 3a, the black comb expresses the measured spectrum when the target is stationary and the red comb is that when the target is moving away from the collimator. It can be seen that every comb has a reduced offset due to the Doppler effect. As the target's velocity is much lower than that of light, the Doppler shift generated at 10 GHz repetition frequency is too small to be seen in the spectrum. In order to show that the repetition frequency has changed, we translate the red comb and move its center frequency to 110 MHz in Figure 3b. Hence, the changes of frequency interval can be seen on the high-order comb teeth due to the cumulative effect. It is thus proved that the Doppler effect causes a change in the repetition frequency, which can be used to calculate the velocity of a moving target. To show the directional discriminability, Figure 3c shows the frequency spectrum when the target is stationary, as well as leaving from and approaching the collimator with *v* = −100 mm/s and *v* = +100 mm/s. As can be seen, for the same teeth of the combs, the green line is on the left and red line is on the right. This proves that approaching leads to an increase in the repetition frequency and leaving leads to a reduction in the repetition frequency.

**Figure 3.** Beat notes when the target is moving and stationary. The black line is the spectrum when the target is stationary; the red line is the spectrum when the target is moving away from the collimator. (**a**) Fourier transforming the waveform obtained by the oscilloscope; (**b**) translating the red comb and moving its center frequency to 110 MHz; (**c**) frequency spectrum of the target in different states.

When processing the data, we selected the two teeth at 110 MHz and 126 MHz to calculate the change in repetition frequency. This means taking the average of 16 repetition frequency data and increasing the resolution by 16 times. The greater the distance between the two teeth involved in the calculation, the higher the resolution. However, we saw only 16 sidebands in the spectrum analyzer; the other sidebands were suppressed due to their weak intensity. Hence, we selected those two teeth to make the calculation. We have conducted experiments to set the moving target at *v* from 100 to 300 mm/s with an interval of 50 mm/s and in the directions toward and away from the collimator. Figure 4 shows the results of velocity measurement. As shown in Figure 4, the velocity of the target is accurately measured by the dual optical comb LDV with a maximum deviation of 0.44 mm/s. Due to the minimum speed of the electric rail being 100 mm/s, we were not able to measure at a lower speed. In addition, due to the limit of the length of the electric rail, the sampling time was too short at a higher speed to collect enough data.

**Figure 4.** Results of velocity measurement. (**a**) The target is leaving the collimator; (**b**) the target is approaching the collimator.

Theoretically, the upper limit of the speed measurement is limited by the difference between the two optical comb repetition frequencies. In this experiment, a difference of 1 MHz limited the maximum speed of the measurement at 15 km/s. Additionally, the resolution was limited by the frequency stability and FFT resolution. Figure 5 shows the Allan deviation of a 1 MHz frequency, which is mixed with the output of the two RF drivers at 110 MHz and 111 MHz. The Allan deviation is well below 14 mHz, with the averaging time larger than 1 s, which corresponds to a resolution of 0.21 mm/s. In our data processing, the FFT resolution was 5 Hz, which corresponds to a resolution of 75 mm/s. Since we selected 16 teeth to calculate the change of repetition frequency, the resolution could be improved to 4.69 mm/s.

**Figure 5.** Allan deviation of 1 MHz frequency.

#### **5. Conclusions**

In order to achieve a wide range and high accuracy in the field of velocity measurement, an electro-optic dual-comb Doppler velocimeter was proposed and studied in this work. In principle, the proposed velocimeter which was derived is feasible and can further improve velocity measurement accuracy. Experimentally, we locked the repetition frequencies of the two optical combs in the microwave frequency domain and beat them to make the detection. In this way, we verified its high accuracy in the range of 100–300 mm/s. The measured range of 4.69 mm/s to 15 km/s was theoretically derived. The proposed velocimeter combines the merits of high precision, high accuracy, and wide range. In addition, since the repetition frequency used for the measurement is traceable to a rubidium clock, its potentially superior traceability can be utilized in velocity metrology. It is of great meaning and necessity because it helps to provide an available velocimeter with high stability and an extremely compact configuration, making a potential contribution to the velocimeter in practical engineering applications.

**Author Contributions:** B.X. designed the experiment. H.Z. and T.Z. performed the experiment. H.Z. and T.Z. analyzed the data. H.Z. and T.Z. wrote the original manuscript. B.X. and H.J. did the review and editing. **Funding:** This research was funded by the Research Project of Tianjin Education Commission (No.JWK1616). **Acknowledgments:** We thank Zhiyang Wang and Kai Zhang for their help in performing the experiment. **Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Monolitic Hybrid Transmitter-Receiver Lens for Rotary On-Axis Communications**

**René Kirrbach 1,2,***∗***, Michael Faulwaßer 1, Tobias Schneider 1, Philipp Meißner 1, Alexander Noack <sup>1</sup> and Frank Deicke <sup>1</sup>**


Received: 29 January 2020; Accepted: 19 February 2020; Published: 24 February 2020

**Abstract:** High-speed rotary communication links exhibit high complexity and require challenging assembly tolerances. This article investigates the use of optical wireless communications (OWC) for on-axis rotary communication scenarios. First, OWC is compared with other state-of-the-art technologies. Different realization approaches for bidirectional, full-duplex links are discussed. For the most promising approach, a monolithic hybrid transmitter-receiver lens is designed by ray mapping methodology. Ray tracing simulations are used to study the alignment-depended receiver power level and to determine the effect of optical crosstalk. Over a distance of 12.5 mm, the lens achieves an optical power level at the receiver of −16.2 dBm to −8.7 dBm even for misalignments up to 3 mm.

**Keywords:** hybrid lens; optical wireless communications; Li-Fi; freeform lens; optic design; rotary interfaces; rotary joint; wireless rotary electrical interface; rotating electrical connectors; full-duplex data transfer; Gigabit-Ethernet; industrial communications; real-time

#### **1. Introduction**

Reliable, real-time connectivity is the backbone of industrial automation. Data transmission over rotating parts is required in a broad range of applications such as wind turbines [1], industrial communications [2,3], surveillance radars [3–7], military [2,8], aerospace [9] and many more. Table 1 gives an overview of the most important transmission principles used in rotary communication links. Slip rings were widespread in the past. However, due to mechanical contact they suffer wear and tear which limits their durability [1,2,5,6,9–11] and thus increases maintenance costs. Precious brush materials and lubricants are used to extend lifetime [12] at the expense of increased system complexity and higher costs. Therefore, contactless data transfer is favored nowadays [6,10]. Life times of several hundreds of revolutions are common and rotation speeds in the order of magnitude of 10<sup>3</sup> rpm or even 10<sup>4</sup> rpm are reached with contactless methods.

Capacitive-based near-field transmission links are known as reliable and cost-effective [11]. However, the system proposed by Doleschel et al. [13] shows that system complexity of modern solutions is clearly not negligible. Data rates of up to several Gbit/s are possible [11,13]. Practical systems support conventional Gigabit-Ethernet and industrial protocols like ProfiNET and EtherCAT for instance. Thus, devices with data rates ranging from 500 kbit/s to 1 Gbit/s were developed [13]. The maximum transmission distance is in the range of 1 cm. Inductive coupling is mainly employed for power transfer and only rarely used for high-speed data transmission [14,15].


**Table 1.** Common principles for rotary data transmission and their typical state-of-the-art performance. Values should be understood as orders of magnitude rather than exact values. \* Higher values are possible but have not been published yet; ‡ this work is limited to on-axis scenarios.

Several systems using conventional low-power radio-frequency (RF) technologies were proposed [1,10] with data rates in the range of tens of Mbit/s or below. Standards like 802.11ac and 802.11ad might be able to provide data rates in the range of Gbit/s. Their main problem is reliability and robustness in terms of ensuring a bandwidth and low-latency data transfer [2,3,13]. Future millimeter-wave based communication [20] standards like IEEE 802.11ay might even reach data rates in the range of several Gbit/s to several tens of Gbit/s [11]. However, their practicality and cost-effectiveness has to be proven.

Highest data rates are reached with fiber optical rotary joints (FORJ). Besides their superior data rate in the range of Gbit/s up to multiple tens of Gbit/s per channel [6,7,13], these links provide immunity against RF. Single-fiber systems only consist of an optical transceiver at both sides and optical fibers in-between. However, due to sophisticated mechanical alignment [8], these systems are expensive. Multi-fiber links offer even higher data rates but exhibit a very high complexity [6].

Optical wireless communications (OWC) aim to combine the advantages of rotary FORJ with relaxed mechanical tolerances, reduced system complexity and thus lower costs. Initially only light emitting diode (LED) based systems with data rates in the lower Mbit/s range [2,3] or laser diode based uni-directional links were demonstrated [21]. Faulwaßer et al. [19] introduced a full-duplex link for data rates of up to 10 Gbit/s for on-axis rotary data transfer. Similar to fiber-based communications, data rates are likely to increase in the future. The demonstrated rotation speed from 0 rpm to 1400 rpm [19] was limited by the test equipment. There is no OWC-exclusive factor limiting the speed. The lifetime of OWC links is expected to be similar to other contactless methods, since there are no significant aging effects. The optoelectronic components, i.e., light emitting diodes (LEDs), laser diodes (LDs) and photodiodes (PDs) are known for high reliability and long lifetime [22–24]. The key element of the transceiver in [19] is a monolithic hybrid lens that acts as transmitter (TX) and receiver (RX) optics in parallel and thereby relaxes the mechanical alignment to several millimeters. The form factor of the transceiver is only 5 mm × 5 mm × 5 mm [19].

This article investigates the potential of OWC for bidirectional, full-duplex, on-axis rotary scenarios and describes how to design a monolithic, hybrid TX-RX lens. In Section 2 a channel model is used to derive some adequate figures of merit. The use of a hybrid lens is motivated by discussing several realization approaches of rotary OWC. Next, the design procedure of a hybrid lens is described and the choice of design parameters is discussed. In Section 3, the performance of the lens and a second system is evaluated and compared using optical ray tracing simulations. Thereby, the alignment-depended optical signal power at the receiver and optical crosstalk is studied. The results are discussed in Section 4. Finally, Section 5 provides a short summary.

#### **2. Materials and Methods**

#### *2.1. Fundamental Concepts*

#### 2.1.1. Channel Model

OWC use optical emitters like LEDs or LDs at the TX to convert an electrical signal into the optical domain. A PD is used for back-conversion at the RX. Similar to FORJ, both transceivers are placed in front of each other [25]. OWC use lenses instead of optical fibers to enable larger mechanical tolerances. The key goal for the designer is to increase the optical power *P*PD that falls onto the PD in order to improve the RX signal-to-noise ratio (SNR) [26] and to minimize the bit-error-rate (BER). If the SNR is already sufficient, the excess can be converted into a higher data rate by increasing bandwidth or applying a multilevel modulation scheme.

There will always be a misalignment between both transceivers due to positioning, assembly tolerances or vibrations. The rotation axis might even exhibit a nutation, i.e., nonideal motion around the ideal rotation axis. The combination of these nonidealities and the communication distance leads to a minimum field of view (FOV) that is required for robust operation. The FOV sizes are denoted by the half-opening angles *θ*TX FOV and *θ*RX FOV. In order to ensure eye-safe operation, i.e., to classify the system as laser class 1 according to IEC 60825-1:2014 (DIN EN 60825-1:2015-07) [27], the optical transmitter power *P*TX is limited. Therefore, efficient transceiver design is required to meet link-budget requirements, ensure eye-safety and to maximize data rate. The designer tries to maximize the dynamic range by keeping the TX and RX performance constant within a plane perpendicular to the optical axis [28]. In other words, TX has to provide constant irradiance *E*TX and RX has to detect the same signal level within this plane.

For a moment we assume the distance *z* between both transceivers is large compared to their apertures. Although this does not fully apply for short distances, the assumption is useful to show the fundamental dependencies. As a result, we can assume that parallel rays are incident onto RX. For a homogeneous FOV, the optical power *P*PD can be expressed as product of the irradiance *E*TX(*z*), the effective receiver input aperture *A*RX,eff and the geometrical coupling coefficient <sup>c</sup> as shown in Equation (1). For the ideal optical link, TX and RX have the same FOV size and both FOVs are placed on the optical axis. In practice, both FOVs may differ or they might be misaligned. The geometrical coupling coefficient <sup>c</sup> quantifies this misalignment. It becomes crucial for short-distance communication as we know from the challenging assembly tolerances of FORJ. <sup>c</sup> is defined as the fraction of the illuminated area of TX, which is overlapping with the FOV of the RX divided by the total illuminated area *A*FOV TX. For the ideal optical channel, i.e., equally sized TX and RX FOVs with no misalignment, <sup>c</sup> = 1 applies.

$$P\_{\rm PD} = E\_{\rm TX}(z, \theta\_{\rm TX \, FOV}) \cdot A\_{\rm RX,eff}(\theta\_{\rm RX \, FOV}) \cdot \varepsilon\_{\rm \mathcal{C}} \tag{1}$$

The effective receiver input aperture *A*RX,eff(*θ*RX FOV) is expressed as product of active PD area *A*PD, optical gain *g*(*θ*RX FOV) and efficiency *η*RX(*θ*i) at the angle of incidence *θ*<sup>i</sup> as it is shown in Equation (2). A large PD area *A*PD is favorable for the link budget but goes along with a large PD capacitance, which limits the modulation bandwidth (BW) [26,29]. Consequently, a PD with large *A*PD but sufficient BW is chosen.

$$P\_{\rm PD} = E\_{\rm TX}(z\_{\prime}\theta\_{\rm TX\,FOV}) \cdot A\_{\rm PD} \cdot \lg(\theta\_{\rm RX\,FOV}) \cdot \eta\_{\rm RX}(\theta\_{\rm i}) \cdot \varepsilon\_{\rm c} \tag{2}$$

The RX FOV should always be chosen as small as possible, since the optical gain *g*(*θ*RX FOV) decreases with increasing FOV due to conservation of Ètendue [30]. Moreover, a restricted RX FOV improves the robustness against noise and interchannel interference. Next, *g*(*θ*RX FOV) is substituted by the maximum theoretical optical gain [30] as shown in Equation (3). Now *η*RX(*θ*i) is used as a figure of merit for the optical RX efficiency. The ideal lossless receiver achieves *η*RX(*θ*i) = 1 for all *θ*<sup>i</sup> ∈ [0, *θ*RX FOV]. The angle *θ*RX,PD,max denotes the maximum coupling angle from the RX optics to the

PD surface normal. In the next step, the irradiance *E*TX(*z*, *θ*TX FOV) is substituted by the product of the optical TX power *P*TX and the TX efficiency *η*TX divided by the illuminated spot area *A*FOV TX.

$$P\_{\rm PD} = \frac{P\_{\rm TX} \cdot \eta\_{\rm TX}}{A\_{\rm FOV \,\,\rm TX}} \cdot \left(\frac{n\_1 \cdot \sin \theta\_{\rm RX, PD}}{n\_{air} \cdot \sin \theta\_{\rm RX \,\,FOV}}\right)^2 \cdot A\_{\rm PD} \cdot \eta\_{\rm RX}(\theta\_{\rm i}) \cdot \varepsilon\_{\rm c} \tag{3}$$

*A*FOV TX is replaced by the corresponding triangular relationship, which includes the communication distance *z* and the tangent of the TX FOV *θ*TX FOV. Finally, Equation (4) is a useful expression for the most important geometric dependencies of *P*PD.

$$P\_{\rm PD} = \frac{P\_{\rm TX} \cdot \eta\_{\rm TX}}{\pi \cdot (z \cdot \tan \theta\_{\rm TX \, FOV})^2} \cdot \left(\frac{n\_1}{\sin \theta\_{\rm RX \, FOV}}\right)^2 \cdot A\_{\rm FD} \cdot \eta\_{\rm RX} \cdot \varepsilon\_{\rm C} \tag{4}$$

The performance merits from Equation (4) are *η*TX, *η*RX and c. The transmitter efficiency *η*TX specifies how much of the emitted power *P*TX is concentrated into the FOV. Its loss mechanisms are *ζ*TX M and *ζ*TX F. *ζ*TX M describes rays that strike the target plane outside the FOV. *ζ*TX F describes back-reflected rays due to Fresnel-reflections.

Since *η*TX contains no information concerning irradiance homogeneity, we additionally introduce the effective transmitter efficiency *η*TX eff. It is defined according to Equation (5) by using the minimum irradiance within the FOV *E*min [28,31]. The difference between *η*TX and *η*TX eff is called the inhomogeneity factor *ζ*TX H.

$$
\eta\_{\text{TX\\_eff}} = \frac{E\_{\text{min}}}{E\_{\text{min,ideal}}} = \frac{E\_{\text{min}} A\_{\text{FOV\ TX}}}{P\_{\text{TX}}} \tag{5}
$$

The sensitivity of RX is typically limited by the interaction of several internal noise mechanisms that exhibit a Gaussian probability function [26]. The signal can additionally be corrupted by optical crosstalk, i.e., adjacent communication channels. This nonGaussian noise has a limited range of variation. This effect introduces a power penalty *PP* [26] that reduces the usable peak-to-peak amplitude of the signal. The optical power that falls onto the PD consists of a signal part *P*PD and another part arising from crosstalk *P*PD cross. The corrected power value *P*PD PP takes the eye-closure effect into account by subtracting the *P*PD cross from *P*PD [26]. It holds *P*PD PP < *P*PD as soon as crosstalk is present. In this work, all optical power values are understood as average values to ensure comparability with literature. When dealing with the power penalty the signal strength is considered in a peak-to-peak manner. It is still valid to consider average values as long as the extinction ratio of the signal and the crosstalk signal is equal. This assumption applies to our case, since both link directions are designed equally and the crosstalk arises from the signal itself.

#### 2.1.2. Ideal Arrangement

The ideal arrangement consists of TX and RX placed at the same position on the optical axis as shown in Figure 1. Both FOVs are equal in size and they fully overlap. The arrangement can be realized by using an LED. On the one hand, applying a forward-bias to the LED causes a forward-current and leads to photon generation. Biasing the PN-junction reversely enables fast photo-detection on the other hand. As a result, the transceiver only needs a single optoelectronic component for TX and RX. Data rates of up to 150 Mbit/s were demonstrated for close distances [32]. However, the LED is a rather bad PD. It has a low responsivity, small area *A*PD and low bandwidth [32]. Although a bidirectional link is realizable, data transfer is restricted to half-duplex mode, because the link direction is switched in time domain.

**Figure 1.** Ideal rotary arrangement: transmitter (TX) (blue) and receiver (RX) (green) of a transceiver are placed on the optical axis at the same position. The field of view (FOV) of TX and RX overlap ideally.

For high-speed bidirectional data transfer in full-duplex mode, TX and RX are separated. The optical arrangement becomes more challenging and design trade-offs have to be met. The following section introduces several geometrical arrangements for optical wireless rotary communication scenarios and compares their performance.

#### 2.1.3. Geometrical Arrangements

Figure 2 shows three geometrical approaches for rotary OWC. In Figure 2a TX and RX are placed next to each other separated by a spacing *d*. In Figure 2b, TX and RX are radially separated regarding the optical axis. Third, both elements arranged along the optical axis and the front element are transparent as illustrated in Figure 2c.

**Figure 2.** Schematic illustration of arrangements of separated TX (blue) and RX (green). (**a**) TX and RX are placed next to the rotation axis. The FOV of TX and RX overlap only partially. (**b**) TX and RX are placed at the optical axis and they are separated in radial direction. (**c**) TX and RX are placed in a line along the optical axis and the front element is transparent.

Approach (a) is used by commercial low data rate IrDA transceivers like Vishays *TFBS4711* [33] but also by a high-speed transceiver demonstrated by Faulwaßer et al. [34]. The axis of rotation is placed trough one of the elements or between them. This leads to misaligned FOVs. The designer tries to minimize the spacing *d*. Next, both FOVs are chosen large enough to ensure a sufficient c. As we learned from Equation (4), *P*PD scales with 1/ tan *θ*<sup>2</sup> TX FOV and 1/ sin *<sup>θ</sup>*<sup>2</sup> RX FOV. This penalty is typically significant. A numerical example is given in Chapter 3.

This penalty is not present for the radial separation from Figure 2b. The approach directly provides aligned FOVs and a high c. TX can be placed in the center surrounded by RX as shown in Figure 2b or vice-versa.

The approach shown in Figure 2c avoids shadowing of the back component by designing the front element transparent. If the front component is TX, it has to have separated emission and absorption bands, i.e., it must exhibit a Stokes shift similar to the fluorescence materials [35]. The issue here is the back plane of the emitter: on the one hand, it has to be reflective to direct the transmitted signal towards the other transceiver. On the other hand, it has to be transparent for the incoming signal. This contradiction does not seem to be easily resolved.

Clearly the radial separation from Figure 2b has superior performance over Figure 2a and has no fundamental concept issue like architecture Figure 2c. However, what is the best way of realizing radial separation? Figure 3 illustrates three principles.

**Figure 3.** Schematic illustration of different radial separations of TX (blue) and RX (green) with an optical element (grey). (**a**) Direct integration of TX and RX into a plane. (**b**) Stacking an emitter chip onto a PD. (**c**) An optical element is used to homogenize the TX ray bundle and move it to the optical axis.

Designing a transmitter element surrounded by one or more high-speed PDs or vice versa on a single chip, like it is shown in Figure 3a, comes with many design challenges including process compatibility. Thereby, it seems easier to produce the chips separately and stack them afterwards. A small emitter die is bonded onto a large area PD chip as it is depicted in Figure 3b. The emitter is contacted with bonds or directly trough the PD chip. The PD could be separated into multiple parts to ease the contacting and to reduce the transit time of the electrons and holes as they might limit the bandwidth [26]. The main issues of this approach include crosstalk between TX and RX, shadowing of the PD by the emitter and the fact that both chips are custom designs.

By using an optical system like it is shown in Figure 3c, conventional emitter and receiver chips can be used. Those dies are placed next to each other and a hybrid TX-RX lens is used to redirect the rays to achieve radial separation. The spatial separation of both chips is favorable to reduce electrical crosstalk. Injection molding allows the fabrication of the lens in high volume [36] and low cost. Ultra-precise drilling and milling [36] is used to produce the mold tool.

Since a part of the lens is used to direct the TX rays, the maximum theoretical gain *g*max cannot be reached. From the PDs point of view, the solid angle of the TX lens part is not used for optical concentration. In Equation (4) this is expressed by a reduced *η*RX.

In summary, using a hybrid lens for radial separation is most promising: besides the potential for <sup>c</sup> ≈ 1, commercial emitter and PD chips can be used. Since the hybrid lens can be fabricated at low cost by injection molding, the system costs are expected to be low.

#### *2.2. Hybrid Lens*

#### 2.2.1. Concept

In order to achieve separation in radial direction, the lens is divided into a TX and RX part, i.e., a center part and a surrounding one. Generally, both parts consist of nonrotationally freeform surfaces at the top and bottom of the lens to form a constant irradiance pattern *E*TX and provide a homogeneous gain *g*. In order to limit the lens thickness *t*, Fresnel-structures could be applied to the top and bottom surface. However, it is favorable to keep the top flat to improve reliability, since cavities tend to fill up with particles.

There are two possible arrangements of TX and RX:

1. The emitter is placed centrally and the PD is positioned off-axis as depicted in Figure 4a. In this case only the emission profile of the emitter has to be adjusted. This includes the homogenization of the profile and an adjustment of the angle *θ*TX. This can ideally be achieved with a single freeform interface. Therefore, the lower surface can be used for beam shaping and the top surface can be flat. The downside of this approach is a challenging RX lens design: the focus point of the RX lens part is off-axis.

2. The PD is placed centrally and the emitter is positioned off-axis as shown in Figure 4b. The TX lens part fulfills two tasks: first, it compensates the displacement of the ray bundle regarding the optical axis. Second, it reshapes the ray bundle to the anticipated FOV. Because two surfaces are required, the top aperture of the lens cannot be flat. On the other hand, the design of the RX lens part is simplified, because the focal point is on the optical axis. If the TX part is neglected, the RX lens can be designed to be rotationally symmetrical. However, the shadowing effect of the TX part introduces a nonrotationally symmetric factor. Theoretically, this can be partly compensated by a nonrotationally symmetric RX lens part. The shadowing effect is also mitigated by minimizing the size of TX.

**Figure 4.** Ray path in the hybrid lens system. (**a**) Center TX (blue) and off-axis RX (green). (**b**) Center RX and off-axis TX.

We choose Approach 2, due to the simplified RX lens design. In this configuration, the lens thickness *t* typically results from the vertical distance between both TX surfaces. A low *t* is desirable for size, weight and cost reduction. The costs are reduced, because less material is needed and due to faster processing [36,37]. Nevertheless, a certain thickness *t* is required to keep the refraction angles at the TX surfaces low in order to limit undesirable Fresnel-reflections. If the RX part determines *t*, the surface can be split into several Fresnel-grooves to reduce *t*.

#### 2.2.2. Optic Design Methods

The hybrid lens is a nonimaging optical system. It can be designed by two fundamental approaches [38]: numerical optimization and direct calculation.

Numerical optimization is a straight-forward approach for designing complex optic modules. Modern optic simulation tools like Optic Studio Zemax enable forming and optically simulating arbitrarily shaped optics by overlapping parametric objects. Optimization algorithms like the Levenberg–Marquardt algorithm are used for adjusting parameters of those objects until a sufficiently good result is achieved. Due to the large amount of variables, the optimization is typically inefficient, because of many local minima in the merit function [38]. As a result, it is easy to find a solution, but its performance is very limited, especially if the systems become more complex.

In contrast, direct calculation algorithms follow well-defined design procedures and yield deterministic outcomes. Thereby, they provide better results than numerical optimization methods [38], especially if the systems are complex. A great variety of design methods are known, for example ray mapping [39–46], forming surfaces using Cartesian ovals [47], the simultaneous multiple surface method in 2D [48] and 3D [38,49] or the tailored freeform design method proposed by Ries and Muschaweck [50]. Nowadays ray mapping approach, i.e., the combination of energy mapping in

conjunction with geometrical surface construction, is in the focus of illumination research [39–46]. Here, a mass-transfer problem is solved by transforming the source power irradiance *E*<sup>s</sup> into the target power irradiance *E*t. This transformation is represented by a mapping *φ* : Ω<sup>s</sup> → Ω<sup>t</sup> from the source domain Ω<sup>s</sup> to the target domain Ωt. Then, the laws of refraction and reflection are applied for *k* × *l* points to calculate a corresponding vector field *N* containing the surface normals *ni*,*<sup>j</sup>* with *i* ∈ [1, *l*], *j* ∈ [1, *k*]. The challenge is to find a mapping *φ* which yields a vector field *N* that satisfies the integrability condition for a continuous surface. This condition is shown in Equation (6) [41,42,50]. It states that *N* has to be curl-free or exhibit at least minimum curl.

$$\mathbf{N} \cdot (\nabla \times \mathbf{N}) = 0 \tag{6}$$

Nowadays, parametrization and consecutive optimization are widely employed for generating a mapping *φ* [28,31,39,40]. Circular shaped FOV are formed with the mapping shown in Figure 5a. An equi-flux grid around the source in spherical coordinates *θ*s, *φ* is mapped onto a target grid in polar coordinates *β*,*r*. In 3D, mappings like these typically lead to a normal vector field *N* with substantial curl. Therefore, an optimization procedure purposely distorts the target grid, for instance by varying *r*, to improve the performance. It was shown that this approach works well for on-axis scenarios [28,39]. In case of an off-axis placed emitter and two optical surfaces, the scenario is more challenging due to the nonparaxial nature. Hence, the result will deviate from the anticipated irradiance pattern. An additional variable *β* could be used in lateral direction as illustrated in Figure 5a. The downside is a slower optimization process.

**Figure 5.** Energy and ray mapping. (**a**) Equi-flux grids at the source and target. Adapted from Wang et al. [39]. (**b**) Mapping from non equi-flux source to anticipated target irradiance. (**c**) Input vectors *vi*,*<sup>j</sup>* and output *oi*,*<sup>j</sup>* vectors are used to construct the surface geometry by subsequently calculating points *Pi*,*<sup>j</sup>* and their normals *ni*,*j*.

For the TX lens part of the hybrid lens, inefficient optimization can be avoided by taking the curl of *φ* directly into account. An initial curl-free mapping *φ*<sup>0</sup> is generated, assuming that the resulting vector field *N* exhibits minimum curl [42]. Figure 5b illustrates the irradiance in front of the source *E*s and the anticipated irradiance at the target plane *E*t. The mapping *φ* has to ensure that the infinitesimal area elements at *σ*<sup>s</sup> and *σ*<sup>t</sup> are passed by the same flux. This is achieved by expansion or contraction of the area elements. Mathematically spoken, the mapping has to satisfy Equation (7) for every *σ*<sup>s</sup> of the source grid [42,46].

$$\det(\nabla \phi(\sigma\_{\sf s})) \rho\_{\sf t}(\phi(\sigma\_{\sf s})) = \rho\_{\sf s}(\sigma\_{\sf s}) \tag{7}$$

The term det(∇*φ*(*σ*)) represents the expansion and contraction of the area element [46]. Although many mappings *φ* might satisfy Equation (7), there is only a single one that minimizes transport cost [46]. Solving Equation (7) turns out to be nontrivial [42,46]. In this article, the algorithm proposed by Wu et al. [46] is used since it provides good convergence.

For the RX lens part, a rotationally symmetric concentrator lens is designed. It consists of two sections: the center is based on refraction, whereas the outer section works with total internal reflection (TIR). TIR is generally superior over refraction for large angles *θ*RX PD i, since it limits the lens diameter and reduces Fresnel-losses [51]. Similar to the TX lens part, the surface is calculated from input and output vectors *vi*,*<sup>j</sup>* and *oi*,*j*. In the first attempt, the edge ray principle [30] was applied for generating the input vectors *vi*,*j*. Due to the TX lens part and the discontinuity between both RX lens sections, the edge ray principle is not valid. Therefore, the gain will vary over the FOV and may drop at certain alignments. In the second attempt, this issue is addressed by defining the input vectors for a range of angles of incidence rather than only for the maximum incidence angle. The output rays *oi*,*<sup>j</sup>* are derived from coupling angles *θ*RX PD i to the PD active area. Ideally, they cover the whole half-space in front of the PD.

The normal vector fields **N** of both lens parts and an initial point **P**1,1 for each surface is used to calculate a finite number of surface points **P***i*,*j*. The result is only a point cloud representation for the optical surfaces. Nonuniform rational B-splines (NURBS) [52] are used for interpolation. NURBS are very flexible and they are supported in the most popular computer aided design formats, which makes them suitable for optical simulation in third-party software and for subsequent fabrication.

#### *2.3. System with Separated TX and RX*

In order to show the potential of the methodology and the concept of the hybrid TX-RX lens, a second optical wireless link based on the principle shown in Figure 2a is developed. A concave lens is chosen for TX to widen up the beam. A convex RX lens is used for optical concentration. The distance *d* between the TX and RX determines the maximum size of both lenses. The RX lens must not be too small to enable a sufficient optical gain *g*. Therefore, *d* will be larger compared to the hybrid lens approach. The exact optical parameters are determined by numerical optimization. The hybrid TX-RX lens and the fully separated TX-RX lens system are denoted by H TX-RX and S TX-RX respectively.

#### *2.4. Simulation Parameters*

For optical simulations, Monte Carlo ray tracing in *Optic Studio Zemax 17* is used. For each simulation, 5 · 105 rays are traced. Thereby, polycarbonate lenses are used with a refractive index of *n*<sup>1</sup> = 1.57 at *λ* = 940 nm. We especially investigate Fresnel-reflections which lead to direct optical crosstalk *P*PD cross if the reflected rays reach the PD. The detector is a circular shaped PIN-PD with a radius of 100 μm. The LD has an output power of *P*TX = 4 mW = 6 dBm at *λ* = 940 nm. First, the transmitter part is considered. Then, the full channel is characterized. In order to give a comparable result to other rotary communication technologies, the link distance is set to 12.5 mm.

Conventionally, optical concentrators are characterized by estimating the gain over the angle of incidence by assuming parallel light [53]. However, Figure 6a shows a large divergence of the incident rays. Thus, the assumption of parallel rays does not apply for the present arrangement. The RX performance depends directly on the TX characteristics. Therefore, it is evaluated by the full-channel simulation.

#### **3. Results**

Figure 6a shows a render of a cross-section through the lens. It is 6.0 mm in diameter and has a thickness of *t* = 2.0 mm in the center and *t* = 2.7 mm at the groove. The PD is placed on-axis at the origin. The LD position is (*x*, *y*)=(−0.5 mm, 0 mm).

For the second system, the numerical optimization leads to a separation between LD and PD of *d* = 1.9 mm. The LD is placed at *x* = −0.95 mm and the PD at *x* = 0.95 mm. The TX and RX lens have a diameter of 1.3 mm and 2.2 mm and a focal length of −9 mm and 2.7 mm, respectively.

**Figure 6.** Render of the cross-section through the optical system with TX rays (blue) and RX rays (green). Both devices are placed on the rotation axis (black dotted line) with no misalignment. Only RX rays that hit the detector are shown. (**a**) Hybrid TX-RX lens (H TX-RX). (**b**) Fully separated TX and RX (S TX-RX) in worst-case orientation, i.e., TX lens facing TX lens and RX lens facing RX lens.

#### *3.1. TX Performance*

Figure 7 shows the irradiance at a distance of 12.5 mm, measured from the TX aperture. Table 2 lists the detailed merits of both profiles. Figure 7a shows the donut-like shaped profile of the LD without optics. The off-axis placement of the LD is directly observable as profile displacement along the x-axis. Figure 7b shows that the hybrid TX-RX lens is able to remove the displacement and homogenize the profile. The effective transmitter efficiency *η*TX eff is 57.6 % with a minimum irradiance of 117 μW/mm2 at (*x*, *y*)=(1.5 mm, 1.5 mm). *η*TX and *η*TX eff are both reduced by *ζ*TX M = 24.7 % and *ζ*TX F = 12.1 %. Moreover, *η*TX eff is lowered by another *ζ*TX H = 5.6 % due to inhomogeneity within the FOV.

Figure 7c shows the irradiance for the system with fully separated TX and RX as described in Section 2.3. Due to the profile displacement, *ζ*TX M = 36.0 % of the power misses the FOV. Moreover, the spherical lenses are not able to correct the donut-shaped profile resulting in a high *ζ*TX H = 38.6 %. As a result, the irradiance is very inhomogeneous within the FOV and it drops down to 25.9 μW/mm2 at the right side. This corresponds to *η*TX eff = 12.8 %.

**Figure 7.** Irradiance over *z* = 12.5 mm. (**a**) For the LD without any optics (LD). (**b**) For the LD with the hybrid TX-RX lens. (**c**) For fully separated TX and RX. The black circle highlights the anticipated TX FOV. Note the different color bar scales for the subfigures.


**Table 2.** TX merits for the laser diode (LD), the hybrid TX-RX system (H TX-RX) and the separated TX-RX lens system (S TX-RX).

#### *3.2. Full-Channel Performance*

Figure 8a–c displays *P*PD, *P*PD cross and *P*PD PP at a lens-to-lens distance of 12.5 mm for the hybrid lens. The values are determined by misaligning the receiving transceiver, whereas the transmitting one is placed at the origin. Table 3 shows numerical values for different misalignments in negative x-direction. Figure 8a shows a rotationally symmetric performance in its fundamental structure. However, *P*PD is not fully homogeneous within the FOV for a revolution. The largest variation Δ*P*PD over one revolution is reached if the misalignment is between 2 mm and 3 mm. There, Δ*P*PD is in the range of 3 dB to 3.29 dB. The crosstalk power *P*PD cross in Figure 8b is similarly distributed as *P*PD with a power level which is about 10 dB lower than *P*PD. As a result, the optical power with applied power penalty *P*PD PP in Figure 8c is similar to *P*PD. This can also be observed in Table 3: the difference between *P*PD and *P*PD PP is in the range of 0.1 dB... 0.9 dB (1.023...1.230). The crosstalk *P*PD cross consists of two parts: a constant part *P*PD cross 1 of −32.4 dBm and an alignment-depended part *P*PD cross 2.

**Figure 8.** Incident power *P*PD and *P*PD cross over misalignment at *z* = 12.5 mm. The "x" marks the axis of rotation. (**a**) *P*PD/dBm for the hybrid lens. (**b**) *P*PD cross/dBm for the hybrid lens. (**c**) *P*PD PP/dBm for the hybrid lens. (**d**) *P*PD/dBm for the system with fully separated TX and RX (Section 2.3) in worst-case orientation, i.e., TX lens facing TX lens and RX lens facing RX lens. Note: The graphs are clipped below −30 dBm to highlight features within the FOV. Therefore, *P*PD cross 1 cannot be seen in Figure 8b.

The system with fully separated TX and RX is shown in Figure 8d. It suffers from a displacement of the communication-area in x-direction. The effect can be seen in Table 3; a communication is only possible for a misalignment of 0.5 mm... 1 mm, depending on the data rate.

**Table 3.** Minimum *P*PD depending on the misalignment in negative x-direction (*y* = 0). Values are determined by choosing the minimum value *P*PD on a circle around the center with a radius of the misalignment. Values lower than 40 dBm are very noisy due to the finite number of simulated rays.


#### **4. Discussion**

#### *4.1. TX Performance*

Figure 7 proves the suitability of the ray mapping method based on curl-free mapping for the TX part. In order to assess the performance of the design, it is meaningful to have a closer look at the loss mechanism: the main loss is represented by *ζ*TX M = 24.7 %. This magnitude is quite common. It results from the extended source effect from the LD with regards to the TX lens part and the remaining curl in **N**. Furthermore, the overall ray mapping performance depends on the extent of the TX output aperture and the distance to the target plane [41]. The reason for this is that ray mapping is an optical far-field design method that neglects the rays' position vector on the output aperture.

The inhomogeneity within the FOV is with *ζ*TX H = 5.6 % very low. Generally, a *ζ*TX H below 10 % is a good result. The major part of the remaining inhomogeneity results from the FOV edge, where the irradiance starts to decrease. The Fresnel-loss *ζ*TX F is with 12.1 % in a common region for two material interfaces. Although an anti-reflection (AR) coating might reduce this effect by a factor of 3 to 4, it is challenging to homogeneously coat the nonplanar TX surfaces.

#### *4.2. Full-Channel Performance*

*P*PD PP is in a sufficient range for data transfer, but how does it correspond to the data rate? For a bit error rate of 10−<sup>12</sup> Säckinger calculates a sensitivity of −26.5 dBm for a 2.5 Gbit/s link and −20.5 dBm for a 10 Gbit/s link [26]. Tzeng et al. [54] demonstrated a sensitivity of −21.5 dBm for a 10 Gbit/s link. It can be concluded from Table 3 that *P*PD PP is sufficient for data transmission in the Gbit/s data rate range and even provides a margin for ageing effects and other nonidealities. The optical crosstalk introduces a power penalty of up to 0.9 dB. The magnitude of *P*PD cross depends directly on *P*TX. A lower *P*TX leads to reduced *P*PD cross. The downside is that the SNR cannot be improved by increasing *P*TX if the crosstalk is the dominant noise factor. The constant crosstalk part *P*PD cross 1 is alignment-independent. Therefore, it has to be the internal optical crosstalk. In contrast, the back-reflected signal *P*PD cross 2 from the opposite transceiver depends on the transceiver alignment. The crosstalk and thus the power penalty is reduced if *z* is increased. In the far field, *P*PD is decreasing with approximately *P*PD ∝ *z*−<sup>2</sup> and the crosstalk scales with *P*PD cross 2 ∝ *z*<sup>−</sup>4, since the back-reflected rays have to travel twice the distance. Although this relation is not fully correct for the near field, the trend is still valid.

The crosstalk *P*PD cross 2 can be lowered by reducing the Fresnel-reflections *ζ*RX,F. The planar top surface is well suited for an AR coating. The problem is that the nonplanar top surface of the TX lens part is also affected by the coating. If one is able to solve the TX coating problem, a single coating is twice as effective, because the ray passes the coating two times: first at TX and second at RX. The link-budget is improved by approx. 6 % per coating instead of only approx. 3 %. Moreover, the crosstalk *P*PD cross 2 is reduced. Alternatively, the top surface of the RX lens part could be designed nonplanar in a way that back-reflected rays miss the transmitting device.

Although TX exhibits a quite homogeneous performance, *P*PD PP varies about 6.9 dB along radial direction for misalignment ≤ 3 mm according to Figure 8a and Table 3. Thus, there is clearly some room for improvement for the RX lens part. The main issue results from the mismatch between near and far field in terms of ray optics. This can be seen in Figure 6a: the incident ray bundle (green) exhibits a large divergence. The situation is different at *z* = 50 mm, for instance. The solid angle of incidence is much smaller and the incident ray bundle exhibits a smaller divergence. The larger the distance, the better the design approach with the edge-ray principle works. Another issue is the nonrotationally symmetric shadowing effect of the TX lens part that manifests itself as a local minimum around (*x*, *y*)=(−2.5 mm, 0 mm) in Figure 8a,c. The variation Δ*P*PD = 3.29 dB is not crucial for data transmission. Assuming a misalignment of −2.5 mm and a rotation speed of 10 000 rpm, the link moves 9.42 μm over the surface in Figure 8a during one bit-duration of 1 ns (1 Gbit/s). The change of *P*PD over a sequence of bits is small enough and does not influence the transmission drastically. Ideally, the receiver circuit includes a decision-point control mechanism for continuous adaptation of the decision level to improve the BER [26].

#### *4.3. Suitability for Rotary Scenarios*

Faulwaßer et al. [19] already reported data rates of up to 10 Gbit/s. However, what data rates are generally possible compared to FORJ? From an electrical point of view, OWC is able to achieve similar data rates as single-FORJ. In contrast to FORJ, the PD has typically a larger area and thus a larger capacitance, which limits the bandwidth. Choosing a smaller PD, will reduce the maximum tolerable misalignment. The hybrid lens from this work exhibits 14.5 dB link loss at (*x*, *y*)=(0 mm, 0 mm). This link loss is a part of the active link concept. The signal is directly recovered at RX by amplification and optimally by subsequent analog-to-digital conversion. The magnitude of the OWC link loss depends on the magnitude of tolerable misalignment. Thus, a higher data rate requires a smaller FOV. In summary, the data rate of OWC links might be slightly below single-fiber FORJ due to a higher link loss.

With regards to the communication distance, the hybrid lens is flexible and not restricted to *z* = 12.5 mm. If the lens is designed for larger ranges, the TX beam exhibits a lower divergence and RX should be designed for smaller angles of incidence. Thereby, the hybrid lens can be tailored to the ideal distance of the rotary system.

As mentioned in Section 2.1.3, the hybrid lens approach has the potential to be very cost effective: in contrast to fully separated TX and RX, only a single lens has to be fabricated, potentially coated and assembled.

The proposed hybrid lens works in on-axis configuration like FORJs. Future work will deal with off-axis optical links to improve flexibility. Only data rates in the range of kbit/s to a few Mbit/s have been demonstrated [2,3], which cannot compete with modern capacitive links [6]. Another interesting field of research is the realization of multi-channel optical wireless links similar to multi-FORJs. In nonrotary scenarios, data rates of several hundreds of Gbit/s have already been demonstrated [55].

#### **5. Conclusions**

This work has shown the suitability of OWC for full-duplex, high-speed data transfer in on-axis rotary scenarios. The proposed hybrid lens is able to provide a sufficient RX signal level *P*PD PP of more than −16.2 dBm even for misalignments of up to 3 mm at a communication distance of *z* = 12.5 mm. OWC is therefore able to provide a robust data transfer without the strict mechanical tolerances compared to FORJs. The results show a maximum power penalty resulting from optical crosstalk of 0.9 dB within the FOV. The approach is promising since it allows low-cost fabrication. Besides the electronical components, only a single optical component is required that can be fabricated by injection molding in high volume.

#### **6. Patents**

The hybrid TX-RX lens and its derivatives are covered by several patents including DE102018205559 B3 [56] (WO19197343A1 [57]). Further patents are submitted.

**Author Contributions:** Conceptualization, R.K., A.N.; methodology, R.K., T.S.; software, R.K.; validation, R.K. and T.S.; formal analysis, R.K.; investigation, R.K.; data curation, R.K.; writing–original draft preparation, R.K.; writing–review and editing, M.F., T.S., P.M., F.D. and A.N.; visualization, R.K.; supervision, R.K.; project administration, R.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **Abbreviations**

The following abbreviations are used in this manuscript:


#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Review* **Key Roles of Plasmonics in Wireless THz Nanocommunications—A Survey**

#### **Efthymios Lallas**

General Department of Lamia, University of Thessaly, 35100 Lamia, Greece; elallas@uth.gr Received: 20 October 2019; Accepted: 10 December 2019; Published: 13 December 2019

**Abstract:** Wireless data traffic has experienced an unprecedented boost in past years, and according to data traffic forecasts, within a decade, it is expected to compete sufficiently with wired broadband infrastructure. Therefore, the use of even higher carrier frequency bands in the THz range, via adoption of new technologies to equip future THz band wireless communication systems at the nanoscale is required, in order to accommodate a variety of applications, that would satisfy the ever increasing user demands of higher data rates. Certain wireless applications such as 5G and beyond communications, network on chip system architectures, and nanosensor networks, will no longer satisfy speed and latency demands with existing technologies and system architectures. Apart from conventional CMOS technology, and the already tested, still promising though, photonic technology, other technologies and materials such as plasmonics with graphene respectively, may offer a viable infrastructure solution on existing THz technology challenges. This survey paper is a thorough investigation on the current and beyond state of the art plasmonic system implementation for THz communications, by providing in-depth reference material, highlighting the fundamental aspects of plasmonic technology roles in future THz band wireless communication and THz wireless applications, that will define future demands coping with users' needs.

**Keywords:** wireless NoC (WiNoC); graphene based WiNoCs (GWiNoCs); wireless nanosensor networks (WNSNs); surface plasmon polariton (SPP); GFET; multiple-input-multiple-output (MIMO); graphennas; THz transceiver

#### **1. Introduction**

Wireless data traffic has experienced an unprecedented boost in the past years, and it is expected to increase sevenfold up to 2021 [1]. Data traffic forecasts in wireless communication networks will account for more than 60% of the overall internet traffic by then [2]. Current wireless communications handle data rates of tens of Gbps per link or even more, and the prospect for the future demands will be 100 Gbit/s within 10 years [3], with multiplexed rates well beyond 100 Gbit/s, and eventually Tbit/s. Wireless communications seem to be in advance against conventional wired communications. By 2030, wireless data rates will be sufficient enough to compete with wired broadband rates [4]. Therefore, the use of even higher carrier frequency bands in the THz range is required, via adoption of new technologies equipping future THz band wireless communication systems at the nanoscale, in order to accommodate a variety of applications, that would satisfy the ever increasing user demands for higher data rates. Certain wireless applications such as 5G and beyond communications, NoC system architectures and nanosensor networks, will no longer satisfy their speed and latency demands with existing technologies and system architectures. Apart from conventional CMOS technology, and the already tested, still promising, photonic technology, other technologies and materials such as plasmonics with graphene material, may offer a viable solution on existing THz technology challenges.

At the moment, wireless traffic in the access 5G networks exploits millimeter wave (mmW) bands. In order to accommodate the continuously increasing traffic demands of 5G and beyond

communications, researchers have been focused on taking advantage of higher regions in the radio spectrum, pointing to the THz band communication and infrastructure, as a promising solution to equip 5G plus networks, thus enabling efficient operation of bandwidth hungry applications, that are not feasible at the moment with current infrastructure.

Wireless NoC (WiNoC) [5] with its inherent broadcast capability, appears as a promising approach to overcome all abovementioned bottlenecks of ancestor technologies. Ultra small miniature sizes of plasmon based antennas and other nanolink components as well, with considerably much less wiring, are the desired features of this technology, in order to enable the integration of one or multiple antennas per core, paving the way for dense, scalable NoC schemes, as required by future applications. Graphene based WiNoCs (GWiNoCs) is probably the most updated promising approach for THz nanoscale wireless communications, and it is therefore considered to be the basis for implementing future on chip network architectures. Alternatively, hybrid optical wireless schemes, may be also proved to be a promising NoC solution [6], by combining the best assets of these two worlds: low loss dielectric waveguide media, and miniature sized plasmonic material oscillating at THz rates.

Last, wireless nanosensor networks (WNSNs) is another established THz nanoscale application, with basic similarities as in WiNoCs, such as core to core or to memory communication, but also with other unique characteristic types of communication, mainly between nanosensors and nanomachines in the THz band [7]. Such EM communication in the THz band, is usually enabled by plasmonic materials, as graphene for implementing plasmonic nanotransceivers and nanoantennas, as in the WiNoC case. These three important wireless THz nanocommunication applications can be seen altogether in Figure 1.

**Figure 1.** Applications for wireless THz nanocommunications.

Nonetheless, the THz band as the last undiscovered frontier of the total EM spectra range, and THz wireless nanoscale communications, are still urging for an efficient, compact and standardized interconnect solution for generating, transmitting, propagating, and detecting the THz wave information. Despite the fact that new efficient methods have been introduced based on modern system architectures via the utilization of new technologies for manipulating THz Band signals by the research academia, still there are many challenges to face, such as the very high propagation signal loss, the impedance mismatch between THz link components, the limited size restrictions along with integration potentials, associated with high bandwidth availability and ultrafast operating data rates with minimum latency requirements. Conventional CMOS based electronic interconnects are definitely far from the target to meet THz speed, low propagation signal loss, and the impedance match between THz link components. Current CMOS scaling technology growth has restricted the cut-off frequency and the maximum oscillation frequency of the device to several hundreds of a GHz. Traditionally, the operating frequency should be much below the cut-off frequency when designing

mixed-signal circuits at high frequencies, and the CMOS device technology applied, should be capable of continuously scaling down the size and values of on-chip components (e.g., transistors, capacitors, and inductors), in order to achieve higher throughput and reduced circuit footprints [8]. Evidently, at THz frequencies this approach has weak potential for improvements, which are further limited by the loss encountered in on-chip metal structures [9]. The alternatives for such a case, would be either to search for a new implementation technology that offers better scaling prospects and a higher intrinsic transit frequency, or to exploit the non-linearities of the device for efficient power generation at higher-order harmonics. As concerns the latter, there are approaches that use novel circuit to generate, radiate, and control THz frequencies widely adopted in academia and industry [10], however it is high time for other technologies to take this on, at this critical stage.

Artificial intelligence (AI) computing with its hot AI chip topic may be an alternative for CMOS wired interconnection and NoC architectures, against CMOS process and device bottleneck and Von Neumann and memory wall bottlenecks. In a broader sense, the AI chip is the adoption of AI principles in computing processing systems in the form of accelerators, in order to boost their computing performance. AI approaches such as processing-in-memory (PIM), machine learning (ML), and especially neural network (NN)-based accelerators, such as FPGA, GPU and ASIC, are considered as mature solutions for speeding up computing performance [11]. PIM or near data computing (NDC), is a promising solution to tackle with the memory wall bottleneck. PIM architectures put additional computation logic in or near memory by leveraging 3D memory technologies to integrate computation logic with memory, thus speeding up NN computations on larger memory capacity and bandwidth, via in-memory data communication, at the same time. The metal-oxide resistive random access memory (ReRAM) has showed great potential to be used for main memory, with its crossbar array structure, capable for performing matrix vector multiplication efficiently, and accelerating NN computations [12].

However, as processing system scalability increases, AI processing loads are getting more and more data-intensive and demand higher bandwidth and heavy data movement between computing logic and memory. Hence, when the scale of the NN computation and accelerator increases, the NoC-based data communication within NN accelerators would evidently have to deal with a performance bottleneck. In addition to performance, the energy consumption of the NoC in an NN accelerator may be also a big challenge to deal with. Particularly, when all this AI processing is done at the edge level, embedded in sensors, smartphones or general IOT equipment, there are more strict power requirements and the need for much more specific hardware implementations [13]. Compared with cloud applications, the application requirements and constraints of edge devices are much more complex, and special architecture design may be needed for handling different situations. Among them, the most important feature and at the same time, request for current edge IOT devices, is their ability to locally perform "inference", relieving thus processing burden from cloud servers and reducing delay [13]. However, in such a case, the demand for training in edge embedded devices is not very clear, given that in the future, all these wearable IOT devices should be capable to perform efficient inference computing, which in turn, requires them to have sufficient inference computing ability, so as to achieve a certain intelligence threshold under the strict power and cost constraints of the edge area, in order to meet the challenges of various different AI application scenarios. Efforts from the research community have been made towards the direction of locally reducing accuracy, and computational complexity, by combining some data structure transformations, such as FFTs to reduce the multiplication in matrix operations, or table lookup to simplify the implementation of multiply-and-accumulate (MAC) operations. Moreover, various low power methods have been applied to AI chips of edge devices to further reduce the overall power consumption, such as the clock-gating applied to MAC. Nowadays, industry has been focused in developing specialized AI chips and all kinds of IOT devices with enhanced inference capabilities at low power and costs. The collaborative training and inference among cloud and edge devices would be an interesting direction to be explored by research academia [13].

As current NN training accelerators relied on conventional wired NoCs, seem to have to deal with certain limitations, especially as time goes by, WiNOC on the other hand, with its inherent features, as broadcast support and multiple access to the shared medium by beamforming and antenna beam narrowing, spatial multiplexing within package, reduced latency, as wireless channel is a distance independent communication means, flexibility in a sense of virtually mapping different topologies within a cycle, and scalability potential as systems scale linearly with the number of cores, may be also considered as an alternative, for being exploited as an interconnection means for these accelerators [14]. Since NN training accelerator parallelized computation nature in a many-core-like environment, is similar to one-to-all, or all-to-all WiNoC communication nature, focusing on exploring the matching points between these two pillars, may be an interesting direction to be investigated by research academia. In this way, the improvement of the efficiency of a WiNoC architecture should be sought not only on the implementation of the miniature antennas and transceiver wireless equipment, but also on the proper design of the NN architecture, as concerns the intensive data movement between processing core and memory units.

In general, the AI chip concept is a complex multi-variable issue, lying in the middle of a whole layer stack, with demanding tasks ranging from providing efficient support for higher layer cloud applications and algorithms, up to orchestrating entities based on AI principles in low level architectures, consisting of devices and circuits, processes and materials, and hence there are a lot of unsolved issues and unanswered uncertainties that may be set under consideration [13], which is out of the scope of this work. Despite the fact that AI chips have made significant progress in the area of ML and NN computations, it is still in its infancy stage, and there seems to be a long way to go, before achieving a generic standardized AI framework; the so called artificial general intelligence (AGI), capable for solving out heterogeneous nature AI applications, especially at the edge network [13].

The photonic based interconnect solution is undoubtedly a viable approach for providing high data rates at low propagation losses, still, their component size is one with two orders of magnitude larger than what is required for the THz band case. Plasmon based THz link components on the other hand, due to their extremely small size and their ability to operate at ultra-high rates, may be a promising approach for equipping wireless THz nanoscale communication systems. Moreover, they could be perfectly combined with photonic technology, and particularly with dielectric waveguiding, as plasmonic waveguiding is quite lossy for long interconnect distances.

To this end, there is still an urge to have a comprehensive view on the current progress and recent advances in the wireless THz communications field, that would help researchers to have a reference point, and based on that, to expand their own ideas and directions, and find motivations to further develop research in this field. This work is a thorough investigation on current and beyond state of the art plasmonic system implementation for THz communications, by identifying the target nanoscale applications and major open research challenges, as well as the recent research achievements. It is the aim of this comprehensive survey then, to highlight the key roles of plasmon based technologies on equipping future competitive THz nanoscale communication systems hosting wireless THz nanoapplications, namely NoCs, WNSNs and beyond 5G communications. This survey paper may be well considered as a complementary work of [15], which had emphasized on key roles of plasmonics and silicon photonics, on equipping wired ultra-high bit rate interconnects, ranging from nanoscale intra and inter-chip interconnections, up to board to board and rack to rack interconnections between data centers (DCs).Particularly, the current work aims to complete the fundamental roles of plasmonic elements and mechanisms referred in the previous work, by associating the currently under investigation, THz system infrastructure in wireless communications. The potential of the THz communications is highlighted by illustrating the basic design issues in equipping these three important THz applications, that will define future wireless application demands coping with users' needs. Moreover, key roles of plasmonics for equipping each single, individual part of a future wireless THz nanocommunication link, namely the antennas and the transceiver parts, are also highlighted.

The rest of this paper is structured according to these two pillars, wireless THz nanoapplications and the implementation of future THz transceiver components. Sections 2–4 are each dedicated accordingly on these three major wireless THz nanoapplications, namely NoCs, WNSNs and beyond 5G communications. In each of these sections, in-depth reference material is provided, which includes the latest literature findings regarding the fundamental aspects of plasmonic technology roles and accompanied photonic technologies whenever required, for each one of these wireless THz nanoapplications, respectively. In Section 5, we focus individually on each critical plasmon based, or hybrid component part of a wireless THz nanocommunication link, namely the antennas and the transceiver parts. Finally, we conclude the paper in Section 6, by also providing two summary tables, with all the information aggregated, including all, state of the art and beyond state of the art characteristic plasmon based and hybrid achievements, for equipping future competitive THz nanoscale communication systems and wireless THz nanoapplications, accordingly.

#### **2. WiNoCs**

#### *2.1. WiNoC Architectures Potential*

As the number of processing cores within a chip area increases in pace with the increased requirements of running applications, communication needs increase as well, and given the fact that the chip area is limited, the node architecture complexity increases accordingly. More complex and sophisticated system designs are required, not only for cores, but also for memory parts to communicate with each other efficiently. Evidently, multi-level memory hierarchy communication needs with multi-core architectures, may be causing a communication bottleneck, as they grow in size.

Nowadays, state of the art multi-core architectures are based on wired NoC paradigm designs [16]. The first multi-processing core interconnections were shared bus architectures [17] which later on were replaced by on-chip CMOS electrical wired interconnections, according to the NoC framework. These were originally implemented via metal traces over a substrate forming a PCB [18]. In the last decade, many alternative fabrics and technologies have been progressively proposed in order to deal effectively with the NoC communication bottleneck, such as 3D NoC [19], RF signals over on-chip transmission lines [20], FSO communication systems at IR frequencies and above [21], photonic NoC [22], nanophotonic NoC [23], and recently WiNoC [24,25], or hybrid WiNoC [26,27]. Three dimensional NoCs are definitely an advantageous network architecture with desired features such as low distance, multiple variety horizontal or vertical interconnections, allowing integration of different technologies at different layers. This technology, however, requires thermal management to deal with the increased heat density due to the superposition of active layers and complex alignment methodologies for the precise positioning of the vertical interconnects. On chip RF schemes allow the interconnection of multiple cores over the same channel with dynamic bandwidth allocation, but they don't have much scalability potential, as they require an increased area and power overhead for the implementation of complex multi signal transceivers, and they also have to deal with the energy reflections at the line terminals. FSO systems on the other hand, may be a promising solution for providing high data rates at large bandwidth and operating at high frequencies accordingly, and they still have to tackle with a few issues such as the low transmission power budget due to eye-safety limits, the impact of several atmospheric effects (e.g., fog, rain, etc.) on signal propagation, and the strict alignment between transmitter and receiver that limits the achievable data rates [28]. Photonic and nanophotonic NoCs are definitely suggested for providing ultra-high data rates and bandwidths, they are CMOS compatible, but there are some parts within a NoC chip area, that are difficult to be implemented all optically, such as buffers, memories, and header controllers.

In general, it seems that as the number of cores on a chip increases and hence the communication performance requirements increase accordingly, all conventional wired interconnection and NoC schemes are inadequate to provide at the same time guaranteed desired latency, throughput, bandwidth, and energy efficiency, while wireless or hybrid wireless optical NoC solutions may be proved to be

more promising alternatives. Specifically, WiNoCs, due to its inherent broadcast and multicast features, should be capable of providing improved performance in terms of scalability, flexibility and area overhead for multi-core systems. Only a single wireless transceiver along with integrated antennas and considerably less wiring equipment are required for interconnecting and sharing resources, among all the chip components, instead of many individual wired connections that would otherwise be required in conventional wired NoCs. It is a critical target for WiNoCs, to be able to manage efficiently wireless communication requirements at the core level, by exploiting miniature sizes of plasmon based antennas and other transceiver parts, in order to enable the integration of one or multiple antennas per core, as seen in Figure 2.

**Figure 2.** Wireless NoC(WiNoC) critical target—one nanoantenna per core.

Such antennas are mostly graphene based planar antennas, which radiate signals at the THz band, and utilize the minimum chip area than other conventional metallic counterparts [29]. Evidently, wireless interconnects are feasible to reduce wire equipment and parasitic and area occupation, as well as the power dissipation of long global, or multi hop short wires that would normally be required at wired competitors, providing the same high bandwidth and low latency communication [5]. Moreover, as wireless schemes natively enable all-to-all communication, they deal effectively with many other interconnection related issues such as multi-core interconnection with single memory, data coherency, consistency, and synchronization. Indeed, memory ordered execution operations, and cache coherence operations which involve a single memory image accessible to all processors, are critical in terms of latency, especially as the number of cores on a chip increases, in which case, traditional wired NoCs would be insufficient for guaranteeing such latency conditions, while wireless NoCs may do offer a promising solution [5].

#### *2.2. GWiNoCs*

WiNoC's main enabler is considered to be its on chip antenna, which it is integrated with a proper transceiver. Originally, WiNoC implementations were based on millimeter on-chip antennas radiating in the GHz band, integrated with adequate high frequency transceivers [30]. Nowadays, the research community has been focused on advanced wireless communication at the THz era. By increasing the communication frequency from GHz to THz domain, first, we anticipate for smaller footprints of the transceiver and the antenna, thus improving the integration potential of the system, and second, we anticipate for larger available transmission bandwidth and higher achievable data rates. At this critical crossroad there seem to be two reasonably strong trends to act jointly as a promising solution: the considerable reduction of the size of the current metallic antennas and other transceiver components, so as to operate at very high resonant THz frequencies, which are mostly implemented via graphene material [31], and the adoption of an hybrid, wireless optical approach for WiNoCs, based on seamless integration of optical and wireless links on chip, enabling wireless multicasting and broadcasting of data, at optical frequencies [32].

GWiNoC, is a relatively recent approach that relies on graphene material for implementing not only nanoantennas [33], but also any other THz wireless transceiver part, for fully equipping the interconnection between the cores of a multiprocessor. As mentioned above, it is not feasible to reach extremely high resonant THz band frequencies by simply scaling down current metallic antennas [34], at the expected size of a nanosystem (a few μm) [35], as it would result in a huge channel attenuation. Graphene based nanoantennas on the other hand, are inherently just a few micrometers in size, i.e., two orders of magnitude below the dimensions of future metallic on-chip antennas, and hence they could provide inter-core communication in the THz band (usually between 0.1–10 THz). These graphene inherent features would offer both size compatibility with each continuously shrunk processor core, as well as adequate bandwidth for massively parallel processing.It seems that the ultimate WiNoC design target has been already set, consisting of a single graphene based nanoantenna and a nanotransceiver interconnected for each individual processing core, for managing the data of outgoing and incoming transmissions to the antenna respectively.

Graphene based antennas or graphennas, have shown excellent behavior as far as concerns the propagation of surface plasmon polariton (SPP) waves in the THz band. SPPs are coupled electron-light oscillations at the interface between a dielectric and a metal, that can propagate at the speed of light. SPPs in graphene are confined much more strongly than those in conventional noble metals, and they are electrically and chemically tunable by electrical gating and doping [8]. Hence, graphene can be considered as an appropriate THz tunable material for building THz resonator devices [36]. Graphene based nanoantennas and transceivers are a hundred times smaller in size than conventional microstrip antennas, with equal or higher bandwidth and gain [37]. Its long plasmon lifetime and the very high propagation velocity characterize it as an ideal material for implementing plasmonic waveguides for on-chip communication. Moreover, graphene has been found to be an appropriate material to enable the elaboration of GFET, providing higher speed and lower energy than conventional CMOS devices [38], and what's more, it is CMOS compatible. All graphene THz transceiver components can be combined with graphene-based THz antenna arrays, to achieve dynamic beam forming and steering enabling all wireless NoC scenarios. Therefore, apart from antennas, graphene has been equivalently proposed to build all types of THz transceiver components, as described in [39], and will be discussed in the next subsection.

#### *2.3. Hybrid Optical Wireless NoCs*

On the other hand, optical technology when combined with chip scale wireless interconnections may be as well considered as a promising hybrid NoC solution to overcome the performance bottlenecks of the current state of the art NoC architectures. Unfortunately, all plasmonic based solutions proposed in the literature for wireless applications do not overcome the problem of integration with SOI based NoC platforms. Moreover, plasmonic waveguides display high propagation losses and, therefore, they are not suitable for implementing long range on chip interconnections. An appropriate solution would be based on the adoption of the hybrid combination of plasmonic resonators as nanoantennas, while keeping dielectric waveguides as the feeding elements [6]. Hence, the employment of plasmonic nanoantennas adjusted to dielectric waveguides for building nano-optical wireless links instead of conventional plasmonic waveguide links, with short range propagation limitations would be a promising solution.

Another key design issue for building successfully such hybrid NoC architectures, is the implementation of the perfect coupling of plasmonic antennas with conventional silicon waveguides, guaranteeing full compatibility with Si photonic and nanophotonics circuitry standards. Waveguide

coupled plasmonic antennas may become a drastic solution for a successful coupling without losses [40,41], enabling a hybrid optical wireless approach in the NoC design. The efficient coupling between plasmonic antennas and SOI waveguides is a non-trivial issue, as an on-chip, point-to-point connection normally requires matched directive nanoantennas. The nanoantenna shape and size should be properly designed so as to ensure impedance matching to the waveguide, and directional emission in the desired direction [42].

Moreover, hybrid wireless optical on chip communication takes advantage of the entire WDM spectrum when propagating in the optical wired links, guaranteeing even higher multiple capacities, as required by intra-chip communications [43]. Various configurations of plasmonic nanoantennas for supporting wireless-optical on chip communication have been proposed in the literature, such as plasmonic horn nanoantennas [44], a directional plasmonic Yagi-Uda nanoantenna placed on a dielectric waveguide [45], or a plasmonic nanoantenna array on a dielectric waveguide [46], or various configurations of plasmonic Vivaldi antennas (double, or an array of them) to name but a few [47,48]. Plasmonic antennas will be described more analytically in the upcoming section.

#### **3. Wireless Nano Sensor Networks**

WNSNs is another established THz nanoscale application under the internet of nanothings (IoNT) framework [49], which encourages, not only the core to core or to memory communication as in the WiNoC case, but also the interconnection between other nanoscale components, mainly nanosensors and nanomachines. These nanoscale networks rely on the THz band communication between its different components, which as mentioned, could be either nanosensors or nanomachines [7]. Nanosensors are capable of detecting events with unprecedented accuracy, while nanomachines are dedicated to tasks ranging from computing and data storing to sensing and actuation [50]. Hence, WNSNs are composed by integrated nanomachines and nanosensors, which interact with each other through EM communication [51]. EM communication in the THz band are mostly enabled by graphene based plasmonic nanotransceivers and nanoantennas, as in the WiNoC case.

Main features of theWNSNs are: (i) the size of nano-devices, which range from one to a few hundred nanometers, (ii) the exploitation of graphene based nanoantennas for THz band communication, (iii) extremely high bit rates (Tbit/s), and (iv) very short transmission ranges (tens of millimeters) [51]. Evidently, the THz band is considered as the natural domain for the operation of nanosensor components, as this frequency range supports very high transmission bandwidths within a short range. Alternatively, in the event of transmitting at lower frequencies (e.g., the MHz range), nanosensor devices would have to communicate over longer distances, but the energy efficiency of such a process to mechanically generate EM waves for remote control of these devices would be very low, and hence, communication by using the MHz frequencies wouldn't be an appropriate solution. Consequently, nanosensor devices would properly communicate with each other in the THz band [35].

As mentioned, apart from graphene-based THz antennas, graphene is also preferred for the development of other transceiver components in a scale ranging from one to a few hundreds of nanometers, such as: nanoscale FET transistors, nanosensors, nanoactuators, and nanobatteries. With the exploitation of graphene material, the integration of these nano-components in a single device of just a few micrometers in size is feasible, and will result in implementing autonomous nano-devices, able to perform specific tasks at the nanoscale, such as computing, data storing, sensing or actuation [35].

Depending on the measured parameters, nanosensors could be categorized in three types; namely physical, chemical and biological nanosensors [35]. Physical nanosensors such as pressure nanosensors [52], force nanosensors [53] or displacement nanosensors [54], are used to measure magnitudes such as mass, pressure, force, or displacement accordingly. Chemical nanosensors are used to measure magnitudes such as the concentration of a given gas, the presence of a specific type of molecules, or the molecular composition of a substance. Their working principle of both types is more or less the same and it is usually based on the change of the electronic properties of nanotubes and

nanoribbons when they are used in a FET configuration, whose on/off threshold voltage changes as well by alteration of the value of each measured magnitude.

Last, biological nanosensors are used to monitor biomolecular processes such as antibody/antigen interactions, DNA interactions, enzymatic interactions or cellular communication processes. A biological nanosensor is usually composed of a biological recognition system or bioreceptor, such as an antibody, an enzyme, a protein or a DNA strain, and a transduction mechanism, e.g., an electrochemical detector, an optical transducer, or an amperometric, voltaic or magnetic detector [55]. There are mainly two subtypes of biological nanosensors based on their working principle: electrochemical biological nanosensors which work in a similar way to chemical nanosensors and photometric biological nanosensors. The latter subtype working principle is based on the use of noble metal nanoparticles and the excitation using optical waves of surface plasmons.

More specifically, a typical generic architecture of a WNSN node as seen in Figure 3 [35], would be consisted of: (i) Sensing unit: graphene material and its derivatives, namely, graphene nanoribbons (GNRs) and carbon nanotubes (CNTs) [56], provide outstanding sensing capabilities and they are the basis for implementing many types of sensors [57]. (ii) Actuation unit: an actuation unit will allow nanosensors to interact with their close environment. Several nanoactuators have also been designed and implemented so far [58]. (iii) Processing unit: nanoscale processors are being enabled by the development of different forms of miniature FET transistors in the nanometer scale. They were mostly implemented via CNTs and GNRs nanomaterials. (iv) Storage unit: graphene has shown excellent performance in a number of applications from supercapacitors [59] to photomechanical actuators [60], however, so far, its potential in nanomemory construction has not been adequately explored. (v) Power Unit: there are two types of nanobatteries [61] for feeding nanomachines: (a) harvesting the energy from the environment via nanoscale energy harvesting systems [62] and (b) wireless energy induced from an external power source [63]. (vi) Communication unit: this consists of nanoantennas and transceivers for guaranteeing EM communication between nanosensors. The working principle of energy harvesting is based on the conversion of mechanical or vibrational or hydraulic energy into electrical energy. The mechanical energy is produced by the human body movements, or muscle stretching, the vibrational energy is generated by acoustic waves or structural vibrations of buildings, and finally the hydraulic energy is produced by body fluids, or the blood flow. This energy conversion is achieved by the piezoelectric effect seen in zinc oxide (ZnO) nanowires, as they are bent, when a voltage appears in the nanowires (Figure 4) [35].

**Figure 3.** Wireless nanosensor networks (WNSN) node architecture.

**Figure 4.** Energy harvesting is based on the piezoelectric effect seen in zinc oxide (ZnO) nanowires.

As mentioned, nanoantennas are mostly implemented via CNTs and GNRs nanomaterials. Concerning the latter case, the propagation of EM waves on a graphene sheet was first analyzed in [64] while in [34], nano-patch antennas based on GNRs and nano-dipole antennas based on CNTs were quantitatively compared. EM transceivers of nanosensors are embedded systems that include all the necessary circuit equipment which processes the transmitted or received signals from the free space through the nanoantenna with proper functioning such as frequency conversion, filtering and power amplification. Several GFET transistors capable of such functioning and operating in the sub-THz or THz band have been demonstrated so far [65]. Other materials, such as Au and Ag have been successfully used for plasmonic sensors in the visual [66,67] and near infrared [68], as well. Ge plasmonic material in the mid-IR, has been also proposed as a promising material for replacing silicon, as a substrate for MOS devices [68].

According to [35], the applications of WNSNs can be classified in four main groups: biomedical, environmental, industrial, and military applications. As far as concerns in the environmental and industrial application domains, various THz nanosensors have been used for the detection of pollutants and empowering the technology of food preservation and food processing [69]. Nanowire based nanosensors are suggested for sensing ambient intelligence, such as vertically bridged nanowires, laterally bridged nanowires, ultra-sharp Ga2O3 nanowire, the nanowire FET chem-biosensor, and the CNT biosensor [70]. As far as concerns military applications domain, various THz nanosensors, detectors and cameras have been suggested for security applications, and specifically for the detection of weapons, explosives, as well as chemical and biological agents [71].

As far as concerns in the biomedical applications domain, graphene based nanoantennas have been used for wireless communication between nanosensors, deployed inside and over the human body, resulting in many bio-nanosensing applications [72]. In-vivo wireless nanosensor networks (iWNSNs) at the THz band, is a characteristic application for providing fast and accurate disease diagnosis and treatment [73]. These networks, fully equipped with nanoscale components (e.g., nodes, routers, gateways, links and interfaces) as in regular networks, as seen in Figure 5, are operating inside the human body in real time, achieving precise and real time medical monitoring and medical implant communication. Apart from iWNSNs, graphene plasmons have also been used in biological sensing applications, monitoring the rotational and vibrational modes of DNA molecules and many large proteins in the THz and far-IR [74,75]. Moreover, as far as concerns imaging applications, and especially the THz band spectroscopy, nanosensors have been used for backscattering techniques for monitoring the dynamics of large biomolecules [76]. A THz biosensor communicating with the biological agents through a graphene plasmonic waveguide is described in [77]. In this sensor, SPPs in the waveguide must be launched by either an external near-field source or an antenna. In [78], another biosensor system has been proposed, where all the required elements (the sensor, frequency modulator, and antenna for energy harvesting) are packaged in a single module [78]. This structure has been designed as a dual resonant antenna part of a GFET.

**Figure 5.** In-vivo wireless nanosensor networks (iWNSNs) for healthcare applications.

#### **4. Beyond 5G Networks: Towards to THz Band Communications**

At the moment, wireless traffic in the access 5G networks exploits wide radio bands such as the mmW frequencies and systems, respectively. The short term roadmap for 5G and beyond communication would anticipate the establishment of many high-rate small cells forming the access link, operating in the mm wave spectrum, which are in the order of several Gb/sec, thus, in total, the aggregation capacity of the fronthaul/backhaul link should be several times higher, so as to guarantee reliable and fast data delivery from multiple users, which are connected to the small cell. Hence, in such circumstances, the THz band available bandwidth resource would be highly appreciated, while the high propagation loss of THz band fronthaul/backhaul links would be compensated by the extremely high antennas directivity [79]. In order to accommodate the continuously increasing wireless traffic demands of 5G communications and even beyond, researchers have been focused on taking advantage of higher regions in the radio spectrum (above 300 GHz), pointing to the THz band communication and infrastructure, thus enabling efficient operation of bandwidth hungry applications, that are not feasible for these systems at the moment. Hence, in the case of THz band communication, the supplied bandwidth required, would be one order of magnitude above current mmW systems, offering faster data transfer and download speeds, lower latency, and higher link directionalities [80] with non-line-of-sight (NLoS) propagation [81], as THz waves could penetrate thin objects, able to carry the data, such as the data transmission/reception by a smartphone in a pocket.

Despite the fact that current research works have been focused on various 5G scenarios based on photonic assisted wireless communication systems with very high data rates, the capacity demand of THz wireless systems has not been achieved yet. Uni-travelling photodiodes (UTC-PDs) and comb sources, [82,83], in the W-band (75–110 GHz) incorporating optical polarization division multiplexing (PDM) [84], and spatial multiple-input-multiple-output (MIMO) techniques, are considered as state of the art photonic sources, achieving data rates of 100 Gbit/s and beyond [85,86].

However, as mentioned, higher regions in the radio spectrum at higher operating rates are required, in order to reach THz communication system goals, seeking for brand new drastic solutions. Reference [21] is a survey dedicated to hybrid radio frequency, FSO systems with THz/O links, considering it as a viable approach for equipping future THz wireless communication. Integrated

microwave photonics (IMWP) in the THz range [87], have been also proposed as an enabling approach for equipping 5G wireless systems, by providing optical signal generation and distribution of mm waves towards antenna terminals [88], dynamic filtering [89], optical control of antenna arrays [90], and many more other functions [91]. 5G applications, however, pose very stringent requirements on the speed of IMWP circuits, for processing mm waves or even sub-THz frequencies, in order to access the multi-GHz bandwidths required for high data rates [91].

In practice, there are a few more challenges a designer has to consider, when applying these suggested photonic based technologies in THz wireless communication systems. Specifically, given that free space path loss increases as the square of the frequency, directive antennas and directional transmitters focused on individual users, such as high gain steerable phased arrays, will be required in order to compensate for the large free space path losses of EM waves. Beam steering techniques [92] and massive MIMO antenna array elements arranged on terminal devices [93] may be promising solutions for increasing antenna directivity gain. There are also other seamless integration issues of wireless links combined with photonic infrastructures that might appear, such as the connection of optical fibers to THz transmitter and receiver front ends, which consequently requires optical to THz (O/THz) and THz/O converters with high bandwidths (well above 300 GHz). As far as concerns O/THz conversion, UTC-PDs integrated with sub-THz waveguides are considered an established solution [82,94] but the opposite THz/O conversion is still an issue that needs reconsideration, as it requires modulators with electro-optic bandwidth well above 300 GHz, high power management, and very high linearity. Plasmonic modulators may at this point prove to be a viable solution, as they are characterized for their ultra-compact footprints [95] (10 s μm2), ultra-low power consumption (2.8 fJ/bit at 100 GBd) [96], and flat frequency responses up to 170 GHz [97] and 325 GHz [98]. In [99], the experimental demonstration of a plasmonic MZ modulator with sub-THz frequency responses (up to 500 GHz), high power handling, and high linearity is described. The first demonstration of a THz link seamlessly integrated into a fiber optic network using direct THz/O conversion at the wireless receiver, is described in [100]. An ultra-broadband silicon POH modulator, is used for THz/O conversion of WDM signals in [100]. Alternatively, graphene based THz components have shown very promising results in terms of generating, modulating as well as detecting THz waves [101], and hence they may be considered as appropriate THz band transceivers [102].

#### **5. Plasmonic THz Wireless Nanoscale Link Components**

The main components of a typical WiNoC layout are basically the THz band antenna along with transceiver components, used as a feeding element. Specifically, the transceiver is held responsible for preparing the information for outgoing transmissions to antenna, and for demodulating incoming transmissions from antenna respectively.

#### *5.1. Plasmonic THz Antennas*

#### 5.1.1. Design Issues

There are many challenges when designing a THz band plasmonic antenna, such as the material parameters and properties, the size of the antenna, the impedance mismatch with the coupling waveguides, and its integration potential with processing cores and transceiver parts. As far as concerns the basic parameters that characterize the antenna material, such as the dynamic complex conductivity and permittivity, as well as the propagation properties of SPP waves on the nanoantenna, such as the confinement factor and propagation length, and finally the antenna geometry parameters, such as the length and radius, all these design issues should be taken into consideration before ending up to the proper antenna implementation choice, such as state of the art metallic or hybrid antenna structures [103].

The conductivity of plasmonic materials such as graphene, gold or silver is a complex-valued parameter which affects the global oscillations of electrical charge in close proximity to the surface of the antenna, resulting into the excitation of electromagnetic SPP waves. The frequency at which SPP waves are excited depends on the material conductivity of the antenna components. For example, graphene supports SPP waves at frequencies as low as in the THz band (0.1–10 THz), whereas in noble metals such as gold or silver, SPP waves are only observed at tens of THz and above [104]. There are many analytical models describing the conductivity of metals, among them, the Kubo formula is considered the most appropriate model, associating the complex conductivity and permittivity as functions of the frequency. Particularly for terahertz frequencies, the Drude model contribution is applicable, which takes into account the intra-band electron transitions within the metal energy band structure. According to this model, plasmon material conductivity acquire a different response at THz band, because of its intrinsic kinetic inductance, associated to the imaginary part of the conductivity, that plays the role of negative real permittivity in a bulk material. At these high frequencies, moderate changes in the chemical potential can significantly alter graphene's conductivity and change the sign of its imaginary part [64]. Apart from frequency dependence and band structure, graphene and other metals conductivity also depends on a set of parameters such as chemical doping, Fermi energy (chemical potential), electron mobility, and relaxation time. The relaxation time is the interval required for a material to restore a uniform charge density after a charge distortion is introduced, while the chemical potential refers to the level in the distribution of electron energies at which a quantum state is equally likely to be occupied or empty. These two variables have a strong impact on the resonant frequency and radiation efficiency of the antennas. The value of the chemical potential can be altered by applying an electrostatic bias or chemical doping, providing also significant reconfigurability potential. The bias injects electrons or holes on the active area of the structure, modifying graphene's chemical potential [29].

An example of graphene's dispersive conductivity is shown in Figure 6. In general, the chemical potential variation is directly related to the applied voltage variation. As implied in Figure 6, higher chemical potential leads to better performance through an increase of the conductivity. Figure 7 shows the resonance characteristics of the antenna as a function of the chemical potential. The increase of chemical potential leads to a significant shift of the resonant frequency and to an enhancement of the antenna response without changes in the radiation pattern [105].

**Figure 6.** Graphene's dispersive conductivity at THz frequencies.

**Figure 7.** Nanoantenna resonance characteristics as a function of the chemical potential.

Absorption enhancement based on metal nanoparticle dispersive properties of metal nanoantenna structures is also a critical design issue that needs also consideration. These metal nanoparticle dispersive properties response of different metals at THz range are also described by the Drude–Lorentz model by relative complex permittivity of metal nanoantennas as a function of frequency as seen in Figure 8 [106], with ε1 the real part of the relative permittivity, and ε2 is the imaginary part of it. Different metal nanoantennas like gold, silver, copper and aluminum are considered.

**Figure 8.** Complex permittivity of different metal nanoantennas at the THz range. (**a**) real part and (**b**) imaginary part of relative permittivity.

An approximate 20% increase in absorption performance is obtained. The frequency that corresponds to the maximum absorption is dependent on the nanoparticle material. The maximum absorption is observed in gold nanoparticle material, as seen in Figure 9 [106]. The frequency of the maximum absorption moves toward lower frequencies as physical dimensions of the nanoparticles increase. Hence, the choice of the best plasmonic material for a given application is a subject of discussion and research. It has been also proved that the radiation frequency of the graphene nanoantenna can be tuned in a wide spectral range by adjusting the dimensions and particularly its length. According to [107], the absorption cross section increases up to a given limit with the substrate size, and hence, a larger substrate improves the performance of the graphene nano-antenna. Moreover, the absorption cross section increases as the graphene patch is located closer to the side of the substrate, while the resonant frequency becomes higher when the patch is farther from the center [107].

**Figure 9.** Absorption response at THz range for different metal nanoantennas.

Therefore, the optimal location for on chip graphene nano-antennas may be near the edge of the substrate, in order to maximize their efficiency. As it concerns chemical potential, there is a tradeoff between the amount of power the graphene nano-antenna can absorb and its resonant frequency. Specifically, graphene nano-antennas with zero chemical potential, resonate at a low frequency but with a small absorption cross section, thus limiting their radiation efficiency. On the other hand, nano-antennas with a higher chemical potential experience greater absorption performances with an increased resonant frequency. Thus, there exists a compromise between these two parameters to be taken into consideration.

In [108] there is an exhausting investigation and comparison of the performance in transmission and reception of metallic nano-dipole antennas, implemented via various metals such as Cu, Al, Ag and Au. Taking into account each metal property, such as its dynamic complex conductivity and permittivity, the propagation properties of SPP waves and the antenna geometrical features, as length and radius, a mathematical framework is developed in order to analytically derive critical transmission and reception performance parameters such as the generated plasmonic current in reception and the total radiated power and efficiency in transmission.

As far as concerns the THz antenna size future requirements, and given that the available bandwidth is inversely proportional to the antenna size, it requires only a few micrometers in size, in order to build an appropriate nanoantenna, which is almost two orders of magnitude below the dimensions of current on-chip antennas. Hence, the approach of integrating one antenna per core seems rather unfeasible for CMOS technology, especially as the core sizes continue to shrink to a few hundreds of micrometers [109].

Silicon integrated antennas could be considered as a mature option for on chip antennas, but their size ranges from a few to ten mm's, such as the zig-zag monopole antenna of axial length 1–2 mm proposed in [110], or the demonstration of a miniature on-chip antenna operating at the range of 100–500 GHz in [111]. CNT based antennas have been also considered as an alternative approach for implementing THz antennas. Normally the CNT antennas are capable for equipping WiNoCs, as they have small size, low power losses, high transmission power and they operate at very high rates. Unfortunately, they are characterized by significant manufacturing difficulties for implementation [112]. Ultra-wide broadband (UWB) and multi-band antennas have been also proposed for on-chip wireless communication, due to their multi-channel capability, that can be shared among multiple nodes. A CMOS UWB based antenna for WiNoC has been proposed in [113], its short transmission range, however, does not give much prospect for long distance communications. Horn and paraboloid antennas have been also proposed for transmission at 300 GHz, with a radiation bandwidth in the order of 10 percent of their center frequency, their geometry however, makes them not suitable for mobile and personal devices [28]. The very small size of a THz Band antenna is an uncompromising necessity required for future wireless THz communications, so as to allow the integration of a very large number of antennas with very small footprints, forming very large antenna arrays. New antenna array patterns, such as massive MIMO schemes based on graphene plasmonic material for doping, may be a promising approach [93].

#### 5.1.2. Graphennas

Graphennas are just a few micrometers in size, thus enabling size compatibility with each processor core, and providing enough bandwidth, and hence, they are appropriate candidates for inter-core communication in the THz band. Graphene is a one-atom thick layer of carbon atoms in a honeycomb crystal lattice, and it is considered as an attractive solution, due to its unique electrical and optical characteristics [114]. Graphennas show excellent behavior, as far as concerns the propagation of SPP waves in the THz band. As known, SPPs are electromagnetic waves guided along a metal-dielectric interface and generated by means of an incident high frequency radiation [115]. A graphene layer supports transverse magnetic SPP waves with an effective mode refractive index. A basic configuration of graphene THz nanoantenna is shown in Figure 10. The nanoantenna is basically composed of a graphene layer, which is the active element, along with a metallic flat surface which is the ground layer, and a dielectric material layer in between the former two layers. An antenna feeding mechanism is also required in order to complete the nanoantenna layout [115]. By adjusting the dimensions of the graphene nanoantenna, the radiation frequency can be tuned in a wide spectral range.

**Figure 10.** Graphene THz nanoantenna configuration.

As mentioned in the previous section, all plasmon based THz solutions are characterized by relatively high losses of the supported SPP modes, especially during waveguide propagation, hence hybrid combinations of plasmonic resonators with dielectric waveguides appear to be an attractive alternative solution. We may at this point, consider a broader generic hybrid combination of surface plasmon with dielectric wave modes, providing better results between mode confinement and propagation loss in the THz band, as seen in Figure 11 [116]. In other words, graphene may be used either for the development of plasmonic waveguides, or antennas in hybrid combinations with dielectric material. Specifically, as far as concerns graphene based waveguide structures, and given that THz plasmons can be confined laterally in a graphene sheet [117], this property has led to an opportunity of implementing many different graphene based waveguide structures, consisting of a number of graphene layers, mixed with dielectric materials [118], or with dielectric-metal structures [119], obtaining excellent field confinement results. Other approaches involve the use of graphene layers to form wedges [120], or coat grooves carved on the dielectric [121], or a waveguide consisting of a dielectric with high permittivity, on top of a low index dielectric-graphene-dielectric stack [122].

**Figure 11.** Hybrid combination of surface plasmon with dielectric wave modes for THz nanoantenna implementation.

As far as it concerns graphennas, such structures are based on a number of graphene layers, on a metallic flat surface, a dielectric material in between, and a feed to drive the signals to the antenna, as seen in Figure 10. Other antenna structures, such as patch antenna, and dipole designs, where the feeding mechanism lies in between the two identical graphene patches, have been also proposed in [107,123], respectively. These structures are either based on an ideal photomixer with high impedance as the feeding mechanism, or on more advanced THz sources based on photoconductive materials [124], or on high electron mobility transistors [102]. Furthermore, graphene antennas in MIMO configurations with a considerable number of radiating elements, may be also considered as an updated efficient THz solution [125]. Alternatively, other works are based on graphene potential of tuning ability, rather than acting as a radiating antenna element. Specifically, in [126], graphene sheets are employed between the source and a metallic radiating element to retain the tunability, while in [127], a novel antenna design is proposed, based on hybrid graphene-metal structure for enhancing reconfigurability capabilities of THz antenna. Different hybrid graphene-dielectric antenna structures have been also proposed in [128], such as a two graphene monolayers separated by a thin dielectric structure, or an hybrid structure with two graphene monolayers (H2G), consisting of a layer with a high index material (HIM) for a dielectric mode, close to the graphene layer for a plasmonic mode, and separated by a spacer with a low index material (LIM), as seen in Figure 12d.

**Figure 12.** Pure plasmonic (**a**,**b**) and hybrid graphene-dielectric antenna structures (**c**,**d**).

#### 5.1.3. Other Plasmonic Nanoantennas

Other plasmon based nanoantennas have also been proposed, based on hybrid wireless-optical on chip communication. When designing such hybrid structures, large impedance mismatches between the resonant nanoantennas and the waveguides should be completely minimized. Proper impedance matching elements of certain permittivities are needed to be carefully employed as wireless transceiver parts, and positioned within the connecting gap between the nanoantenna and the waveguide. Classic antenna layouts applied in RF communications such as dipole, Yagi-Uda and phased array configurations [129–131], have been also applied in optical wireless nanolinks, using plasmonic waveguides as matching elements, between pure or hybrid plasmonic antennas and their feeding silicon waveguides. Recently, in [132], a dipole loop plasmonic nanoantenna has shown an increased operating bandwidth compared with a single loop antenna. Plasmonic horn nanoantennas proposed in [41], are impedance matched to the feeding waveguides, and show superior performance against those using dipole nanoantennas, for point-to-point optical wireless nanolink communications. A typical plasmonic horn nanoantenna is shown in Figure 13.

**Figure 13.** A typical plasmonic horn nanoantenna.

Vivaldi antennas have been also considered as an updated solution for such hybrid communications. They are usually applicable in the microwave and radio frequencies, but also in infrared/optical frequency domain. Vivaldi plasmonic antenna is formed by a slotted microstrip deposited above a silica substrate, along with a hybrid Si-plasmonic coupler, as an impedance matched element to feed silicon waveguides [43]. By increasing the number of Vivaldi antennas, an increase in the directivity and the gain is anticipated too. Hence, double Vivaldi broadside antenna [47], and antenna array configuration based on tilted plasmonic Vivaldi antennas [48], are improving the total antenna radiation performance. A double Vivaldi antenna and its coupling details are shown in Figure 14a,b respectively.

**Figure 14.** A double Vivaldi antenna and its coupling details ((**a**,**b**) respectively).

All these Vivaldi hybrid structures are characterized as SOI integrated as the optical signals could be propagated through silicon waveguides and plasmonic nanoantennas wireless links, thus avoiding integration of electronic devices and electro-optical conversions, and reducing complexity and energy costs [47]. However, efficient coupling between plasmonic antennas and SOI waveguides is a non-trivial issue, and proper plasmon based impedance matched elements are required to tackle with this issue. Other SOI integrated structures are proposed in [133], such as an antenna array consisting of a series of hybrid plasmonic nanoantennas with subwavelength footprint, that is highly compatible with a low loss silicon waveguide, which feeds light from the bottom of the nanoantenna.

Alternatively, plasmonic nanoantennas may be fed by hybrid plasmonic waveguides, instead of pure silicon waveguides, such as the plasmonic nanopatch antenna, fed by a hybrid metal insulator metal (HMIM) multilayer plasmonic waveguide, for achieving proper impedance matching, as seen in side and top views of Figure 15 [134]. Such proposed hybrid plasmonic waveguide (HPW) based nanoantennas, show high efficiency and directivity, and improve the efficiency by minimizing losses [40].

**Figure 15.** Hybrid metal insulator metal (HMIM) multilayer hybrid plasmonic nano patch antenna.

#### *5.2. THz Band Nanotransceivers*

In general, as known, a transceiver is a set of components which is responsible for generating and modulating the outgoing information to the antenna, via appropriate transmitters at the front end of the nanolink, and also responsible for detecting the incoming from the antenna information via appropriate receivers at the back end of the nanolink. A block diagram of a typical plasmonic transceiver architecture is seen in Figure 16 [102]. Apart from the basic transceiver components, namely, the signal generator, the transmitter which is normally a modulator source, and the detector-receiver, there may be other components acting complementary, in order to propagate the information signal across the link with the minimum loss, such as interconnects, switches, filters, high tunability phase shifters, mixers, frequency multipliers, and impedance matched elements, placed between the antenna and the waveguide of the link. As can be seen from Figure 16, at the front end, there is an electric signal generator that generates the electric signal that will later on, feed the transmitter source. Then the outgoing signal, which in our case of plasmonic THz sources is normally a modulated SPP wave, is feeding the plasmonic nanoantenna, which converts the SPP wave into an EM wave. At the receiver end, there is another plasmonic nanoantenna with a similar role, that inversely converts the EM wave again into an SPP wave. Finally, a plasmonic nanoreceiver converts the SPP wave into an electric signal, which will be demodulated to the original information data by the signal detector [102].

**Figure 16.** Plasmonic transceiver block diagram.

#### 5.2.1. THz Band Transmitters

For the last few years, silicon photonic devices have been widely adopted for equipping current THz band transceivers, with desired features such as large bandwidth, high transmission power and high detection sensitivity. However, high path losses at the THz band is still considered a major unsolved challenge. As far as concerns THz band transmitters, photonic III/V (InP) devices, such as UTC-PDs with mW power levels at 300 GHz [135], and quantum cascade lasers (QCLs) [136], are considered as state of the art sources. The latter, however, requires an external laser for optical electron pumping, thus limiting the area overhead, while also performing poorly at room temperature [137]. SiGe-based heterojunction bipolar transistors (HBTs) have been also proposed in [138], for equipping THz transmitters operating at 820 GHz.

QCLs at the THz band, such as plasmonic QCLs, or THz QCLs [139] with metallic cavities [140], which belong to the plasmonic laser family, may be a promising candidate for THz transmission, but researchers have to tackle with tunability potential at these high frequencies. Recently, a new tuning mechanism called the antenna feedback mechanism, has been developed for single mode metal-clad plasmonic lasers, based on the principle that the refractive index of the laser's surrounding medium affects the resonant cavity mode as much as the refractive index of gain medium inside the cavity [141]. This mechanism leads to the generation of hybrid SPPs propagating outside the cavity of the laser with a large spatial extent. The emission frequency of the plasmonic laser and its tunability potential are dependent strongly on the effective propagation index of the SPP mode in the surrounding medium, by coupling the resonant SPP mode to a highly directional far field radiation pattern, and integrating it, with the hybrid SPPs of the surrounding medium [142].

Such an antenna feedback principle is shown in Figure 17. Specifically, the general principle of a DFB can be implemented in a THz QCL by introducing periodic slits or holes in its metallic cladding. A parallel-plate metallic cavity is illustrated (Figure 17a). Due to the periodicity of conventional DFBs there is a phase-mismatch between successive apertures for SPP waves on either side of the cladding (Figure 17b). The designed grating period of antenna-feedback effectively couples a single-sided SPP wave that travels in the surrounding medium with the SPP wave traveling inside the active medium (Figure 17c). The antenna-feedback scheme leads to a buildup of in-phase condition at each aperture between counter propagating SPP waves on the either side of metal-cladding. Emission from each aperture adds up coherently and constructively to couple to far-field radiation in the z direction (Figure 17d).

**Figure 17.** Antenna feedback principle of a DFB laser.

#### 5.2.2. THz Band Receivers

As far as concerns state of the art THz band receivers, waveguide integrated detectors using GaAs Schottky barrier diodes (SBDs) [143], are considered the most common implementation choice. However, they pose limitations on their size and available bandwidth, and behave poorly at room temperature, as QCLs [144]. Silicon CMOS technology has also been proposed in [145], for equipping oscillators at 870 GHz, and sub-harmonic detectors between 790 and 960 GHz. These structures require power amplifiers implemented with InP or GaN technologies, in order to challenge path losses, such as the InP based high electron mobility transistors (HEMT) amplifiers with 10 dB gain at 640 GHz [146], or GaN, GaAs or InGaAs HEMTs [147–149]. Unfortunately, all these components require more chip space and they also pose performance limitations when operating above 1THz.

As far as concerns plasmon based THz wave detectors, hybrid plasmonic schemes consisting of plasmonic waveguides along with silicon waveguides, all integrated with conventional nanoantenna structures, such as dipole or bowtie antennas are considered to be promising, beyond state of the art approaches. A POH slot waveguide integrated with a bowtie antenna is proposed in [150], as a THz wave detector. In this structure, a taper is used to connect silicon strip waveguide with the electro-optical polymer refilled plasmonic slot waveguide, all integrated with a bowtie antenna. Plasmonic slot waveguides are capable of guiding electromagnetic waves at subwavelength scale modes, thus bypassing the diffraction limit bottleneck of conventional waveguides. Consequently, they are functioning as nanocouplers, and hence they can be used as excellent matching elements with the state of the art on chip sources such as nanoLEDs [151] and nanolasers [152], which are all belong to the quantum emitters family. Recently, plasmon based nanocouplers, such as plasmonic grating couplers [153], plasmonic waveguide tapers [154], and nanoantennas [155] have been proposed for integrating silicon photonic chip components with plasmon based components, providing a promising approach for manufacturing plasmonic integrated circuits (PICs) [156].

Alternatively, plasmonic internal photoemission detectors (PIPED) structures, which are actually transmitter and receiver packages, monolithically integrated on a common silicon photonic chip platform, have been also proposed in [157], for THz wave signal generation and coherent detection at frequencies of up to 1 THz. Finally, other more novel approaches are related to the employment of a FET operating at THz frequencies (TeraFET), as a THz detector. Specifically in [158], a TeraFET operation with identical radiation amplitudes at the source and drain antennas but with a phase shift induced asymmetry is proposed, based on the principle that a phase difference between THz signals coupled to the gate and source, and gate and drain terminals of a FET accordingly, enhances device plasmonic resonances. Such a TeraFET structure operates between 200–600 GHz band, and could be used for 5G and beyond, communication systems.

#### 5.2.3. Graphene Based THz Transceiver Components

Apart from graphennas, graphene has been also considered as a promising technology for building THz band transceivers operating at 1–10THz band and even at higher frequencies. In fact, graphene may be used for implementing not only the basic transceiver components, namely sources and detectors, but also for other components such as interconnects, modulators, switches, filters, and phase shifters with high degrees of tunability [101]. Moreover, by applying the same graphene nanomaterial for all these THz transceiver components, the integration potentials of packaging compact THz band nanotransceivers can be highly boosted. Specifically, all graphene THz components can be combined with arrayed THz graphene antennas, so as to enhance total antenna directivity [125]. Evidently, the adoption of the same building material is expected to minimize mismatches, and consequently, the losses between the interconnected components. Many graphene-based components have been proposed for feeding THz link nanoantennas, such as plasmonic switch based on GFET [159], THz LPF [160], BPF [77], and phase-shifters [161]. All these feeding components are based on the same GFET principle, being modeled as a tunable transmission line for the propagation of SPPs, with adjustable operation at different lengths and bias levels each time, depending on the application.

Graphene can be efficiently applied for the implementation of HEMT used as THz source. Specifically, in [102], the proposed THz source is based on a III-V semiconductor based HEMT, enhanced with grapheme, so as to generate the necessary SPP waves that drive a plasmonic nanoantenna with satisfactory impedance matching, thus enabling compatibility between all participating plasmonic THz interconnection components. Moreover, as graphene is a material characterized by ultrafast carrier relaxation recombination dynamics, low pumping threshold level, and wide THz tuning range at room temperatures, it appears to be a promising technology for implementing new types of laser sources, against current semiconductor lasers [162]. It may be used as a THz modulator as well. Specifically, in [163], a single layer graphene THz modulator is proposed, based on a monolayer graphene sheet, lying on a SiO2/p-Si substrate and biased with metal gates. A voltage applied between the gates, injects carriers on graphene, thus modulating the chemical potential at THz rates. Other modulators, based on the same principles, have been proposed at these rates [164–166], as well as modulators at infrared frequencies [167,168]. Finally, another graphene based plasmonic waveguide phase modulator proposed in [169], is also based on the principle of electronically control the propagation speed of an SPP wave, by modifying the chemical potential of the graphene layer. Last, GFETs operating at room temperatures, have been also proposed and demonstrated as ultrafast THz detectors [170]. A graphene-based THz detector based on a log-periodic circular toothed antenna between the source and the gate of a GFET, is proposed in [171]. The THz oscillating electric field is fed between the gate and the channel of the GFET, inducing a DC signal between source and drain, that is proportional to the received optical power. Finally, a compact graphene slot photodetector on SOI with high responsivity, is proposed in [172].

Apart from graphene other materials, such as black phosphorus (BP) has been characterized by attractive features, such as the high carrier mobility, the in-plane anisotropy, and the tuning capability via electrical gating, 0.3 to 1.7 eV. It can be exploited as a perfect THz detector when being integrated in a microscopic FET for a wide range of THz frequencies from 0.26 THz to 3.4 THz. Such a BP-based FET can be used as a plasma-wave rectifier, a thermoelectric sensor, or a thermal bolometer [173]. Chalcogenides compounds containing at least one of the chalcogen elements, namely, sulfur, selenium or tellurium, and specifically, chalcogenide glasses are characterized as phase-change active plasmonic devices appropriate for active plasmonic switching/modulation functionality for future nanophotonic device applications. Experimental demonstrations employing gallium anthanum sulfide as a photo-active medium, implement that CMOS/SOI-compatible, chalcogenide glasses used for such functionalities [174]. They may be also used as topological insulators for protecting metallic surfaces of plasmonic structures, as they are immune to scattering from disorder and defects, they can be dynamically controlled via external electric, magnetic or optical excitations, and their optoelectronic response is highly sensitive to the polarization state of incident light.

#### **6. Summary and Conclusions**

The THz band has been characterized as the last undiscovered frontier of the total EM spectra range that urges for exploration and investigation, since current data traffic and bandwidth hungry applications will no longer satisfy their speed and latency demands with existing technologies and system architectures. On the other hand, wireless communications seem to be in advance against conventional wired communications. Therefore, the migration to higher carrier frequency bands and specifically in the THz band is required, via adoption of new technologies, equipping future THz wireless communication systems at the nanoscale, in order to accommodate a variety of applications that would satisfy the ever increasing user demands for higher data rates. Hence, wireless THz band communications and modern THz wireless nanoscale applications, such as beyond 5G communications, NoC system architectures and WNSNs, are still urging for an efficient, compact and standardized interconnect solution for generating, transmitting, propagating, and detecting the THz wave information.

In this paper, a comprehensive survey has been presented for THz wireless communications and applications, as an attempt to identify unsolved issues and challenges in THz region, such as the very high propagation signal loss, the impedance mismatch between THz link components, limited size restrictions along with integration potentials, associated with high bandwidth availability and ultra-fast operating data rates and minimum latency requirements. Pointing to this direction, the most efficient compact technology, or the hybrid combination of competitive technologies, such as conventional CMOS electronics (photonic and plasmonic), is under investigation, in order to properly equip future THz nanoscale communication systems, hosting modern wireless THz nanoapplications. Among competitive technologies, CMOS-based electronic interconnects are definitely out of the competition, in order to meet THz speed, low propagation signal loss, and the impedance match between THz link components. A photonic solution is indeed; a viable approach for providing high data rates at low propagation losses, still the component size is one with two orders of magnitude larger than what required for THz band case. Plasmon based THz link components, on the other hand, due to their extremely small size and their ability to operate at ultra-high data rates, seems to be a promising approach for equipping wireless THz nanoscale communication systems. Moreover, they could be perfectly combined with photonic technology and particularly with dielectric waveguiding, as plasmonic waveguiding is quite lossy concerning relatively long interconnect distances. Therefore, this survey work, has provided in-depth reference material of the current fundamental aspects of plasmonic technology and hybrid combinations, highlighting plasmon future roles in THz band wireless communication. It is a thorough investigation on current and beyond state-of-the-art plasmonic layouts, implementing THz nanoscale communication systems, and wireless THz nanoapplications accordingly.

Specifically, as far as it concerns NoC architectures, many alternative technologies have been progressively proposed, in order to deal effectively with the NoC communication bottleneck, such as 3D, RF signals over on-chip transmission lines, FSO systems at IR, and photonic and nanophotonic NoC—all these considered to be state of the art technologies, as seen in Table 1. THz band communications however, anticipate for smaller footprints of the transceiver and the antenna for more efficient integration, and for larger available transmission bandwidths, and higher achievable data rates. Given these tight requests, a promising, beyond state-of-the-art solution would be based on reduced size plasmon nanoantennas and other plasmon based THz transceiver components as well, so as to operate at very high resonant THz frequencies. This can be achieved, as it seems, mostly via the use of graphene material supporting graphene based WiNoCs architectures, or alternatively, based on the combination of plasmonic resonators with dielectric waveguiding, supporting wireless-optical on chip communication, and hybrid optical-wireless NoC architectures respectively (Table 1).


**Table 1.** SoA and beyond SoA technologies for wireless THz applications.

\*SoA—state-of-the-art.

GNR and CNT based nanosensors, nanoprocessors, nanoantennas, and nanotransceivers, or other material nanosensors such as Au and Ag plasmon sensors, and nanoscale energy harvesting systems, are considered to be the state of the art for implementing todays WNSNs. Basically, in the near future, as far as concerns the communication unit of a WNSN system, it is likely that graphene nanoantennas and other graphene-based THz transceiver components would be required, so as the system to operate at these high resonant THz frequencies (Table 1). At the moment, beyond 5G communication systems are mostly equipped via silicon photonic technology, and particularly via Si photonics-based UTC-PDs, and comb sources operating at mmW. Many research attempts have been proposed in order such systems to operate, well above 300GHz frequencies, at the THz band. These approaches include FSO systems with THz/O links, IMWP in the THz range, graphene MIMO antennas structures, and plasmonic MZ modulators with sub-THz frequency responses, or ultra-broadband silicon POH modulators, required for THz/O signal conversions (Table 1).

Consequently, plasmonics play a critical role for equipping each single, individual component part of a future wireless THz nanocommunication link, namely the antennas and the transceiver parts, as also can be seen in more detail in Table 2. As far as concerns antenna implementation, silicon integrated antennas, CNT based antennas, UWB and multi-band antennas, have been considered as mature options for on chip antennas, but the tight uncompromising limitation of a very small size of a THz band antenna requires other technologies, such as plasmon or hybrid combinations. Hence, graphene based nanoantennas (patch antenna, dipole, MIMO), or hybrid graphene-dielectric antennas(H2G), or plasmonic antennas with dielectric waveguides feeding elements (dipole loop, horn, Vivaldi, hybrid plasmon-dielectric array), or inversely, plasmonic nanopatch antenna, with HMIM plasmonic waveguide as feeding element, are strong and promising candidates for implementing THz nanoantennas, as also seen in Table 2.

As far as concerns other THz transceiver component parts, silicon photonic devices have been widely adopted for equipping current THz band transceivers, with photonic III/V (InP) devices, such as UTC-PDs and QCLs to be considered as state-of-the-art sources. SiGe-based HBTs have been also proposed for equipping THz transmitters. However, high path losses of these technologies, at the THz band is still considered a major unsolved challenge. THz plasmonic lasers such as plasmonic QCLs, THz QCLs with metallic cavities, and single mode metal-clad plasmonic lasers with their antenna feedback tuning mechanism, may be a promising candidate for THz transmission sources. Hybrid transmitters, such as III-V semiconductor based HEMT, enhanced with graphene, may be as

well considered to be a promising THz transmitter approach. Moreover, plasmonic slot waveguides may be used as a perfect nanocoupler matching element with current state of the art on chip sources (nanoLEDs, nanolasers) (Table 2).


\*SoA—state-of-the-art.

Lastly, as far as concerns state of the art THz band receivers, waveguide integrated detectors using GaAs SBDs are considered as the most common implementation choice. However, they pose limitations on their size and available bandwidth, and behave poorly at room temperature. Si-CMOS technology with HEMTs has also been proposed for equipping detectors between 790 and 960 GHz, as they pose performance limitations above 1THz. Hybrid plasmon structures, such as POH slot waveguide integrated with a bowtie antenna, or graphene slot photodetector on SOI, or GFET based structures such as plasmonic teraFET, are considered to be promising, beyond state of the art, THz wave detectors (Table 2). GFET based structures can be also used for equipping other THz transceiver parts, such as switches, LPFs, BPFs, phase shifters and modulators. Alternatively, PIPED structures, which are actually end to end, transmitter and receiver packages, monolithically integrated on a common silicon photonic chip platform, have been also proposed for THz wave signal generation and coherent detection, at frequencies of up to 1 THz (Table 2).

Tables 1 and 2 include, in summary, all state of the art and beyond state of the art, plasmon based technology, exploited for the implementation of future THz band nanocommunication systems, as have been exhaustively presented in this work. Apparently it seems that, in order to fully equip future THz nanocommunication applications, miniature size transceiver components are required, with uncompromising features such as low propagation signal loss, impedance matching between link components, strong integration potentials and compatibility with ancestor technologies as CMOS, with high bandwidth availability and ultra-fast operating data rates and minimum latency. This comprehensive survey paper has highlighted such an objective, by qualitatively presenting in such detail, the latest plasmonic and accompanied photonics technologies on equipping future competitive THz nanoscale communication systems, hosting wireless THz nanoapplications, namely NoCs, WNSNs and beyond 5G communications, providing at the same time, motivation for research academia to seek for efficient solutions towards this direction.

**Conflicts of Interest:** The author declares no conflict of interest.

#### **References**


© 2019 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Real-Time Train Tracking from Distributed Acoustic Sensing Data**

**Christoph Wiesmeyr 1,\*, Martin Litzenberger 1, Markus Waser 1, Adam Papp 1, Heinrich Garn 1, Günther Neunteufel <sup>2</sup> and Herbert Döller <sup>2</sup>**


Received: 31 October 2019; Accepted: 23 December 2019; Published: 8 January 2020

**Abstract:** In the context of railway safety, it is crucial to know the positions of all trains moving along the infrastructure. In this contribution, we present an algorithm that extracts the positions of moving trains for a given point in time from Distributed Acoustic Sensing (DAS) signals. These signals are obtained by injecting light pulses into an optical fiber close to the railway tracks and measuring the Rayleigh backscatter. We show that the vibrations of moving objects can be identified and tracked in real-time yielding train positions every second. To speed up the algorithm, we describe how the calculations can partly be based on graphical processing units. The tracking quality is assessed by counting the inaccurate and lost train tracks for two different types of cable installations.

**Keywords:** DAS; fiber optic sensing; train tracking; pattern recognition

#### **1. Introduction**

Railway safety is an ever-increasing issue, as traffic demand increases world wide with railways playing an important role. This is especially true in the light of the urgent need for the decarbonization of traffic. The accurate tracking of a train's real position on the track is the basis for all modern railway safety concepts. In state-of-the-art rail operations, the train's position is only known within a so-called "block" between two signals. In one block, there is only one train allowed to operate at a time, which is sufficient to prevent collisions of trains. The disadvantage of this concept is a reduced efficiency in the use of the available track infrastructure because long portions of the track stay unused. Future traffic demand calls for more flexible, thus efficient, concepts. Novel safety concepts will, therefore, be based on virtual- or moving-blocks that virtually enclose the train while in motion, providing enough safety separation between trains at all times [1,2]. An accurate, reliable, and redundant train tracking technology is an important basis for these novel concepts.

Conventional train tracking is mainly achieved by communication between train-side equipment, track-side equipment, and the interlocking system. Block-based systems need axle-counting sensors installed next to the track. They deliver a signal for each axle that passes through the detector, to ensure that a block is unoccupied before the next train enters this block. The European train safety system, ETCS, can use moving blocks in its level 3 implementation and relies on continuous positioning information of the train [3]. High accuracy differential satellite positioning systems (D-GNSS), such as GPS, Glonass, or Galileo, become available and are certainly planned to be applied in future railway systems [4], but the high safety requirements of railway operation call for highly reliable and redundant positioning systems. Furthermore, the coverage of D-GNSS and the necessary communication link back to the

interlocking system is never 100% guaranteed: Tunnels, deep valleys, and tall metallic structures next to the track can degrade, or even block, radio frequency signals.

Non-coherent optical time domain reflectometry (OTDR) technology has been used for a long time for long-range monitoring of the quality and the integrity of fiber optic cable infrastructure [5], e.g., to locate break points in underwater fiber cables. Distributed acoustic sensing (DAS), a special variant of coherent (or phase sensitive) OTDR, is not only able to assess the cable quality but also to detect and locate strain and temperature changes along the cable with high sensitivity. The measurement and signal processing methods presented in this paper are based on DAS and offer an alternative or redundancy system for accurate train positioning and tracking. Other than the aforementioned technologies, it does not need equipment installed on the train nor communication between train and interlocking system. Furthermore, it does not need extra track-side equipment, with the exception of a fiber optic cable that is already in place for data connection and communication purposes. In this paper, we investigate advanced signal processing of the DAS data received from the so-called optical interrogator device for accurate train tracking. The interrogator injects a series of laser light pulses into the fiber cable and measures the back scattered (Rayleigh backscatter) light at the same end of the cable. The measured signal contains the optical path length change resulting from a refractive index change due to compression of the glass and from contraction or elongation of the fiber due to ground deformation [6]. These effects can be received from any point along the cable over a range of up to 40 km and can be measured with a positional resolution down to 0.5 m. The reader interested in the different types of DAS devices and their physical principles is referred to the survey of Bao and Chen [7].

OTDR and DAS are used, for example, for oil well monitoring [8], intrusion detection [9,10], or even for rock slide detection [11]. Vibrations of train movements near the railway track can also be monitored using a DAS device, where in most cases, fiber cables are already installed there for communication means. Several methods for train tracking using DAS signals have been proposed [12–14]. However, these algorithms were never evaluated for longer recordings and for different ground and installation conditions. All the presented methods are based on the variance of the measured DAS signals. In this contribution, we propose a method that is based on machine learning methods to detect vibrations, which is more flexible and can be more easily adapted to different recording conditions. Furthermore, the algorithm which is the basis for [15,16] has a five-second lag in time. The goal of this manuscript is to provide a detailed description and evaluation of an algorithm that can be used for real-time train position monitoring.

We will give the details on the data acquisition of DAS test data used for this investigation in the following section. The algorithmic details, as well as a discussion of the accuracy of the proposed method will be given in the later sections. We will further provide an outlook for future work and improvements.

#### **2. Materials and Methods**

In this section, we will present the signal processing methods used for train tracking, i.e., we will describe in detail how to process the raw measurement data to obtain train trajectories. The train tracking algorithm works on the basis of one-second chunks of data and is based on two main steps, i.e., vibration detection and object tracking. Before we go into details on these, we will give a short introduction on how to interpret the raw DAS signal from the Fotech Helios DAS (https://www.fotech. com/products/helios-das/ retrieved 31 October 2019)) interrogator device.

#### *2.1. DAS Test Data Acquisition*

We recorded long-term DAS data with a Fotech Helios DAS interrogator device in two different railway test sites. At one site, the track was located in a tunnel, with the single-mode fiber optic cable installed in a cable trench; in the other site, there was an open track with the single-mode fiber cable directly attached to one rails' foot with clips. For all experiments, already existing telecommunication fiber cables were used, and no special installation of cables was done for this work. At these two sites, we recorded a total of approximately 1000 h of data with 3000 trains passing the infrastructure. The total rail track length monitored during the recordings was approximately 20 km. Refer to Section 3.3 for details on the sites. Apart from these long-term recordings, two shorter recordings were available, of around one hour each, from an open track with the cable in a cable trench. The standard telecommunication fibers were used for the recordings, and the effect of fiber darkening over time was not evaluated in this study.

The Helios DAS is a coherent (C-OTDR) phase sensitive interrogator device that delivers the optical detector voltage, encoding the laser light phase change, as a 16 bit resolution signal, sampled with 150 MSamples/s. This results in a physical spacing of the DAS segments of 0.68 m, given by the light speed in the fiber. The effective spatial resolution of the measurement is limited by the laser pulse length of 100 ns, resulting in a size of the laser pulse of 20.2 m, over which the phase measurement is averaged. Table 1 summarizes the relevant measurement parameters. The cable used for test site 1 is a single-mode, stranded mini cable 60 × 5 × 12 E9/126, the cable at the second test site is a single-mode stranded fiber optic cable 5 × 12 E9/125 A-DF(ZN)2Y(BN)2Yv 5 × 12 E9/125-G652D.

Due to the nature of the optical measurement principle, neither the absolute phase nor the fiber strain is directly accessible from the signal output. Furthermore the signal is delivered with an ambiguity of 2*π* (equal to the "fringes" observed in a Michelson interferometer), which results in a non-linearity in very strong signals. A detailed description of the signal generation for a similar DAS setup is found in [9]. The laser pulse repetition frequency of the Helios DAS device was set to 2000 Hz for all recordings. This data stream is the input to our signal processing stages.


**Table 1.** Summarized parameters of measurement, optical fiber, and laser pulses used for this work. ADC = Analog-to-Digital Converter.

#### *2.2. Annotations*

For all the different test sites, we annotated vibrations and background in the raw data. In the subsequent steps, we describe how these annotations are used to automatically detect vibrations in the DAS signal. An annotation for a test site contains 10,000 samples for background and vibration signal each, where one sample consists of the raw signal recorded in one second for one cable segment.

#### *2.3. Vibration Detection*

The vibration detection takes as input the raw measurements acquired within one second for all the cable segments along the optical fiber. Given a sampling rate of 2000 Hz and a spatial resolution of 0.68 m with a monitored fiber length of 40 km (corresponding to around 60,000 channels), the raw data consist of 2000 × 60,000 samples each second. In the following, we will denote the data matrix containing all the raw measurement data for second *i* as *Mi*. The result of the vibration detection for this second in time, *vi*, is a vector of Boolean values of length 60,000, where each element of this vector indicates whether vibration is present at the second *i* or not for that specific cable segment. Therefore, the vibration detection step corresponds to a data reduction from *Mi* to *vi*.

In the following, we will describe the steps we used to perform vibration detection. We use machine learning techniques to decide whether the vector *Mi*,*<sup>x</sup>* containing the sensing samples for a given cable segment *x* and a given second in time *i* contains vibration or not. The decision is based on the spectral distribution of the raw signal *Mi*,*<sup>x</sup>* and is done based on the following steps detailed below. Note that the decision has to be made independently for all the cable segments without any spatial averaging; in the following, therefore, we will describe the procedure for one cable channel.


**Computation of spectral energy.** The energy of the frequencies of the signal *Mi*,*<sup>x</sup>* are estimated using the Fast Fourier Transform (FFT). The squared absolute values of the Fourier transformed signal |*Mi*,*x*(*ξ*)| <sup>2</sup> are then summed up individually for ten bins

*bk* = ∑ *ξ*∈*Bk* |*Mi*,*x*(*ξ*)| 2, (1)

where *Bk* denotes the *k*-th bin with *k* ∈ {0, ... , 9}. The frequency bins are linearly spaced and span the frequencies between 10 Hz and 990 Hz.

**PCA computation.** We use PCA to reduce the number of features from the frequency binning from 10 to 2. This is done by computing two linear combinations of the 10 original features optimally with respect to the variance of the data they can explain. Computationally, this PCA can be represented as a 10 × 2 matrix. The data used for computing the principal components are from the manually annotated test dataset described in Section 2.2.

**SVM classification.** SVM classification is used for each cable segment to categorize it as vibration or background for each second in time. The SVM classifier is trained on manually annotated data. We use the SVM implementation of openCV 2.4 with an RBFkernel [17].

#### *2.4. Train Tracking*

The train tracking algorithm works on the inputs from the vibration detection and works on a onesecond basis. It takes the classification vector *vi* and outputs the current positions of the trains present in the signal denoted by *Pj*, where the index *j* runs through the active trains at the given second *i*. The train tracking algorithm itself consists of three steps, which we will discuss individually subsequently:


**Edge Detection.** The edge detection is based on a K-means clustering of wavelet responses of the vibration detection vector *vi*. We use Ricker wavelets of two different widths, namely 128 and 256, cable segments to compute the convolution with the detection vector *vi*. From both of these convolutions, we extract 512 consecutive values as feature values for the K-means clustering algorithm. A prototypical dataset of vibration detections is used to train the algorithm with four centers which correspond to background, leading edges, trailing edges, and train signal. Note that at this point leading edges are edges that mark the start of a train viewed from the interrogator device, while trailing edges are edges that mark the end of a train. If the train moves towards the interrogator, this leading edge will also be the leading edge of the train; if the train moves away from the interrogator, the leading edge viewed from the interrogator will be the trailing edge of the train. In the following, we will call all these edges *detected edges*.

**Edge Assignment and object creation.** The tracking algorithm works internally with two types of objects: tracked edges and tracked trains, which we will denote by *Ei*, *i* ∈ *I* and *Pj*, *j* ∈ *J*, respectively. Note that tracked edges are different objects from the detected edges identified in the previous step. In the following, we will describe how the detected edges are assigned to the tracked trains and the tracked edges. This is repeated every second with the corresponding detected edges. Therefore, all the

steps described in this paragraph correspond to one given second in time. The algorithm starts to assign detected edges to the list of tracked trains; this is done for detected leading and trailing edges separately. We start by defining cost functions *cL* and *cT* for a given assignment of an edge *e* and a Train *P* by the distance between the edge and the train's leading edge if *e* is itself a leading edge and the distance between the edge *e* and the train's trailing edge if *e* itself is a trailing edge. Furthermore, we define a threshold *T*, which denotes the maximum distance between *e* and the corresponding edges of the trains, where we still allow an assignment to be made. The formulas for these two cost functions read as

$$c\_L(e, P) = \begin{cases} |e - P\_{\text{Lead}}| & \text{if } |e - P\_{\text{Lead}}| < T \\ \infty & \text{if } |e - P\_{\text{Lead}}| > T \end{cases} \tag{2}$$

$$\mathcal{c}\_{T}(\mathcal{e}, P) = \begin{cases} |\mathcal{e} - P\_{\text{Trail}}| & \text{if } |\mathcal{e} - P\_{\text{Trail}}| < T \\ \infty & \text{if } |\mathcal{e} - P\_{\text{Trail}}| > T. \end{cases} \tag{3}$$

Therefore, the two cost functions equal infinity if the detected edge is too far away from the corresponding edge of a given train. If the cost function *cL*(*e*, *P*) = ∞, the edge *e* will not be assigned to the leading edge of the train *P* (this is analogous for the function *cT* and the trailing edge of *P*). The assignment problem for a given set of detected leading edges (the assignment works analogously for trailing edges) *ei*, *i* ∈ *Ie* to a given set of trains *Pj*, *j* ∈ *Ij* iteratively finds the minimum in the matrix *Ci*,*<sup>j</sup>* = *cL*(*ei*, *Pj*) for *i*, *j* in the defined index sets. The leading edge and train with the respective indices *i*<sup>0</sup> and *j*<sup>0</sup> corresponding to this minimum, if smaller than ∞, are assigned to each other and the *i*0-th row and the *j*0-th column of the cost matrix *C* are set to ∞. This step is repeated until no more assignments can be made, i.e., the matrix *C* contains only ∞. The same procedure is then also repeated for the trailing edges of the trains.

After the assignment of leading and trailing edges to the trains, a subset of the originally detected edges has been assigned. The remaining detected edges are then assigned to the currently tracked edges *Ei* using the same iterative procedure as described for the assignment of edges to trains. We define distance functions analogous to Equation (2), (3), which leads to a definition to an assignment matrix *C*. With the same greedy iterative algorithm as described above, we can determine the assignments of found edges to tracked edges.

A (possibly empty) subset of the found edges cannot be assigned to a train or to a tracked edge. These edges will become tracked edges in the next time step.

The assignment process described in this section is an approximate greedy approach to the more complex quadratic assignment problem where one tries to find all the assignment index pairs (*i*, *j*) such that the corresponding total cost (i.e., the sum of the individual costs for assignments) is minimal. For better readability of the paper, we will not give a rigorous formulation of this optimization problem here since it is computationally not feasible for a real time tracking algorithm.

**Kalman filtering.** Each train and each tracked edge has an underlying Kalman filter. The tracked edges follow a two-dimensional state-space model consisting of position of the edge and its velocity leading to the following matrices corresponding to a constant velocity model

$$F = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \\ \tag{4}$$
  $H = (1, 0),$ 

where *F* denotes the state transition model, and *H* denotes the observation model.

The Kalman filter for a tracked train is based on a slightly more involved state-space model. To model a moving train, we propose a four-dimensional state space consisting of leading edge position, trailing edge position, speed, and the length of the train. A measurement at a given time

results in the leading and the trailing edge of the train. The resulting constant velocity model leads to the following transition and observation model

$$F = \begin{pmatrix} 1 & 0 & 1 & 0 \\ 0 & 1 & 1 & 0 \\ 0 & 0 & 1 & 0 \\ -1 & 1 & 0 & 0 \end{pmatrix} \\ \tag{5}$$
  $H = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{pmatrix}.$ 

#### *2.5. GPU Computation for Feature Extraction*

The computation of the features for the SVM classification needs to be able to process the raw data matrix *Mi* containing the data for second *i* in time and all the channels along the optical fiber. As mentioned above, for a typical setup of 40 km monitoring length and a sampling rate of 2000 Hz this leads to a data matrix of dimensions in the order of 2000 × 60,000. In the case of a Helios DAS interrogator device, the data format is unsigned int 16, which leads to a data rate of 228 MB/s. We then have to cast these data to a float datatype because the first step involves a Fourier transform that doubles the data rate as we use 32 bit float values. This Fourier transform computation has to be done for all 60,000 cable segments in parallel. The highly parallelizable nature of these computations makes general purpose GPUs (GPGPU) highly suitable for the task. The same holds true for the PCA feature reduction, which has to be applied for each cable segment individually. The SVM classifier is not available with GPU acceleration; therefore, the reduced features are downloaded to the CPU before classification. A flow chart of the complete algorithm is depicted in Figure 1.

**Figure 1.** Flow chart of the algorithm. The top box describes the steps of the vibration detection (steps on the GPU are in blue); the bottom box describes the steps of the train tracking. FFT = Fast Fourier Transform; PCA = Principal Component Analysis; SVM = Support Vector Machine.

#### **3. Results**

This section is devoted to the results. We will start to discuss the results for the different stages of the tracking algorithm and then present the results for two long term recordings, as well as tracking accuracy results for a short-term recording where a ground truth of the train positions was available.

#### *3.1. Train Tracking Stages*

To perform the vibration detection, we first annotated vibrations in a recorded raw data file individually for the two different track conditions. These ground truth data are then used to train the PCA for feature reduction and the SVM for classification. An illustration of the vibration detection can be found in Figure 2.

**Figure 2. Left**: Raw subsampled signal for visualization. **Right**: Result of the vibration detection where white pixels indicate vibration and black pixels indicate background. The gaps in the train tracks occur when a train stops since an electric train that is not moving does not emit any vibrations and is, therefore, not visible usind a Distributed Acoustic Sensing (DAS) system.

The first step of the tracking algorithm is the edge detection. In Figure 3, we give an example of the result of the edge detection from the first 35 s of the dataset shown in Figure 2. After the edge detection, the assignment and Kalman filtering are performed. For the same example dataset, we show the tracking result in Figure 4.

**Figure 3.** Leading and trailing edge of the object in blue and orange, respectively, plotted over the vibration detection results in black and white; note that the depicted signal consists of the first 35 s of the signal shown in Figure 2.

**Figure 4.** Train tracking result for an example dataset (colored lines), detection result in black and white; note that the depicted signal is the same as in Figure 2.

#### *3.2. Evaluation of Tracking Accuracy*

For two datasets encompassing around 3000 s of recording each, we have available ground truth positional data for two points along the track. The recording was done on an open track with the cable installed in a cable trench. The available ground truth is accurate to one second in time, which leads to an accuracy of around 30 to 40 m considering that this is the speed the objects move on the test infrastructure. The tracking errors in meters can be found in Table 2. The tracking error is in the range of 40 m, which is in the range of the available ground truth considering that both, the ground truth and the tracking algorithm, work on a one second basis that leads to a maximally two-second difference between the two systems, and the trains move at a speed of 30–40 m/s.

**Table 2.** Summarized performance of tracking algorithm for the two two short term recordings for single trains. Given that the ground truth and the tracking are both accurate down to one second, the evaluation is accurate down to 2 s. Considering the train speed, which is around 30–40 m/s, the errors are in the expected range.


#### *3.3. Evaluation of Tracking Reliability*

In this section, we will evaluate the tracking reliability based on two long term recording results. One is done in a tunnel where only one tube in one direction has been measured. The second test site is a standard track with the cable mounted directly on the rail. To evaluate the number of correctly tracked trains without the availability of ground truth from an alternative tracking system, the following steps were taken:


We collect the respective numbers of the correct and lost train tracks for two different test sites, where we were conducting long term measurements in Table 3. In the following subsections, we will describe the specific results for the two test sites.


**Table 3.** Summarized performance of tracking algorithm for the two monitored test sites.

#### 3.3.1. Test Site 1: Tunnel

The monitored section of the cable in the tunnel is from 13,600 m to 25,900 m relative distance from the interrogator device; therefore, we observe a long stretch of cable. This analysis shows that we reach 98% of correct tracks, respectively, for the two evaluated cable lengths. In this test site, no artifact tracks due to non-train-related noise was observed, meaning that the two percent incorrect tracks are attributed to lost tracks.

#### 3.3.2. Test Site 2: Open track

This test site consists of an open track with a cable that is directly mounted on the rail, which leads to a very different scenario from the first test site. In our test period, we were not observing any lost tracks of moving trains on the open track. However, due to the cable installation method, we observed a higher amount of noise in the signal, which leads to a number of false positive detections of trains during the evaluation period. Of the 1071 total tracks, we found 12 to be artifacts, which leads to a total of 99% correct tracks. Generally, mounting the optical cable directly on the rail leads to more noise in the signal that is not caused by trains. This is not surprising since such a cable is exposed to natural elements, such as wind and rain. Furthermore, we observed that certain sections of the cable seem to be coupled, meaning that if a train enters such a section, all the cable segments in that section will vibrate. An illustration of this phenomenon, as well as of the increased noise in the signal, can be found in Figure 5. The coupling of fiber segments leads to decreased accuracy in estimating the train length because the detected leading and trailing edge do not coincide with the start and end of the train but, rather, with the start and end of the coupled segments.

**Figure 5. Top Left**: Subsampled signal through band energy calculation. **Top Right**: Train Detection result for a train on test site 2; the green boxes indicate coupled cable segments; the background without train shows increased noise. **Bottom**: False positive train track due to false positive vibration detections.

#### 3.3.3. Comparison with the Literature

The goal of this section is to discuss the presented results in comparison to the approaches discussed in the literature. The main difference between the algorithms described in the literature and our approach is the flexibility of our method through the use of machine learning.

The papers of Timofeev et al. [12,13] describe an energy based train vibration detection. According to our experience, this approach is not feasible for long stretches of monitored fiber with the recording device we used. The paper does not mention the validation data used for drawing the conclusions, nor does it mention a tracking algorithm. The accuracy of the proposed method with respect to positional accuracy is reported at 15 m. In comparison, we did not have a ground truth available with such a high accuracy.

The paper of Peng et al. [14] describes an algorithm which uses a relative variance for vibration detection. The validation data for the method are two trains recorded over 400 s. The paper does not mention any accuracy measurements for the proposed method, nor do the authors mention which tracking algorithm is used.

The earlier papers of some of the authors of this contribution [15,16] describe a similar algorithm with some important difference. The earlier versions of the algorithm used a filtering over time which introduced a delay. This has been changed in this contribution to Kalman filtering avoiding a time lag of the method. Furthermore, the edge detection has been refined to be more flexible, which is especially important for rail foot cables.

We can conclude that this is, to our knowledge, the most extensive study done on railway monitoring using DAS. In comparison to the other literature, we used flexible methods that allow for efficient re-calibration for different track conditions.

#### **4. Discussion**

In this paper, we presented a real-time train tracking algorithm that runs on the basis of one-second signals without delay. The algorithm is based on two main steps, i.e., the detection of vibrations and the tracking of trains in the signal. The performance of the algorithm was evaluated on two test sites, where one was in a tunnel with a standard cable trench and the other one was on an open track with the cable attached directly to the rail. In the tunnel, we measured a long distance which lead to occasionally lost tracks, especially on a stretch where the cable installation was sub-optimal for transferring vibrations from the train to the cable. On the open track with the cable mounted directly to the rail, we did not observe lost tracks during the test period, which is due to strong signals, because the observed cable length was considerably shorter than in the tunnel. Due to the increased noise with the cable installation method, we observed several artifact train tracks. Nevertheless, for both evaluated tracks we reached accurate tracks in more than 98% of the cases. We conclude from the results that the installation in a cable trench is advantageous for train tracking with a DAS system.

For two shorter recordings, we presented the positional error, which was around 40 m, which was also in the range of the available ground truth. For more accurate evaluations of the positional accuracy of the tracking, it would be important to get better ground truth data and perform a highly accurate calibration of the fiber cable.

For application in the railway sector, DAS shows great potential, especially when combined with other sensors for redundancy. In future research, we will evaluate methods for fusing different sensor modalities to increase the robustness of the tracking.

**Author Contributions:** The authors from NBG Fosa GmbH recorded the data. This involved the planning and calibration of the test sites, as well as understanding the signal behavior in different positions. The authors from AIT Austrian Institute of Technology GmbH developed the tracking algorithms and their implementation. Conceptualization, C.W. and A.P.; Data curation, G.N. and H.D.; Methodology, C.W.; Project administration, M.L. and H.G.; Software, A.P.; Supervision, M.L. and H.G.; Visualization, M.W.; Writing—original draft, C.W.; Writing—review & editing, M.L. and M.W. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was partially funded by the "Wiener Wirtschaftsgantur" through the research project "FOS—Real Time Methods" with the grand ID 1890985

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

1. Gao, S.; Dong, H.; Ning, B.; Zhang, Q. Cooperative Prescribed Performance Tracking Control for Multiple High-Speed Trains in Moving Block Signaling System. *IEEE Trans. Intell. Transp. Syst.* **2019**, *20*, 2740–2749. [CrossRef]


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Article*
