*Article* **Collaborative Robotic Wire + Arc Additive Manufacture and Sensor-Enabled In-Process Ultrasonic Non-Destructive Evaluation**

**Rastislav Zimermann 1,\*, Ehsan Mohseni 1, Momchil Vasilev 1, Charalampos Loukas 1, Randika K. W. Vithanage 1, Charles N. Macleod 1, David Lines 1, Yashar Javadi 1, Misael Pimentel Espirindio E Silva 2, Stephen Fitzpatrick 2, Steven Halavage 2, Scott Mckegney 2, Stephen Gareth Pierce 3, Stewart Williams <sup>3</sup> and Jialuo Ding <sup>3</sup>**


**Abstract:** The demand for cost-efficient manufacturing of complex metal components has driven research for metal Additive Manufacturing (AM) such as Wire + Arc Additive Manufacturing (WAAM). WAAM enables automated, time- and material-efficient manufacturing of metal parts. To strengthen these benefits, the demand for robotically deployed in-process Non-Destructive Evaluation (NDE) has risen, aiming to replace current manually deployed inspection techniques after completion of the part. This work presents a synchronized multi-robot WAAM and NDE cell aiming to achieve (1) defect detection in-process, (2) enable possible in-process repair and (3) prevent costly scrappage or rework of completed defective builds. The deployment of the NDE during a deposition process is achieved through real-time position control of robots based on sensor input. A novel high-temperature capable, dry-coupled phased array ultrasound transducer (PAUT) roller-probe device is used for the NDE inspection. The dry-coupled sensor is tailored for coupling with an as-built high-temperature WAAM surface at an applied force and speed. The demonstration of the novel ultrasound in-process defect detection approach, presented in this paper, was performed on a titanium WAAM straight sample containing an intentionally embedded tungsten tube reflectors with an internal diameter of 1.0 mm. The ultrasound data were acquired after a pre-specified layer, in-process, employing the Full Matrix Capture (FMC) technique for subsequent post-processing using the adaptive Total Focusing Method (TFM) imaging algorithm assisted by a surface reconstruction algorithm based on the Synthetic Aperture Focusing Technique (SAFT). The presented results show a sufficient signal-tonoise ratio. Therefore, a potential for early defect detection is achieved, directly strengthening the benefits of the AM process by enabling a possible in-process repair.

**Keywords:** non-destructive evaluation; in-process robotic NDE; Wire + Arc Additive Manufacture (WAAM); ultrasound testing; total focusing method

#### **1. Introduction**

In 2019, the global metal Additive Manufacturing (AM) market size was valued at 2.02 billion € and was predicted to grow by up to 27.9% annually until 2024 [1]. AM technology plays a critical role in the latest industrial revolution, Industry 4.0, where there is a demand for smart factories capable of fabricating high-quality customized products [2].

**Citation:** Zimermann, R.; Mohseni, E.; Vasilev, M.; Loukas, C.; Vithanage, R.K.W.; Macleod, C.N.; Lines, D.; Javadi, Y.; Espirindio E Silva, M.P.; Fitzpatrick, S.; et al. Collaborative Robotic Wire + Arc Additive Manufacture and Sensor-Enabled In-Process Ultrasonic Non-Destructive Evaluation. *Sensors* **2022**, *22*, 4203. https://doi.org/10.3390/s22114203

Academic Editor: Nachappa Gopalsami

Received: 4 April 2022 Accepted: 28 May 2022 Published: 31 May 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

One such AM technology, called Wire + Arc Additive Manufacturing (WAAM), is a rapidly developing metal AM technology, based on a directed energy deposition process [3], which promises an automated fabrication of structurally complicated three-dimensional (3D) near-net shaped components [4]. The process is aiming to achieve superior costefficiency by reducing energy usage, material waste and time as compared to traditional subtractive manufacturing methods [5]. The technology has attracted the attention of sectors such as aerospace and naval engineering, due to their interest in weight reduction and increased geometrical complexity of solid metal components [6]. Moreover, WAAM offers the potential to build products using otherwise expensive materials such as titanium alloys, steel or nickel-based super-alloys [4].

Conventionally, the quality assurance of WAAM is performed by Non-Destructive Evaluation (NDE) after the full built completion via manually deployed methods such as ultrasound testing [7], eddy-current [8] or X-ray based imaging [9]. These techniques, however, require complex and time taking manipulation between workstations and often pre-processing or machining of the components [9], affecting the production throughput and therefore overall cost if a defect is discovered. Hence, in order to maintain the benefits of the already highly automated WAAM process, the demand for automatically deployed flexible NDE integrated in-process is high [10].

The detection of defects, in-process, facilitates the potential for real-time repair or early scrapping of parts, preventing the manufacturer from time-taking deposition of costly material over defective layers. Moreover, the deployment of automated NDE offers greater benefits such as positional accuracy, repeatability and high rates of inspection as compared to human NDE operators [11].

Recently published research has presented an important advancement in the field of automated in-process NDE of arc-based welding processes [12–14], which are manufacturing methods with similar applicable attributes and challenges to WAAM. The development of a multi-robot welding cell demonstrated the possibility of robotic welding and automated ultrasound NDE [15]. Full automation was achieved by a novel sensor-enabled robotic system based around a real-time embedded controller which enabled: (a) real-time communication, (b) data acquisition and (c) control of the process. Moreover, a UDP (User Datagram Protocol) communication protocol established through the Robot System Interface (RSI) [16], developed by industrial robot manufacturer KUKA, was used for the sensor-based robotic motion correction that could influence the pre-programmed robot's path through the sensor's feedback on the fly. The motion corrections were executed based on a novel developed motion software operating in real-time intervals (4 milliseconds intervals for KUKA Robot Controller (KRC) 4). Therefore, it was reported possible to utilize the cell for automated ultrasound inspection in three modes: (1) post-process continuous, (2) inter-pass in-process or (3) live-arc in-process. Further, the use of Force-Torque (FT) sensor-driven robotic motion for automated NDE of complex geometries was explored in [10]. The FT sensor facilitated path correction required for contact-based scanning of the aircraft wing cover through an as-built surface geometry that was inconsistent with geometry in the original CAD model. Therefore, the research faced similar automation challenges, associated with transducer deployment on the estimated pre-programmed path, applicable to possible automated NDE deployment on near net-shaped WAAM.

The ultrasound-based in-process NDE of welds at elevated temperatures was made possible by the development of novel, phased array ultrasound transducer (PAUT)-based, high-temperature and dry-coupled roller-probes [15,17]. Thanks to its design, based on a PAUT coupled through a water delay line and a flexible silicone rubber, the roller-probe was reported capable of withstanding temperatures of up to 350 ◦C, which made this device well suited for in-process NDE of arc-based manufacturing processes, where the resistance to elevated temperatures is highly desired [17]. This was a significant advancement in ultrasound NDE transducer development, given a typical commercial PAUT can only operate up to around 60 ◦C, while commercial delay lines can offset this limit to the temperatures only up to around 150 ◦C for a short period of time [18].

Further, the roller-probe technology has also been developed to couple with an as-built surface of WAAM components without the use of liquid couplants [19]. However, when considering the roller-probe inspection approach, new challenges emerged, as the as-built WAAM component features a non-flat and varying surface geometry (in both the scanning and the traverse direction) resulting in high contact forces being required to assure full compliance of the roller-probe tyre to the surface. Hence, the design must facilitate the transmission of the maximum possible ultrasonic energy without suffering signal losses. This resulted in the alteration of the internal liquid delay line with a heat-resistant solid core (delay line) made of Polyetherimide Polymer.

The key advantage of the novel WAAM roller-probe NDE approach is the removal of the post-deposition processing stages, which often included waiting for the sample to cool down, machining operations and manipulation between workstations. The inspection would then be performed through a flat surface either using direct gel-coupled contact with the sample or in water immersion tanks using gantry systems [7,20,21].

Owing to the novel WAAM dry-coupled roller-probe ultrasound NDE concept, the research has presented a possibility to detect Lack of Fusion (LoF) defects as small as 5 × 0.5 × 0.5 mm (width, length and height), through an as-built surface of the WAAM wall [22]. The ultrasound data acquisition called Full Matrix Capture (FMC) enabled the collection of raw time-domain data without consideration of any refractive boundaries or couplant conditions as the imaging was executed at the post-processing stage [23]. The developed ultrasound post-processing algorithms, based on the Synthetic Aperture Focusing Technique (SAFT) and Total Focusing Method (TFM) made it possible to overcome complications associated with multiple refractions present at a non-flat surface of WAAM and internal components of the roller-probe [24]. These algorithms, also called the SAFT-TFM package, were based on Delay and Sum (DaS) computational logic, where at first the Time of Flight (ToF) elapsed, between a PAUT element and a targeted image pixel, was calculated. Subsequently, the signal response from the corresponding time sample of an elementary A-scan was summed to the pixel. ToF calculation was repeated for every transmit–pixel–receive combination, thus, a fully focused image of the WAAM interior was formed. This novel ultrasound NDE approach was, however, presented on static inspection of WAAM components and was not yet deployed in-process on high-temperature builds.

Therefore, in this paper, the authors present an experimental multi-robot cell designed for WAAM deposition and automated in-process dry-coupled ultrasound NDE using a custom WAAM roller-probe. Within the cell, the plasma-arc WAAM process was controlled by the deposition software while a full external control of the NDE process was achieved by the sensor-enabled adaptive motion control package adapted to in-process WAAM NDE. The automated high-temperature WAAM roller-probe was deployed within a dwell time, set for inter-layer cooling, while sufficient coupling with the as-built surface of WAAM during the inspection was assured by the FT sensor. In this work, a titanium WAAM straight component (wall) with embedded tungsten reflectors was deposited to evaluate the performance of the in-process NDE approach. The use of tungsten tubes as cylindrical artificial reflectors, with known diameter, for ultrasound inspection technique calibration and evaluation has found its application in the fields of in-process welding inspection [12,13] as well as ultrasound inspection of WAAM [25]. An advantage of the tungsten can be realized by the possibility to manufacture inclusions, closely simulating defects such as Lack of Fusion (LoF) or keyholes, at the desired location [26]. During the in-process NDE, the position encoded FMC data were acquired using a high-speed ultrasound phased array controller. The SAFT-TFM package, then, enabled the highly accurate detection of artificial reflectors presented on an amplitude C-scan image. Cscan imaging provided a top-view over an interior of the WAAM component and was found effective for data review from a large inspection volume [27]. Thus, for the first time, a volumetric in-process ultrasound NDE of as-built WAAM was achieved, directly supporting research on producing right-first-time WAAM parts.

#### **2. The Architecture of the WAAM + NDE Cell**

#### *2.1. Hardware*

The automated robotic WAAM and NDE system depicted in Figure 1 was designed based on 2 × 6 Degrees of Freedom (DoF) industrial robotic manipulators (KUKA KR90 R3100) employed as a WAAM deposition robot and as an inspection robot. Additionally, as a part of the deposition robot, a horizontal rotary positioner (KUKA DKP-400V3) was also located within this cell and utilized as a rotational tooling mainframe and substrate clamping device. The deposition hardware, physically mounted on a deposition robot's end-effector, featured a water-cooled plasma-arc welding torch (controlled by: EWM-TETRIX 552 AC/DC SYNERGIC PLASMA AW welder) integrated into a deposition device with a local shielding [28], as seen in Figure 1a. The local shielding device was an aluminum enclosure with multiple gas outlet channels fitted, that provided an additional supply of the argon shielding gas on a high-temperature WAAM to prevent atmospheric contamination that could result in oxidation of the fresh deposit. Further, a wire-feed outlet with adjustable height was fitted on the deposition device, positioned to supply feedstock into the melt pool. The wire supply was controlled by an automatic wire feeder (EWM T drive 4 Rob 3 Li, EWM) that was attached directly to the deposition robot's arm as well. Lastly, the deposition head was equipped with a high dynamic range welding camera (Xiris XVC-1000) used to remotely assess the deposition quality.

**Figure 1.** Implemented (**a**) WAAM deposition cell with plasma arc process, and (**b**) Roller-Probe based NDE.

An inspection robot, seen in Figure 1b was equipped with an FT sensor (FTN-GAMMA-IP65 SI-130-10, Schunk, Germany) mounted on the end effector. A WAAM roller-probe was, then, attached to an FT sensor serving as an end effector to the robot flange. The roller-probe, depicted in detail on Figure 2, was driven by a high-speed phased array ultrasound controller LTPA (PEAK NDT, United Kingdom) mounted directly on the robot arm. Further, the communication between all hardware was achieved by a network switch (Zyxel Gigabit ethernet switch) enabling control of the WAAM process and NDE via a single ethernet connection plugged into the PC.

**Figure 2.** The internal structure of the roller-probe (**left**) and assembled device (**right**).

#### *2.2. Software Setup*

#### 2.2.1. Deposition

In this work, the deposition robot was controlled by a pre-installed PC with a WAAM-Ctrl (WAAM3D, Milton Keynes, UK) [29] application, streaming the deposition commands (robot paths, deposition parameters) directly to the deposition robot via RSI over an ethernet connection. The tool-path plan was generated using WAAMPlanner Software (WAAM3D, UK) [30], where the desired component was imported as a Computer-Aided Design (CAD), sliced into layers according to the pre-defined layer height, segmented into a set of individual building blocks from which the series of tool-paths was generated. Depending on the variables, such as material, geometry or deposition process, the deposition parameters were given to a WAAMPlanner and the post-processed file was generated, translating the information to a ready-to-stream xml file.

#### 2.2.2. NDE Software

The NDE inspection was guided by a software platform developed in the LabVIEW programming environment [31], which offers reliable communication between instruments and fast prototyping, through several available toolboxes and libraries. The Graphic User Interface (GUI) is presented in Figure 3, where the platform consisted of parallel state machines responsible for executing the program in sequence, controlling the inspection robot kinematics through the FT sensor feedback and ultrasound data acquisition in real-time.

The real-time corrections (every 4 milliseconds) of the robot's motion, based on linear interpolation, and control used for the in-process NDE work were based on a flexible robotic motion framework presented in [15] and developed for in-process inspection and automated NDE purposes. During the inspection, real-time adjustments of the inspection robot velocity, acceleration and contact force were available. Position-determined triggers were implemented to automatically switch between inspection and travel speed of the inspection robot, enabling/disabling the FT sensor-driven motion and data acquisition when needed. The *Z*-axis force control through the FT sensor was used to maintain sufficient contact with the WAAM component while the operator maintained the ability of real-time adjustments of the kinematics. In this work, the *Z*-axis motion corrections, associated with maintaining a steady force at the inspection speed, were calculated by the KRC based on the RSI configuration diagram. The X and Y translation, and A, B and C rotation-axes motion correction always remained in control of the initial inspection path-planning, while the appropriate motion corrections were calculated within the developed motion framework in 4 ms intervals and streamed through the RSI.

Further, taking the advantage of real-time communication with the inspection robot, the timestamped position of the inspection robot during an inspection was encoded to each FMC frame acquired. The FMC data were then processed within a MATLAB environment using a SAFT-TFM algorithm package, enabling positionally accurate analysis.

**Figure 3.** LabVIEW GUI for NDE process control and monitoring.

#### **3. Experimental WAAM Manufacturing**

#### *3.1. WAAM Wall Path Planning and Deposition Parameters*

To demonstrate the WAAM and NDE cell concept, and evaluate its performance, a titanium (Ti-6Al-4V) WAAM wall was chosen and designed for fabrication. The experimental wall was set to be 300.0 mm in length, 25.0 mm wide and a height given as 25.0 mm. However, knowing the nature of the WAAM process delivering near net-shaped components [4], extra material volume post-deposition was expected. Moreover, the height of the wall was not considered important since the goal was to evaluate the inspection of WAAM's interior with a specific volume. Therefore, the built process was stopped when the wall was found sufficiently high for in-process NDE demonstration to be performed.

The path planning designed in WAAMPlanner, seen in Figure 4, consisted of an oscillating deposition strategy [32], where a single bead, with a square zig-zag pattern, was deposited per layer. Relevant deposition parameters can be seen in Table 1 below.

**Figure 4.** Deposition Path Planning for Layer 1 of an experimental WAAM wall.


#### **Table 1.** Deposition Parameters.

#### *3.2. WAAM Wall Deposition*

Figure 5a displays a deposition setup where an experimental wall was built on a Ti-6V-4AL substrate plate, 12.0 mm thick, clamped to the tooling which was mounted on a rotary table of the horizontal positioner. The plate was clamped using welding clamps to prevent bending caused by heat-induced residual stress [33], typical for arc-based manufacturing processes such as welding [34].

**Figure 5.** Deposition clamping setup and a substrate plate with a deposited 1st layer (**a**) and deposition process with an active torch (**b**).

This clamping set-up has created a challenging and restricting working envelope; hence, the first stage of manufacturing was calibration and verification of the path motion by a dry run. At this step of WAAM part fabrication, the robot traveled through the produced deposition paths without an active torch or wire feed. Therefore, the correct positioning of the robot could be assured, knowing that the deposition head would not collide with the clamping. This was extremely important, especially during the deposition of the first few layers, after which the deposition head was high enough not to collide with welding clamps.

Following, Figure 5b shows an active deposition of the 1st layer, while the completed pass on the substrate plate is visible in Figure 5a image. It is worth mentioning, that the height of the first layer was measured to be 3.5 mm.

#### *3.3. Ultrasound Reflector Planting*

To evaluate the NDE defect detection capability, artificial reflectors were embedded into the experimental wall. In this work, tungsten tubes with parameters specified in Table 2 were utilized for this purpose. Two tubes were embedded into layer 3 by producing slots using a portable grinding machine. The tubes were located approximately 55 mm from each other. Tube 1 was placed parallel to the wall, in the approximate centre of the bead. Tube 2, on the other hand, was embedded in the transverse direction to the wall as seen in Figure 6a.

**Table 2.** Tungsten Tube parameters.


**Figure 6.** Tungsten tube embedding into layer 3 (**a**) and a subsequently deposited layer 4 covering tubes (**b**).

Further, Figure 6b depicts the wall after layer 4 where the tungsten rods were fully covered by the freshly deposited titanium. No significant inconsistencies (defects) in the surface quality that could cause a potential failure of the building process were observed once layer 4 was completed. However, a minor material built up was observed which was corrected after the subsequent layer deposition.

#### **4. In-Process NDE of the Experimental WAAM Wall**

#### *4.1. Ultrasound Inspection Parameters*

The ultrasound data were acquired using a roller-probe featuring a solid delay line housed in a silicone rubber tyre. The PAUT, with specifications found in Table 3, was positioned to sit on the top of the delay line.


**Table 3.** PAUT parameters.

The FMC data were collected using an LTPA phased array controller with 200 V excitation voltage and a fixed hardware gain of 60 dB. The time-domain matrix of the signals was formed by 3000 data samples for each transmits–receive pair at a sampling frequency of 50 MHz. During the data post-processing stage, the following acoustic velocities for longitudinal ultrasound waves were used for refraction and time-of-flight computations: (1) Delay line = 2480 m/s, (2) Rubber = 1006 m/s and (3) Titanium = 6100 m/s. These values were obtained by ultrasound pulse-echo measurements of the individual roller–probe's components and titanium coupons cut from a previous trial and heated to 150 ◦C.

#### *4.2. In-Process NDE*

To demonstrate the ultrasound in-process NDE capability on the titanium wall with embedded tungsten tubes, producing an air gap inside the WAAM, two subsequent layers were deposited to build a six-layer-high component. The deposition of two additional layers enabled a natural surface profile common for plasma WAAM deposition [4], without any significant negative influence from previous grinding and tungsten tube embedding.

Figure 7 shows a completed deposition of the experimental wall after layer six with a measured height of approximately 21 mm with an average layer height of approximately 3.5 mm. A width of 28 mm and a length of 305 mm were also measured.

**Figure 7.** Completed experimental wall and its dimensions.

As suggested by the literature [32], there is an optimal inter-pass dwelling time to allow for cooling of Ti-6Al-4V WAAM built using the oscillation deposition strategy. Therefore, the in-process NDE can be integrated into the build process to leverage this inter-pass dwelling time to complete the inspection of the last pass without delaying the built process. Accordingly, a dwell time of 9 min was set to allow inter-pass cooling during the deposition of the experimental wall as suggested by [35]. This time was set to avoid the formation of coarse *α*GB phase grain microstructure, and thus, achieve optimal mechanical properties of

this hypothetical component. Moreover, the time was found sufficiently long for in-process NDE to be performed without causing costly delays in the production process.

Before starting the NDE, the surface temperature was taken using a handheld thermometer. The surface temperature of the WAAM was measured and ranged to be between 180–230 ◦C along the wall, which was much lower than the operational limit of the rollerprobe (resistant up to 350 ◦C).

The NDE was initiated within the first 2 min of the deposition robot's retraction to its home position. Figure 8 shows a step-by-step inspection diagram, where at first, the inspection robot's end-effector approached the wall with a travel speed of 50 mm/s until the position 5 mm above the predicted as-built surface of WAAM was reached.

**Figure 8.** Inspection diagram showing the process and the sequence of robotic motions during an inspection.

The second stage in the diagram shows a contact establishment with the WAAM specimen. This was accomplished by an automatic trigger that recognized the robot's position (5 mm above the expected surface), which was followed by a change of robot speed to an inspection speed (in this work = 2 mm/s) and initiation of FT sensor-driven motion. A command to maintain a constant force of 130 N was sent to the inspection robot from LabVIEW via RSI; thus, the *Z*-axis position correction was no longer managed by a LabVIEW motion framework, but the kinematics corrections were calculated and applied by the KRC. The force applied to the component was set to a maximum force given by the FT sensor operational limit.

During the descending of the inspection robot on the surface of WAAM, the LabVIEW program was set to wait for 2 s, before sending coordinates of the next position. This "wait" command enabled the inspection robot to position itself on the surface with the required set stable force and without further freedoms in the X-Y plane that could result in inconsistent contact with a specimen.

Stage 3 of the inspection was initiated by sending coordinates of the next target position (in this scenario = the end of an inspection, +300 mm in the *X*-axis direction) and enabling encoded FMC data acquisition. The FMC data were acquired while the inspection robot traveled along the path with a steady force by correcting its *Z*-axis position to maintain a given force value with the experimental wall.

Once the end of the path was reached, the termination of the inspection was triggered by the change of the inspection robot's *Z*-axis targeted position. This was given as the *Z*-axis target position offset larger than 5 mm above the predicted WAAM surface. The trigger was used to disable the sensor-driven motion and the ultrasound data acquisition. The process was concluded by retracting the inspection robot to its home position according to the path planning.

The inspection volume from the experimental wall was set to 300 mm, therefore the time elapsed to inspect the component equaled 150 s with an additional approximate 60 s that included the approach to the specimen and the robot retraction back to a home position. It is worth mentioning, that the entire inspection took significantly less time than the period set for a dwelling (9 min), which complemented the objectives required for the in-process NDE of WAAM in this scenario. The total number of positions encoded FMC frames acquired was 200, giving a sample density of 1.5 per mm (sampling frequency = 0.75 Hz).

#### *4.3. Ultrasound Data Post-Processing: TFM Imaging and C-Scan*

After the completion of the in-process NDE, the ultrasound data were processed using a SAFT-TFM algorithm described in [22]. The TFM frames (B-Scan) were computed for a 25 mm × 19 mm region at 6 pixels/mm resolution, which is compatible with the 2 dB Amplitude Fidelity criterion of ASME V [36]. This window represented an internal volume of the desired component between the baseplate and a region 2 mm beneath the surface or just above the interface of layers 5 and 6, where potential defects would be expected. Moreover, this work was focused on the detection of tungsten tubes, therefore there was no interest in detecting and analyzing possible generated true defects from the WAAM process, since the calibration for these defects has not yet been developed.

To achieve a full C-scan, the computation was initiated by the ultrasound surface reconstruction using a SAFT surface imaging and surface finding algorithm. Afterward, the curves representing the WAAM surface contours were augmented into the 3-layer adaptive TFM algorithm to produce the TFM frames before their normalization. Normalizing all the frames used to construct the C-scan aided to visualize the entire image on the same dB scale. Using the raw unnormalized frames, the C-scan was formed by populating a new 2-dimensional array's columns with maximum detected amplitudes from all TFM frame's columns from n number (*n* = 200) of TFM frames.

The size of the C-scan presented in this paper was set to 150 × 200 pixels (Number of pixels in the horizontal axis of the TFM frame × the number of frames). The resulting C-scan image was normalized and plotted on a dB scale from the peak amplitude to an averaged noise level (0 to −12 dB), giving the best visual contrast between a signal from tungsten tubes and interference from the base noise levels.

#### **5. In-Process Inspection Results and Discussion**

In this section, the results of an in-process NDE are presented and discussed. The outcome of the in-process inspection is depicted in Figure 9a, where the signal from Tube 1 and Tube 2, with an internal diameter of 1.0 mm was successfully detected. At a first glance, stronger signal levels are observed from a longitudinally placed 30 mm long Tube 1. Noteworthy that a matching signal extension of approximately 30 mm along the inspection travel direction is also well noticeable. Tube 2, embedded in the traverse direction, shows visually weaker signal strength where the energy from the tube is represented by a concentrated signal in the centre of the corresponding frames approximately 100 mm from the inspection start point.

Following a visual analysis of the results, a maximum amplitude along an *X*-axis was presented in Figure 9b. Based on this plot, a Signal-to-Noise ratio (SNR) of up to 12 dB was achieved from the scanning of Tube 1 while an SNR of 10 dB was seen for Tube 2. Considering the dry-coupling condition, these SNR values were found sufficiently high for the indications to stand out from the background noise and be readily detected by the operator.

**Figure 9.** Results showing: (**a**) C-scan obtained from computed TFM frames and (**b**) extracted maximum amplitude along *X*-axis.

Further analysis shows signal strength variations from Tube 1 signal along the scan path where an SNR drop of only 4 dB was observed. This local signal strength loss can be associated with the possible changes to the contact quality between the rubber tyre and the non-flat wave-like surface profile of WAAM. This means the signal strength propagating into the specimen was fluctuating with the varying profile of WAAM. Further losses of SNR, especially for Tube 2, could be associated with a lack of compensation for the thermal gradient that affects an ultrasound wave velocity during propagation, as also pointed out in [12,37]. This means the image signal amplitude is negatively affected due to the loss of focusing precision during TFM image forming.

#### **6. Conclusions and Future Work**

In this paper, a design and demonstration of a novel multi-robot cell for WAAM and ultrasound in-process NDE was presented. The architecture, based on two robotic industrial manipulators featuring a deployed plasma arc WAAM process and high-temperature PAUT roller-probe was introduced along with a software control package, merging manufacture and NDE into a single continuous process.

The in-process NDE capability was demonstrated on a dry-coupled ultrasound inprocess inspection of the Ti-6Al-4V WAAM wall with embedded tungsten tube reflectors, with an internal diameter of 1.0 mm. Using the FMC data acquisition, a C-scan image of the experimental wall was computed by deploying a SAFT-TFM package. The results of the in-process inspection showed successfully detected embedded tubes, with distinguishable SNR of up to 12 dB.

Therefore, this work demonstrated the ability to detect defects just after the point of generation, which can pave the way for possible in-process repair processes to be deployed in the future. It can also be concluded that the presented research enables the further amplification of WAAM benefits by the deployment of flexible and automated NDE.

The future development is aimed at improving the speed of image forming, by the employment of graphics processing units. Moreover, the performance evaluation is targeted at transduction and automated deployment in various scenarios such as (1) inspection of geometrically complex WAAM components, (2) an investigation of probe deployment while the torch is active elsewhere and (3) at varying robotic NDE speeds. For the in-process defect detection and characterization area, the key aims are based on the development of thermal gradient compensation capabilities, which can further enhance defect detection and accuracy of the inspection approach.

Lastly, the research aims to develop a defect calibration procedure for various materials and a wide range of natural defects to enable automated defect detection and characterization that can further enhance the automation of the WAAM.

**Author Contributions:** Conceptualization, R.Z., E.M. and C.N.M.; methodology, R.Z., E.M. and R.K.W.V.; software, R.Z., M.V. and C.L.; validation, R.Z., E.M., Y.J., S.M. and D.L.; formal analysis, R.Z. and E.M.; investigation, R.Z., S.H. and S.M.; resources, S.F., S.W., C.N.M., S.G.P., M.P.E.E.S. and J.D.; data curation, R.Z., E.M. and D.L.; writing—original draft preparation, R.Z. and E.M.; writing—review and editing, E.M., R.K.W.V., S.W., D.L., C.N.M., M.V., C.L. and S.G.P.; visualization, E.M., D.L. and R.Z; supervision, C.N.M., S.W., S.G.P., J.D. and S.F.; project administration, S.W., C.N.M., S.G.P. and S.F.; funding acquisition, C.N.M., S.G.P. and S.W. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by EPSRC: (I) NEWAM (EP/R027218/1), (II) EPSRC Doctoral Training Partnership (DTP) (EP/R513349/1) and RoboWAAM (III) EP/P030165/1.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** No applicable.

**Data Availability Statement:** The data can be shared upon reasonable request.

**Acknowledgments:** The project was supported by EPSRC: (I) NEWAM (EP/R027218/1) and (II) EPSRC Doctoral Training Partnership (DTP) (EP/R513349/1) and RoboWAAM (III) EP/P030165/1. Further, the authors would like to thank the technical staff at the Lightweight Manufacturing Centre (LMC) (Renfrew, UK) for the support of this work, and the team at KUKA Robotics (Jeff Nowill, Alan Oakley, Steve Hudson in particular) for enabling and supporting the robotics work.

**Conflicts of Interest:** The Authors declare no conflict of interests.

#### **References**


### *Article* **A Novel Complete-Surface-Finding Algorithm for Online Surface Scanning with Limited View Sensors**

**Alastair Poole 1,2,\*, Mark Sutcliffe 2, Gareth Pierce <sup>1</sup> and Anthony Gachagan <sup>1</sup>**


**\*** Correspondence: alastair.poole@strath.ac.uk

**Abstract:** Robotised Non-Destructive Testing (NDT) has revolutionised the field, increasing the speed of repetitive scanning procedures and ability to reach hazardous environments. Application of robot-assisted NDT within specific industries such as remanufacturing and Aersopace, in which parts are regularly moulded and susceptible to non-critical deformation has however presented drawbacks. In these cases, digital models for robotic path planning are not always available or accurate. Cutting edge methods to counter the limited flexibility of robots require an initial pre-scan using camera-based systems in order to build a CAD model for path planning. This paper has sought to create a novel algorithm that enables robot-assisted ultrasonic testing of unknown surfaces within a single pass. Key to the impact of this article is the enabled autonomous profiling with sensors whose aperture is several orders of magnitude smaller than the target surface, for surfaces of any scale. Potential applications of the algorithm presented include autonomous drone and crawler inspections of large, complex, unknown environments in addition to situations where traditional metrological profiling equipment is not practical, such as in confined spaces. In simulation, the proposed algorithm has completely mapped significantly curved and complex shapes by utilising only local information, outputting a traditional raster pattern when curvature is present only in a single direction. In practical demonstrations, both curved and non-simple surfaces were fully mapped with no required operator intervention. The core limitations of the algorithm in practical cases is the effective range of the applied sensor, and as a stand-alone method it lacks the required knowledge of the environment to prevent collisions. However, since the approach has met success in fully scanning non-obstructive but still significantly complex surfaces, the objectives of this paper have been met. Future work will focus on low-accuracy environmental sensing capabilities to tackle the challenges faced. The method has been designed to allow single-pass scans for Conformable Wedge Probe UT scanning, but may be applied to any surface scans in the case the sensor aperture is significantly smaller than the part.

**Keywords:** NDT; free-form surface profiling; autonomous robotic systems

#### **1. Introduction**

Enabling robotised scanning processes is the harnessing of prior knowledge to fully traverse surfaces. For mobile or static-base robots completing NDT scans, knowledge of positions that have not been scanned is essential to ensure completeness of an inspection process that guarantees component integrity. Currently, this is ensured by planning a path over a known surface or part, that is then either verified of modified by an operator to ensure completeness.

Paths for parts equipped with an accurate CAD model can be produced automatically with commercial software. For parts without an accurate digital-twin, such as legacy parts or components with moulding errors, an operator has had to define a path on the robot's teach-pendant manually to capture its unique profile.

**Citation:** Poole, A.; Sutcliffe, M.; Pierce, G.; Gachagan, A. A Novel Complete-Surface-Finding Algorithm for Online Surface Scanning with Limited View Sensors. *Sensors* **2021**, *21*, 7692. https://doi.org/10.3390/ s21227692

Academic Editor: Steven Waslander

Received: 21 September 2021 Accepted: 17 November 2021 Published: 19 November 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 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 (https:// creativecommons.org/licenses/by/ 4.0/).

For one-off scans or for scanning parts with unique moulding errors, this process voids the high speed and repeatability benefits available to robotised NDT. In these cases, robotic platforms must be able to flexibly scan parts through online path planning, and to provide the same guarantee of completeness in surface coverage that is achieved by a human operator manually inspecting the part.

Recently, NDT has been enabled to define a 2-scan process. The first scan reconstructs the part for path planning of a subsequent scan with NDT equipment. The second scan can then commence, fully covering the known surface that is within reach of the robot. Methods of reconstructing part surfaces in the initial scan have been widely researched with respect to both Photogrammetry and in the field of machining.

In the field of Photogrammetry, automated robotised methods for free-form surface profiling have developed significantly. Processes involving 3D or 2D cameras have evolved from requiring user-inputted positions [1] to fully automated 3D model reconstruction techniques. Automated photogrammetry has been applied to a wide range of scales, from fine-detail model reconstruction using robotic arms [2,3] to large-scale reconstruction using autonomous robots with wide-aperture sensors [4]. A recent example of photogrammetry enabling a 2-pass scan within NDT utilising Structure-from-Motion (SfM) [5].

These methods have relied on multiple volumetric inspections of a complex object using wide field-of-view sensors such as traditional RGB or RGB + Depth (RGB/D) cameras. This work has considered surface profiling in the case of limited-range sensors, such as linescanners or ultrasonic devices that have a field of view many magnitudes smaller than the inspected surfaces. In the case of laser scanners, a volumetric pre-scan is not safe for human operators working nearby. Volumetric scanning of curved objects cannot guarantee surface discovery in the case of water-coupled ultrasound devices without lengthy re-scanning processes due to beam divergence and scattering.

Within the field of machining, validation of machining quality or accurate part profiling when there is no available CAD model has been implemented using Coordinate Measuring Machines (CMMs). CMMs utilising limited field-of-view sensors for fullsurface profiling have also been thoroughly investigated [6]. Their use has relied on spline-surface approximations to predict surface positions [7–9], or planar raster-tangent path planning [10]. These methods all require saturation of user-sampled positions, user input to define surface tangents, or rely on tangents defined by a gantry constrained rasterization pattern. The spline-surface approximation method has been successfully applied to ultrasonic-sensor surface discovery [11]. This method requires that the surface can be defined by a global spline, as opposed to an atlas of piece-wise smooth splines. This is disadvantaged when inspecting objects with discontinuities such as holes, as these cannot be captured by a global b-spline representation. Surfaces with global b-spline representations are also known as doubly ruled surfaces.

In aiding accurate offline path planning for Eddy-Current inspections, CMM machinery and software were applied within a manual pre-scan procedure to generate a CAD model [12].

This work has sought to completely remove the reliance on operator inputted information regarding the target surface, except for its maximal curvature. The authors have further aimed to completely automate the surface-profiling process, unconstrained by sensor type, robotic platform, or spline representations of the surface. The only requirement on sensor information is that the position of the surface relative to the sensor and the normal-direction of the surface are recoverable at each scan position. Approximate normal direction extraction requires discovery of at least 3 accurate local surface points.

Enabling full surface discovery requires a search process and memory structure to discover and store potential surface points for later traversal.

A candidate heuristic process are Flood Fill Algorithms (FFAs) that propagate through maps or networks in order to discover all positions within a connected surface or graph. The pseudo-code for two dimensional pixel maps has been presented in Algorithm 1 and accompanied by Figure 1.

#### **Algorithm 1** Flood Fill algorithm on the plane.


#### First iteration Last iteration

Green: Found-points Blue: Open-list Black: Boundary points

**Figure 1.** Colour Flood-Fill on the plane.

This work has generalised planar FFA heuristics to three-dimensional surface traversal, inventing the Complete-Surface Finding Algorithm (CSFA). Whereas FFAs require a preknown data structure, the novel CSFA requires only curvature information about the target surface to ensure complete coverage when applied to sensors of arbitrary dimensions and sensitivity.

Simple stack-based FFA and scanline heuristics are of particular interest in the simulation section. Scanline implementations choose a preferred direction of motion for search until a boundary position is reached. When a boundary position is discovered, the lesspreferable step is then taken until a free path is found in the preferred direction of motion. The resultant path is a traditional rasterization pattern, which is widely utilised within NDT path planning operations.

FFAs have been applied in various contexts, due to their simplicity and versatility. In the context of image processing, FFAs have seen ongoing widespread use in commercial products as a time-efficient method for filling a bounded region with a given colour [13]. The principle of the bucket-fill programme has been inverted to aid segmentation algorithms in 2D and 3D contexts from a user-inputted mask [14–16]. In recent years FFAs have aided machine-learning programmes in object recognition through automatic mask generation [17]. Mixed mapping and network theoretic implementations have been implemented to guide image reconstruction. First, FFAs were shown to be as effective as quality guided algorithms [18], and subsequently used to enhance nearest neighbour node quality optimisation methods in various fields [19–21].

Further, FFA variants have been extensively implemented in robotic path planning and control. Discretised potential field variants such as modified CFill and Flood-Field Methods (FFMs) have been shown to have greater time efficiency in comparison to Potential Field Methods (PFMs) [22,23]. FFAs have gained interest in the context of optimal path planning for 2D platforms [24,25], that has demonstrated flexibility through effective integration with optimal motion planners such as the A\* algorithm [26]. These concepts have evolved in application to optimal motion planning in 3D space for UAVs with an exhaustive search pattern [27]. Further FFA integration and heuristic mirroring has shown to enhance traditional path planning algorithms [28,29]. The above Flood-Fill methods have been implemented on data either with a pre-defined link structure or with a full exploration in each potential direction. For unknown surface profiling constrained by costly rearrangement procedures and a limited field of view, these procedures are either non-applicable or significantly sub-optimal.

#### **2. Method**

The aim of this paper has been to generate a complete set of points that describe the full surface by utilising the simple operations presented in Algorithm 1. To embed planar FFA operations within a 3D context requires the local position and normal direction information at each position.

A point source has been placed with a given stand-off from the surface in the normal direction, and a ray is then generated to intersect with the surface from which the tangent directions have been extracted. The 3D analogue of moving in the 2D principle directions is given by approximating the local surface covered by the sensor array with a tangent plane, defined by the observed points and approximate normal direction. Given a surface normal, the principal axes corresponding to 'UP' and 'DOWN' directions have been calculated through the Gram–Schmidt orthonormalization process [30]. Given a normal vector −→*<sup>n</sup>* = [*nx*, *ny*, *nz*]=[*ni*], and principle directions −→*<sup>e</sup>* <sup>1</sup> = [1, 0, 0], −→*<sup>e</sup>* <sup>2</sup> = [0, 1, 0] and −→*<sup>e</sup>* <sup>3</sup> = [0, 0, 1], the smallest component −→*<sup>x</sup>* has been selected as basis direction;

$$\overrightarrow{\mathbf{x}}^{\flat} = \{ \overrightarrow{\mathbf{e}}^{\flat} \,\_{i} \text{if } |n[i]| = \min\_{k \in \{1, 2, 3\}} |n[k]| \}. \tag{1}$$

The chosen basis direction has then been orthonormalised with the surface normal through the Gram–Schmidt process. The next basis direction −→*y* is taken by cross product of normal and tangent vectors. The basis directions [ −→*x* , −→*y* ] have formed the cardinal directions that planar FFA's utilise of 'DOWN' and 'RIGHT'. The point source traverses the surface in an analogue implementation of the traditional planar FFA, displayed in Figure 2. If no data or insufficient data is available at a given position, the current search point is marked as being in the ambient space with no additional points hypothesised, representing the 3D analogue of a 2D boundary position.

The approximate local surface normal direction can be extracted from at least three distance measurements from a single position with a 2D sensor array. Well-calibrated 1D linear sensors arrays would require two measurement values within a small displacement range, and single-element 0D sensors would require data from at least three positions. The algorithm may be applied to any sensor capable of a surface-tool stand off measurement.

The authors have further adapted the simple embedded stack-based FFA implementation to produce a scanline variation that generates automatic rasterization patterns within post-processing. For surfaces with uni-directional curvature, this has been achieved by retaining the order of the extrapolated *X*,*Y* basis directions. Retaining order on surfaces with significant curvature in two directions, such as the sphere or bowl requires including a 'preferable direction' reference. This is so that when *X* and *Y* surface–tangent directions change their order during traversal, preference is given to the one that lies within a con-

sistent plane in 3D space. On these surfaces, an irregular rasterization pattern emerges without preference vector. Irregular rasterization is not necessarily a negative feature, since for many robots and applications, there is an axial movement limit imposed that prevents multiple circular passes. This has been demonstrated in the results section, while rasterization is achieved in post-processing, online searches will require additional search positions that do not observe the target object in order to define boundary positions.

**Figure 2.** Flood Fill analogue in three dimensions. Grey lines represent iso-lines on the surface.

Finally a continuous surface must be discretised to ensure program closure, requiring a 3D analogue to 2D pixels. This structure allows positions that have been checked to be logged as seen. An Octree structure composed as a collection of boxes, or leaves has been chosen as it is less susceptible to numerical point-collisions present with a hash-table structure [31].

In order to assure full surface discovery, it is required that a step determined by the local information moves to a different Octree-node on the surface. Movements in 3D space under a set of changing basis directions may not align to a granular space oriented to the standard *X*,*Y*, *Z* bases. The undesirable effect of stepping within the same leaf may be prevented by moderating the Octree-leaf widths relative to the operator-specified step size *d*.

To ensure that each step defines a new leaf, the maximum potential length step within a leaf must be less than or equal to the step size. For leaf width *w* and step size *d*, the maximum step size, along the leaf's diagonal can be restricted with Equation (2).

$$w \le \frac{d}{\sqrt{3}}.\tag{2}$$

On high-curvature surface sections the surface will inflect within each Octree leaf, reducing the Cartesian arc-length from one observed position to another. An upper bound for the arc-length reduction for curved surfaces needs to be defined to ensure that each step along the surface defines a new leaf.

Arc-length reduction due to the projection of a line along a curved surface is bounded by the surface's curvature, which defines how a local linearisation deviates from the true surface profile. This term has been defined for a small step-vector −→*dx* by the Second Fundamental Form (SFF) denoted II [32];

$$\text{Arc-length difference} \approx \overrightarrow{d\mathbf{x}}^T \text{ II } \overrightarrow{d\mathbf{x}}/2. \tag{3}$$

The principal curvatures of the surface are eigenvalues of the SFF, and so the maximum possible inflection of a curve bound to the surface is in the direction of maximum principal curvature.

If the maximal principle curvature over the surface is *κmax*, then an upper bound on the minimal required leaf-width for a step size *d* may be derived;

$$kw/d \le \frac{|1 - |\kappa\_{\max}|d/2|}{\sqrt{3}}.\tag{4}$$

Dynamic discrete sampling may apply this principle to calculate minimal necessary Octree leaf-widths and step sizes in highly curved regions [33]; however, in this paper we restrict the analysis to uniform leaf-widths.

Flat surfaces have an over-sampling value of *w* = *d*/ <sup>√</sup><sup>3</sup> (in units of *<sup>d</sup>*), since the maximal principal curvature is 0. This has returned Equation (2), since the step-size in ambient space is equivalent to that of the surface projection, the step taken always contained within the same spatial plane. An example of detrimental point-aliasing when curvature is not considered has been presented in the results section.

Finally, in the case of surfaces with a significantly restricted width, the step size should be limited to less than half of the minimum surface width.

The complete algorithm when simultaneously considering a pulse-echo test has been described in pseudo-code in Algorithm 2.

#### **Algorithm 2** Pseudo-code for the novel CSFA.


The CSFA process results in a single-pass process that reduces the overall number of steps, displayed in Figure 3.

**Figure 3.** The one-step process enabled by the CSFA removes the necessity of accurate digital-twins and world-frame calibration, or lengthy robotic jogging procedures.

#### **3. Robotic Path Planning**

For robotic arm platforms, sections of the surface may lie out of reach, or a given motion may be impossible to execute due to a kinematic singularity [34]. These issues are incurred by a break in the correspondence between Cartesian space and the robot's fundamental coordinates, the possible joint-positions and link structure. In overcoming the spatial limitations of the robotic manipulator, oriented target-points were converted to configuration space coordinates. As a proof of concept investigation for the deployment of the novel CSFA, test pieces were chosen to test the algorithm's ability to ensure full coverage on curved and complex surfaces while minimising the risk of collision. Collision avoidance in the test cases were achieved by placing a motion-length limit. To maintain full coverage in the case of required back-tracking, any motion above this joint-space limit would cause the robot to move safely through a known point above the part. In the case of a convex part, point-to-point motion was considered admissible within one step if the subsequent point did not require motion in the current point's normal direction of more than the sensor-surface stand off. Since the algorithm requires an initial position to be defined along the surface, an initial configuration is given at the start. The path-planner then proceeded to choose the next in Cartesian space, and selected the candidate robotic configuration with the smallest joint-motion. If the selected point induced a configuration motion larger than the allowed threshold, the point was pushed back into the Open-List and another chosen until a suitable point was found or only large-motions were possible. In the latter case, the point with the smallest joint-wise motion was chosen. The robot was then sent joint-wise position command motions, avoiding kinematic singularities and ensuring the reachability of target points.

#### **4. Results**

Tests on shapes with key non-linear aspects have demonstrated the method's total coverage of generalised locally differentiable surfaces. The shapes chosen have been selected on the basis of surface irregularities that present challenges to full scanning. Surfaces with cut-outs that are not captured by a global surface spline representation demonstrated the advantage of the algorithm in handling machined parts, or in piecewise spline produced parts. These are not handled by the nearest available algorithm. Additionally, curved and doubly-curved surfaces were chosen to validate the suitability of the linearisation approximation method. In this section, surfaces chosen demonstrate complete coverage of locally smooth parts and parts with cut-outs. By demonstrating on positive, negative and zero curvature surfaces individually, the iterative and non-recursive algorithm has been validated for all locally smooth and holed surfaces. The process has been implemented in C++, utilising Simon Perrault's Octree structure [35]. Robotic simulations have been generated using RoboDK software with the Universal-Robotics UR10e as a demonstrative platform, with mesh simulations presented in MeshLab.

The CSFA has demonstrated ease in generating raster-motions on aerofoil components with varying step-sizes, displayed in Figure 4. Due to the relative flatness of the surface, a raster pattern was achieved. For more curved surfaces, there will be over-sampling of the space.

**Figure 4.** Demonstration of rasterizing a curved aerofoil component. The robotic path is traced in yellow, demonstrating the raster-like path obtained. (**a**) Sampling distance: 3 mm. (**b**) Sampling distance: 10 mm.

The method has been demonstrated to avoid surface-holes, re-scanning areas previously uncaptured in early-scan stages, displayed in Figure 5. The stack based memory of positions to check allowed effective full-surface discovery in the presence of irregular geometries. Figure 5 demonstrates that the CSFA has a clear advantage over gantry-based delivery platforms, covering complex surfaces without visiting the holed regions while still capturing the whole surface without needing the planar limits of the plate as input.

Repeatedly holed surfaces present multiple points of return, demonstrated in Figure 6.

The CSFA process makes a linear approximation of the surface in the neighbourhoods of discrete points. Displaying the algorithm on surfaces of positive and negative curvature, as in the sphere and bowl, demonstrates that it is robust in cases of local non-flatness. This is displayed in Figure 7.

**Figure 5.** The scan initially misses sections of the pipe due to the shape's cross-sectional hole.The missed points are picked up at the end of the scan as there is memory of surface-positions to check. Points found are marked in blue, the robotic path traced in yellow. (**a**) Initial scan-pass. (**b**) End-of-scan.

**Figure 6.** A complex flat plate holed with differently sized voids. The robotic path in yellow backtracks to allow for full surface discovery, shown by blue crosses, in the presence of surface-discontinuities.

**Figure 7.** Points discovered while simulating a scan on a bowl and sphere of radius 150 mm with a sampling distance of 3 mm. (**a**) Concave shape sampling. (**b**) Sphere sampling.

The irregular rasterization pattern may be seen in Figure 8. Unlike for surfaces of only one direction of curvature such as in Figure 4 or Figure 5, rasterization for double-

curvature surfaces is irregular. This incurs inefficient motions compared to traditional spiral-rasterization patterns.

**Figure 8.** Sampling on a concave shape. The robotic path, that can form irregular patterns without a preferred direction, is shown in yellow. Discovered points on the bowl are shown as blue crosses.

A horizontal rasterization pattern of subsequent circles resembling traditional spiralized patterns may be imposed by using a preferred direction vector; however, they can result in large re-arrangement procedures seen in Figure 9.

**Figure 9.** Sub optimal horizontal rasterization of a concave surface. Yellow trace lines demonstrate costly re-arrangement procedures to discover all the points shown in blue.

Curvature considerations are also demonstrably necessary for full surface coverage of components. Without over-sampling the space based on known surface curvature, full coverage is not guaranteed since taking a step will not necessarily take the algorithm to a new Octree-leaf. In turn, the algorithm stops prematurely as it aliases the points before and after the step within the Octree map. The effect of this is displayed in Figure 10.

**Figure 10.** Points in bold display the extent of discovery with no over-sampling regime. Sampling rate: 1 mm, radius of bowl: 150 mm.

#### **5. Experimental Results**

Complete coverage of locally differentiable surfaces has been shown in simulation when there are no limitations due to the robotic platform or sensor. Two key test pieces were identified to validate the algorithm's practicality in deployment. These were a surface of doubled-curvature and a surface with a cut-out. The doubly curved surface has been chosen to show that with the correct step size, sensors with small ranges may complete the search process, and that the approximation found for the surface normal is a suitable one. Moreover, since the important quantity in Octree sampling to guarantee completeness is the ratio of curvature to step size, the doubly curved surface shows that the heuristic presented is applicable to surfaces of all curvatures, given a step size that does not hinder sensor-surface coupling. The part with a section cut out further validates the approach when the surface is not globally represented by a global b-spline, as is necessary within the nearest algorithm. Since the algorithm utilises an iterative and non recursive heuristic, by demonstrating the process on these surfaces it is also demonstrated to work on curved surfaces with varying curvature and with cut-outs. It is important to note that the hardware chosen for completing the scanning process is the limiting factor, as smaller sensors are necessary to complete scans on objects that have extreme curvatures.

Experimental testing of the CSFA utilised three flange-mounted Panasonic HG-C1030- P lasers, connected to an Arduino board for real-time data collection. The laser's viewing range was 30 mm ± 5 mm, limiting the feasible step size over highly curved surfaces, as height variations of over 5 mm over the step would remove the possibility of further surface discovery. The laser's repeatability did not affect motion planning, as it was in the range of 10 μm. The lasers were held within a 3D-printed cradle displayed in Figure 11. An external laptop collected data from the Arduino and Universal-Robots UR10e robotic platform simultaneously. Connecting through a COM port and Ethernet-enabled TCP/IP connection, respectively, position data and commands were received and sent to the robot. The CSFA, data interpretation, and inverse kinematics solutions were coded in C++. The external laptop had an Intel Core i5 processor with the program built and run from a Visual Studio programming environment. Results were imaged using Meshlab.

To represent a non globally smooth b-splineable surface, laminate plates were placed into a planar pattern with a cut out displayed in Figure 12a alongside the point-cloud of collected data displayed in Figure 12b. Full discovery of the target surface demonstrates the applicability of the CSFA in cases where a direct path along the surface to every point

is not possible. The recollection of hypothesised points to visit allows traversal around corners, completely scanning regions with no direct path to one another.

**Figure 11.** The tri-laser holder, attached to the UR10e flange. The design with rotational symmetry around axis 6 of the robot minimised the footprint of the tool.

(**a**) (**b**) **Figure 12.** Automatic online profiling and scanning of an object with non-smooth shape. After a new point is found, the UT probe is applied to collect data. (**a**) Non-smooth shape created from arranged plates. (**b**) Resultant point cloud collected by the tri-laser and projected to the World-Frame using the live Joint-position of the robot.

> A curved mock-aerofoil segment provided additional experimental data displaying application to a use-case commonly seen within NDT in Figure 13. The total time taken for this use-case was 7 min 30 s for 3 cm spaced collection points. Providing a real-world

use-case for NDT, the full surface discovery of a doubly-curved surface with no-prior path planning provides the proof of concept for single-pass profiling of a complex surface and validation for the linearised surface approximation, while the part is relatively small compared to the robot's reach, the strength of this example is in the surface's extreme curvature. This use-case validates the application to surfaces commonly seen as complex within NDT.

**Figure 13.** Point Cloud of a complex doubly-curved surface profiled in real time, aligned to the CAD model in post-processing.

Finally, the proof of concept for simultaneous non-contact surface profiling with the tri-laser platform combined with Conformable-Wedge-Probe scanning is presented. The process is two-step; the tri-laser discovers the surface, displayed in Figure 14a, the tool reversed and the Conformable Wedge Probe applied to the discovered position, displayed in Figure 14b.

In deployment, sensor ranges provided the most significant challenge. Since the tool's base had a diameter of 5 cm, the curvature of parts observed within that region had to not exceed the viewing range of the laser-sensors in order to ensure the tool and part did not collide.

The main source of risk to deployment was an incorrect laser-tool calibration. During early testing, the sensor's beam had an orientation offset that with larger step-sizes often risking collisions with the part. Scanning the planar part with a re-printed tool that corrected the laser-flange alignment, and calibrated using the four-point method, the

standard deviation of points from the horizontal plane was 0.81 mm with mean signederror of *O* 10−<sup>16</sup> mm .

(**a**) (**b**)

**Figure 14.** Automatic online single-pass profiling of a surface. (**a**) Initial non-contact surface discovery and profiling with the tri-laser. (**b**) Subsequent application of the Conformable-Wedge coupled UT device.

Further, while demonstrations were limited by the lack of a collision avoidance schema, these experiments have proven the algorithm's capability in autonomous scanning processes, and applicability to robotic NDT. The main challenge facing industrial deployment of robotic NDT where parts have no accurate digital-twin is the flexibility of the robotic platforms in use, and their ability to define complete surface coverage. We have proven the ability of this algorithm to overcome this issue in realistic contexts.

#### **6. Discussion**

The authors have successfully implemented an adaptation of the FFA for full coverage of free form surfaces. The implementation has been demonstrated on positive and negative curvature surfaces, highlighting how the linearised approximation is not a detriment to overall surface following capabilities of the algorithm.

In post-processing, the CSFA has been shown to output a raster-path along arbitrarily locally differentiable surfaces. For doubly-curved surfaces, the rasterization pattern becomes irregular and there is an over-sampling of points. However, the method ensures total coverage of the part which is preferable in NDT to sparse sampling. The potential applications of the algorithm are not limited to automatic rasterization procedures. The Octree memory method would allow fully automated discovery and scanning of structures with any robotic platform, such as mobile robots traversing a large structure. Further, the traversal method can be applied with any limited-aperture sensor, enabling a generalised surface-movement strategy when sensor data is limited. Finally, the discrete-point approach allows the method to capture surfaces that cannot be globally splined. The limitation in the case of significant surface discontinuities such as part-edges is that the process will not necessarily find the other side of the part, discovery determined by the sensor's range and aperture size relative to the discontinuity. In practical deployments the sensor range was the key limitation, limiting the sensors step size due to the surface curvature so as to continue full surface discovery. Practical demonstrations applied to complex cut-out surfaces and realistic doubly curved aerofoil mock-ups show the real-world application with limited-range laser sensors. Proof of concept for wedge-probe coupled UT applications provide the NDT specific aims of this paper of removing the need to path plan for full-surface scanning.

For complex surfaces such as aerofoils or machined plates with cut-outs, the algorithm demonstrated is safe for deployment. For more complex shapes such as external pipescans, limited knowledge of the environment is necessary to prevent collisions. Future work will deploy the algorithm using low-cost environmental sensors to prevent collisions and path planning such as Rapidly exploring Random Trees (RRT) algorithms to scan complex components.

Future works investigating online surface profiling will further consider options to remove the necessity for user-inputted curvature estimates and step-sizes entirely. Adaptations to specific sensor types for surface profiling shall also be considered.

**Author Contributions:** Conceptualization and Methodology, A.P. and M.S.; Investigation, Software, Validation, A.P.; Supervision, M.S., G.P. and A.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the Engineering and Physical Sciences Research Council (EPSRC) as part of an ICASE PhD studentship, grant number S513908/1.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** Special thanks to David Carswell for code used within certain applications.

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

#### **References**


### *Article* **Automated Real-Time Eddy Current Array Inspection of Nuclear Assets**

**Euan Alexander Foster 1,\*, Gary Bolton 2, Robert Bernard 3, Martin McInnes 1, Shaun McKnight 1, Ewan Nicolson 1, Charalampos Loukas 1, Momchil Vasilev 1, Dave Lines 1, Ehsan Mohseni 1, Anthony Gachagan 1, Gareth Pierce <sup>1</sup> and Charles N. Macleod <sup>1</sup>**


**Abstract:** Inspection of components with surface discontinuities is an area that volumetric Non-Destructive Testing (NDT) methods, such as ultrasonic and radiographic, struggle in detection and characterisation. This coupled with the industrial desire to detect surface-breaking defects of components at the point of manufacture and/or maintenance, to increase design lifetime and further embed sustainability in their business models, is driving the increased adoption of Eddy Current Testing (ECT). Moreover, as businesses move toward Industry 4.0, demand for robotic delivery of NDT has grown. In this work, the authors present the novel implementation and use of a flexible robotic cell to deliver an eddy current array to inspect stress corrosion cracking on a nuclear canister made from 1.4404 stainless steel. Three 180-degree scans at different heights on one side of the canister were performed, and the acquired impedance data were vertically stitched together to show the full extent of the cracking. Axial and transversal datasets, corresponding to the transmit/receive coil configurations of the array elements, were simultaneously acquired at transmission frequencies 250, 300, 400, and 450 kHz and allowed for the generation of several impedance C-scan images. The variation in the lift-off of the eddy current array was innovatively minimised through the use of a force–torque sensor, a padded flexible ECT array and a PI control system. Through the use of bespoke software, the impedance data were logged in real-time (≤7 ms), displayed to the user, saved to a binary file, and flexibly post-processed via phase-rotation and mixing of the impedance data of different frequency and coil configuration channels. Phase rotation alone demonstrated an average increase in Signal to Noise Ratio (SNR) of 4.53 decibels across all datasets acquired, while a selective sum and average mixing technique was shown to increase the SNR by an average of 1.19 decibels. The results show how robotic delivery of eddy current arrays, and innovative postprocessing, can allow for repeatable and flexible surface inspection, suitable for the challenges faced in many quality-focused industries.

**Keywords:** non-destructive evaluation; robotic NDE; automated eddy current testing; eddy current arrays

#### **1. Introduction**

The global Non-Destructive Testing (NDT) market size was valued at USD 6.3 billion in 2021 with a predicted compound annual growth rate (CAGR) of 13.66% from 2022–2029 to hit a total market value of USD 16.66 billion [1]. This high level of growth can be attributed to the rise of "NDT 4.0", in which greater connectivity across the manufacturing supply chain is sought through the integration of connected sensors of which NDT techniques play a role [2]. To deliver this level of interconnectivity, it is now commonplace to see automated robotic delivery of NDT [3–6].

**Citation:** Foster, E.A.; Bolton, G.; Bernard, R.; McInnes, M.; McKnight, S.; Nicolson, E.; Loukas, C.; Vasilev, M.; Lines, D.; Mohseni, E.; et al. Automated Real-Time Eddy Current Array Inspection of Nuclear Assets. *Sensors* **2022**, *22*, 6036. https://doi.org/ 10.3390/s22166036

Academic Editor: Gui Yun Tian

Received: 15 July 2022 Accepted: 9 August 2022 Published: 12 August 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

The vast majority of the NDT market is based on volumetric inspection of high-value infrastructure and components, such as automotive/aerospace components or public rail infrastructure, primarily through the use of radiographic and ultrasonic testing. Due to this popularity, the automation of volumetric techniques is the most mature in the NDT industry. Further growth in the automation of volumetric NDT is expected to lag behind other NDT techniques, as innovation has shifted towards more novel and complex delivery of volumetric NDT as well as incorporating advanced imaging and post-processing techniques. Examples of these trends include performing the volumetric inspection at the point of manufacture for high-value components [7–12], performing aerial UAV-based volumetric inspection [13–16], optimising the amount of data gathered [17,18], and deploying machine/deep learning in the analysis of the datasets generated [19–21].

By contrast, the automation of surface inspection is far less mature and from 2022–2029 it is predicted to have the highest CAGR of any NDT technique due to the increased adoption of Eddy Current Testing (ECT) [1]. Of the 'big 5 NDT techniques, eddy current, magnetic particle, and penetrant testing were shown to be able to detect surface-breaking flaws, where others in the 'big 5 (ultrasound and radiographic) struggle [22].

Eddy currents are induced in a sample according to Faraday's Law of Induction [23] when a coil carrying an alternating current produces an alternating magnetic field and the conductive sample lies within this magnetic field. The induced eddy current in the sample is of the opposite phase to that of the coil and sets up its own magnetic field to oppose that of the coil. The eddy current density, *J*(*z*), decays exponentially with depth *z* in an isotropic material, and the sensed impedance is directly proportional to the current density [24]:

$$J(z) = J\_0 \exp\left(-\frac{z}{\delta\_{sd}(1+i)}\right) \tag{1}$$

In the presence of a defect, the current density is altered and this change in current density can be sensed as a change in impedance. The magnitude of the eddy current density decays exponentially and when it falls to 1/*e* of its surface value, the depth at which this occurs is known as the standard depth of penetration, *δsd*. The standard depth of penetration is dependent on the frequency of the voltage in the coil, the magnetic permeability, and the electrical conductivity of the component, and is widely viewed as the deepest depth a meaningful change in impedance can be sensed. Due to the exponential decay associated with eddy currents, they are ideally suited to detecting surface-breaking defects. This is in direct contrast with ultrasound where the front wall echo typically masks any shallow defects within a component. With correct eddy current probe design and frequency selection, an eddy current can be created that has a standard depth of penetration greater than or equal to the thickness of some thin-walled components, such as the canisters used in the storage of low-level nuclear waste.

Magnetic particle testing is restricted to the use of ferromagnetic metals and requires the component to be magnetized/de-magnetized frequently. While penetrant testing is not restricted to any material but requires the component to be coated in a penetrant and developer, which is frequently undesirable. Both magnetic particle and penetrant testing are subject to great operator error and do not produce discrete data points as a sensor is rastered across the component's surface making automation unfeasible. However, these drawbacks do not exist for ECT, and hence ECT is well suited for automation. As society moves towards Industry 4.0, automation is becoming increasingly important in surface inspection in the immediate future.

In comparison to volumetric techniques, ECT does not suffer from the health and safety concerns associated with radiographic inspection. Additional technical requirements may also prohibit the use of other inspection modalities. For example, multi-angle accessibility requirements and part size limitations may make computed tomography radiographic testing unfeasible. While for ultrasonic inspection, environmental factors may deter the use of a couplant. ECT has a significant advantage as single-sided access is all that is required, and no couplant is needed to perform an inspection.

Reuse and sustainable business practices are the main drivers behind the increased adoption of ECT, as detecting surface-breaking flaws that occur in operation is becoming increasingly important to prolong the safe operation of key assets for industries such as nuclear and aerospace. Furthermore, due to the lower market size, robotic delivery of ECT is far less common with only a few primitive integration efforts being reported in the literature [25–28]. To keep pace with the high throughput demands of modern production/maintenance lines, increased robotic deployment of ECT is necessary and vital to capitalise on this demand.

This paper presents, for the first time, the automated deployment of an eddy current array, via a flexible robotic cell complete with force–torque control, to scan a canister typical of the ones used in the storage of spent nuclear fuel. Table 1 shows a comparison between previously published papers that feature robotically deployed eddy current inspection and this work. Real-time adaptive control of a 6-axis robotic arm (KUKA Quantec Extra HA KR-90 R3100, Augsburg, Bavaria, Germany [29]) and an external rotary stage (KUKA DPK-400 [30]) with force–torque compensation was accomplished using a framework described in the author's previous work [31]. Force–torque compensation allowed for constant lift-off of the eddy current array during the inspection. This was intentionally carried out as it was shown that robotically delivered eddy current inspection offers far less noise when compared to that of manual eddy current inspection [32]. A commercial 32-element padded eddy current array from EddyFi (Part No: ECA-PDD-034-500-032-N03S, Québec, QC, Canada [33]) with a centre frequency of 500 kHz and an operating frequency range of 100–800 kHz, along with a 64-element Eddyfi Ectane 2 controller [34] were used to perform 180-degree rotary scans of a 1.4404 stainless steel nuclear grade canister with known stress corrosion cracks. Extensive software infrastructure coupled with the Eddyfi Ectane 2 Software Development Kit (SDK) allowed for the impedance data to be logged and analysed in real-time. All data were stored in a proprietary binary file format to allow for further post-processing in MATLAB.

**Table 1.** Comparison between previously published robotically deployed eddy current inspections and this work.


Where denotes yes and ✗ denotes no.

This infrastructure allowed for the acquisition and real-time analysis of impedance data. Novel image post-processing techniques, such as phase rotation and mixing, were shown to increase the Signal to Noise Ratio (SNR) of the resulting C-scan images by an average of 4.56 and 1.19 decibels, respectively. It is envisaged that studies such as this will progress eddy current testing to match the level of flexibility and quality enjoyed in the post-processing of ultrasonic datasets [35,36].

#### **2. Experimental System**

NDT is crucial to safety-conscious industries such as nuclear [37]. Traditionally, the inspection of nuclear assets is highly resource-intensive and complex. The inspection challenge is complicated further when the asset lifetime exceeds the original design intent. This problem is one that is currently being faced in the UK, where government policy has shifted from favouring reprocessing to long-term storage of nuclear assets [38]. Spent nuclear fuel due for reprocessing is now being stored long term as reprocessing facilities are closed down. Some sites store low-level waste in canisters made from 0.9 mm thick 1.4404 stainless steel. These canisters range from 130–150 mm in diameter and are ~300 mm in length. To allow for effective cooling, the canisters are stored in facilities that are partially

open to the environment. Given the coastal location of the UK, stress corrosion cracking is a concern. Due to the points mentioned above, canisters with intentionally induced stress corrosion cracks were scanned with an eddy current array and the acquired impedance data were analysed within this study.

#### *2.1. Hardware and Experimental Summary*

Figure 1 shows the experimental hardware used in the automated deployment of the eddy current system. A nuclear canister with a matrix of 16 stress corrosion cracks shown in Figure 2 is held within a mechanical chuck on top of a KUKA DPK-400 external rotary stage that has an angular resolution of 0.009◦. The padded Eddyfi eddy current array (Part No: ECA-PDD-034-500-032-N03S) is mounted in a bespoke 3D-printed housing which is in turn secured to an IP-65 rated gamma force–torque sensor from ATI Industrial Automation (Apex, NC, USA) [39]. To move the sensor to the height of interest for the inspection, the eddy current array, 3D-printed housing and force–torque sensor assembly, are mounted to the flange of a KUKA KR-90 robot. Both the KR-90 and DPK-400 external rotary stages are controlled via a KRC 4 controller [40].

**Figure 1.** Eddy current inspection hardware.

In order for the eddy current array to be pressed onto the canister surface in the direction of the canister' centre, a calibration tool was manufactured to teach the KR-90 robot a new base coordinate system. The calibration tool was made so that it would align the centre of the chuck to the centre of the rotary stage. Additionally, the calibration tool allowed for the centre of the tool along with 4 concentric radial calibration points at 150 mm in 90◦ increments to be taught to the KR-90 robot. By teaching the KR-90 robot these points, it was able to know where the centre of the canister and rotary stage was relative to its own coordinate system, and ensure motion was performed relative to this point. This effort guaranteed that the eddy current array was always pressed against the canister surface in the direction of the canister's axial centre and helped establish good electromagnetic coupling during the automated inspection.

**Figure 2.** Canisters with a matrix of 16 stress corrosion cracks. Depositions of 5 μL droplets of sea water, 3.03 g/L of MgCl2, 15.2 g/L of MgCl2 and 30.03 g/L of MgCl2 were used to induce the cracks in the top row, left, central and right columns, respectively.

The eddy current array is deployed to the height of interest in the Z-direction of the canister via a variable set by the user on the Graphical User Interface (GUI) of a LabVIEW program using a framework similar to previously published work [31]. Force and torque in and around all three axes shown in Figure 1 are sensed via the force–torque sensor and are transmitted to a LabVIEW program via the robot controller using the Kuka Robot Sensor Interface (RSI) [41]. The transmission of the force and torque characteristics allowed for: (1) the adaptive motion of the eddy current sensor during inspection; (2) the balancing of the eddy current probe and the subsequent triggering for the acquisition of the impedance data to begin; and (3) the triggering of the rotary stage to begin movement. It is important to note that the force–torque sensor was calibrated with all hardware mounted prior to any automated inspection through a program provided by the manufacturer. The calibration enabled the net force and torque values being applied to the eddy current array and mounting assembly to be correctly sensed and subsequently transmitted to the LabVIEW control program for adaptive motion to be performed.

The KR-90 robot presented the eddy current array onto the surface of the nuclear canister at the user-specified height, and a target force and torque of 10 N and 0 Nm were met in the Y-direction and X-axis, respectively, for 3 s. Once this time period had passed, the balancing of all coils within the eddy current array was performed when the probe was stationary. After a further 3 s, the impedance data acquisition along with the rotary stage movement was triggered.

During the inspection, a PI control system was used to monitor and correct both the force in the Y-direction and the torque around the X-axis at the previously mentioned target force and torque values. It was found that *P*- and *I*-values of 0.1 and 0.0 gave an adequate control response. Control of the eddy current probe's orientation in this manner allowed for minimal variations in the lift-off of the eddy current array throughout the inspection providing excellent coupling. Other previously published literature has shown that lift-off can be reduced via advanced signal processing and elaborate probe design [42]. These efforts are often particularly involved and particular to one sample/defect type. As a result, these lift-off compensation strategies are complex to deploy and benefit from. The approach in this paper of utilising a force–torque sensor in combination with a padded ECT array

provides experimental flexibility and passively compensates for any lift-off variation at the point of acquisition giving wide-reaching benefits.

The acquisition of the impedance data was stopped when the rotary stage had completed the angular movement requested by the user from within the LabVIEW program. A singular scan can be summarised by the following process:

	- a. Probe type;
	- b. Probe configuration (axial and/or transversal—See Section 2.2);
	- c. Frequencies;
	- d. Voltages;
	- e. Gain;
	- f. Repetition rate.
	- a. Linear speed of the KR-90 robot;
	- b. Approach speed of the KR-90 robot;
	- c. Angular movement of the canister/external rotary stage;
	- d. Angular speed of the canister/external rotary stage;
	- e. Target force for the KR-90 robot to apply the array onto the canister.

#### *2.2. Eddy Current C-Scan Acquisition*

Figure 3 shows a generic eddy current array layout along with illustrations of the transmit and receive pairings for the axial and transversal configurations. Depending on the probe geometry, there may or may not be an equal number of transversal and axial transmit and receive pairs. Each pairing in each configuration generates a data point of complex impedance data. The probe is linearly scanned perpendicular to the coil columns as noted in Figure 3, and the data points are logged into a complex 2D array. The resulting complex arrays can then be post-processed, and the vertical component of the post-processed complex array can be plotted in a C-scan format to show any defective signals with maximum Signal to Noise Ratio (SNR).

As can be seen in Figure 3b for the axial configuration, coils in the array are excited in one column and reception of the impedance data is performed across the array in the second column. Conversely, the transversal configuration documented in Figure 3c shows coils being excited and reception of the impedance data being performed within the same vertical column of coils.

The coil firing sequence is changed between the axial and transversal configurations to achieve greater sensitivity to differing defect orientations. With reference to the coordinate system in Figure 3, a larger change in impedance would be observed for a defect that is aligned with the X-axis for a transversal configuration over that of an axial configuration. This is due to the defect more severely intercepting the eddy current that exists between the two transmit and receive coils in the transversal configuration over that of the axial configuration. This greater compression of the eddy currents caused by the defect presence will have a large effect on the electromagnetic field and by proxy the sensed change in impedance. The opposite can be said to be true for a defect aligned in the Y-direction. For

further reading, Ye et al. [42] provide a thorough theoretical and experimental investigation of this phenomenon.

**Figure 3.** Eddy current array transmit and receive configurations. (**a**) Generic Eddy current array layout with two vertical columns of coils. (**b**) Axial transmit and receive configuration where **x** (**in blue**) corresponds to the transmit/receive pair centres of the excited eddy current channels in the test part. (**c**) Transversal transmit and receive configuration where **x** (**in green**) corresponds to the transmit/receive pair centres of the excited eddy currents in the test part resulting from the first/odd column of coils, and where **x** (**in orange**) corresponds to the transmit/receive pair centres of the excited eddy currents in the test part resulting from the second/even column of coils.

It is also evident from Figure 3 that the centres of excitation are not aligned between the axial and transversal datasets in the X-direction. Moreover, for each coil column within the transversal dataset, the data centres are also misaligned. As alluded to in Section 2.2, this positional misalignment is corrected within the LabVIEW program and ensures that the resulting complex array for each dataset has the same spatial grid.

Key to the positional compensation is the acquisition rate of the eddy current array and the angular speed of the rotary stage so that each acquisition point aligns with an integer number of divisions of half the array coil column pitch, Δ*x*. The acquisition rate and number of divisions between half of the array column pitch are set by the user, and the coil pitch is defined by the geometry of the array. These three variables are used to set the angular speed of the rotary stage. For example, if an eddy current array has a column coil pitch of Δ*x* = 7 mm, an acquisition rate of 50 Hz, and 50 divisions, the linear speed would need to be <sup>7</sup> <sup>2</sup> <sup>×</sup> <sup>1</sup> 50 / <sup>1</sup> 50  = 3.5 mm/s. This linear speed can then be converted to rotational speed by dividing the diameter of the canister at 150 mm to give the angular speed of the rotary stage at 1.34 deg/s. Whilst individual datapoints are not positionallyencoded, the positional location is extrapolated from setting the angular speed relative to the eddy current probe geometry and acquisition rate as mentioned above. By doing so, it ensures that data are acquired at both the axial and transversal data centre points on the X-axis as the array is linearly scanned.

In order to ensure a common spatial grid, the first and last impedance data points corresponding to a distance of half the coil pitch are discarded within the axial complex array. By discarding the first set of data points that cover half the coil pitch, the axial complex array in the X-direction is synched with the first/odd column of the transversal dataset. Moreover, by discarding the last set of data points that cover half the coil pitch, the axial complex array in the X-direction is synched with the second/even column of the transversal complex array. This discarding of data is shown graphically in Figure 4a. The

resulting data is then linearly interpolated in the Y-direction to align with the Y-coordinates of the transversal complex array.

**Figure 4.** Illustration of complex impedance data positional compensation performed between axial and transversal configurations. (**a**) Axial complex array positional compensation. (**b**) Transversal complex array positional compensation.

The transversal C-scan array is similarly compensated by separating out the first/odd and second/even columns into separate arrays. Impedance data corresponding to a distance of a full coil pitch is discarded from the start of the odd array. Conversely, the opposite operation is performed on the even array where impedance data corresponding to a distance of a full coil pitch is discarded from the end of the array. This process is graphically illustrated in Figure 4b. Once all data are discarded, the odd and even arrays are interleaved together to make one C-scan array that is on the same positional grid as the axial C-scan array.

Once all data were collected and positionally compensated, oversampling is undertaken in the vertical direction of the array. No oversampling is performed in the horizontal scan direction as this is controlled adequately by setting the rotational speed and acquisition rate of the robot as described in the previous paragraphs. The oversampling is performed via linear interpolation of the raw impedance data. It was found that this linear interpolation was fast to implement and produced negligible errors with a maximum error of 2.12% and an average of 0.55% across both the axial and transversal datasets at 250 kHz.

By performing data compensation in this manner, a common spatial grid is established for each dataset configuration, enabling like-for-like comparison and further advanced post-processing techniques such as mixing of datasets.

#### *2.3. Software Infrastructure*

Extensive software infrastructure to control the eddy current Ectane device, as well as receive and process the acquired impedance data in real-time was developed and is documented in Figure 5. Literature has previously well documented the robotic software infrastructure required [31,43] and as a result, the work presented herein will focus on the eddy current software development effort.

**Figure 5.** A flow chart showing the data transfer between different software and hardware elements.

In total 3 programs were developed: (1) A C program that houses the Eddyfi Ectane 2 SDK; (2) A LabVIEW program that receives, post-processes and plots impedance data in real-time as well as saving the data in a binary file format; and (3) A MATLAB reviewer program that reads in the binary file for further post-processing.

Both the C and LabVIEW programs are state machines. States within the C program are evaluated through a switch statement within the main while loop. In addition to the main while loop, the C program contains two threads that each have local host Transmission Control Protocol (TCP) connections. The first listens for standardised comma-separated string commands from LabVIEW and the other sends 32-bit impedance data from the Ectane device to the LabVIEW program. The same infrastructure with reverse logic is mimicked within the LabVIEW program through JKI state machines [44]. The standardised comma-separated string that is sent from LabVIEW is carried out in the following format:

```
state, IPAddress, configuration, acquisitionRate, gain, 
freq1, voltage1, freq2, voltage2, freq3, voltage3, 
freq4, voltage4, freq5, voltage5
```
As can be seen, there are 15 variables housed within the standardised string command. The first of which is the state that the C program should execute, and these are summarised below.


The second is the IP address of the Ectane device in order for the C program to connect to the Ectane device. Third is the configuration of the probe (i.e., will axial and/or transversal datasets be acquired? What probe is being used?). Next is the acquisition rate and gain of all Ectane channels. The final ten are the voltages and frequencies of each Ectane channel. As the Ectane device can acquire 5 datasets at different voltages and frequencies each of these must be specified even if some are unused.

The raw impedance data are received in the LabVIEW program as a series of 32-bit numbers and are immediately queued to be sequentially analysed in two additional threads. Using a 6 core, 2.6 GHz Intel i7-8850 H processor, it was found that the queueing of the received data was performed in 1 ms. As previously, these threads are implemented via two JKI state machines.

The first thread takes each 32-bit number and separates out the first and last 16-bits of data as these correspond to the imaginary and real impedance components. Additionally, the first thread reformats the impedance data into geometric order as the coils are pulsed in a pseudo-random fashion to prevent crosstalk caused by mutual inductance. Moreover, the first thread compensates for offset in coil excitation in the X-direction. Further details of this coil excitation compensation are provided in Section 2.2. It was found that this process was executed in 1 ms on a 6 core, 2.6 GHz Intel i7-8850H processor.

The second thread within the LabVIEW program is responsible for interpolation in the Y-direction between axial and transversal dataset configurations, oversampling, basic mixing of different datasets and live plotting of the impedance magnitude. As before, further details of this Y-direction interpolation and mixing of datasets are provided in Sections 2.2 and 2.4, respectively. Likewise, it was found that this process was executed in 5 ms on a 6 core, 2.6 GHz Intel i7-8850H processor. It is noted that the timings reported should be representative of any array used as the software infrastructure is built for the maximum number of elements, channel pairings, and number of frequencies.

This multi-threaded approach is illustrated in Figure 6 and provides data acquisition, positional compensation, and interpolation of impedance data whilst displaying various impedance magnitude C-scans in real-time to the user, all within the LabVIEW software environment with minimal 7 ms lag. The user can then select a directory to store the acquired data in a binary file format for future post-processing and analysis.

**Figure 6.** Illustration of the multi-threaded C and LabVIEW programs.

#### *2.4. Image Enhancement of Impedance Data*

It was shown in the literature that the impedance plane of the acquired data can be complex to interpret and variations in probe lift-off and wobble can commonly be mistaken as signals from defects [45,46]. Therefore, great care was taken in this work to minimise these adverse effects. Methods such as optimal probe design [47], multi-frequency excitation [45], and phase rotation [46] were shown to reduce such effects. Due to this work utilising commercial off-the-shelf (COTS) equipment, only multi-frequency excitation and phase rotation were performed. Multi-frequency excitation of 4 separate frequencies was conducted as the data were acquired and mixing of the datasets as described in Section 2.4.2 was performed in post-processing. Additionally, phase rotation was performed on the acquired C-scan datasets. All post-processing was performed via the MATLAB review application mentioned in Section 2.3.

#### 2.4.1. Phase Rotation

The signature of adverse effects such as lift-off and wobble experience a phase difference in the response caused by a defect on the impedance plane. It is therefore common to phase rotate the data so that the response from the lift-off aligns with the horizontal axis of the display impedance plane, and plot C-scan images of the resulting vertical component of the impedance [48]. Due to the phase difference observed between the lift-off variations and that of a defect, the resulting C-scan will show any response from a defect clearly.

Mathematically, this is described in Equations (2) and (3). Equation (2) describes the resulting acquired impedance array from Section 2.3, and Equation (3) describes the mathematical operation performed to phase rotate the data by an angle, *θ*. This can be carried out at the point of acquisition or in post-processing. For this study, the decision was taken to phase rotate the data in post-processing to maintain maximum flexibility with the acquired data.

$$Z = \mathbb{R} + iX \tag{2}$$

$$Z\_{rot} = Z(\cos(\theta) + i\sin(\theta)) = (R + iX)(\cos(\theta) + i\sin(\theta))\tag{3}$$

2.4.2. Mixing Eddy Current Datasets

As the impedance data were acquired onto a common spatial grid, mixing of datasets recorded under differing configurations or frequencies can be performed by superimposing the impedance C-scan data. This is graphically illustrated in Figure 7.

**Figure 7.** Illustration of mixing datasets *Z*<sup>1</sup> and *Z*<sup>2</sup> impedance data to make *Zm* mixed data.

Two differing mixing methodologies were performed with the first being a simple sum and the second being a selective sum and average. As the name implies, the simple sum summates complex impedance datasets on a pixel-wise basis. For the selective sum and average, data above a defined noise floor were summated and everything below was averaged. The noise floor was defined as being 5 times the RMS values reported across a non-defective section of one of the impedance datasets to be mixed.

#### **3. Results**

Three 180-degree scans of the canister shown in Figure 2 were undertaken with both transversal and axial datasets being simultaneously acquired at frequencies of 250, 300, 400, and 450 kHz with an amplitude of 2 volts for each frequency channel, 30 dB of gain, an acquisition rate of 40 Hz, and a rotational speed of 1.72 deg/s. Each scan covered an area of 7687.1 mm<sup>2</sup> (array height of 32.625 mm × half the circumference of a 150 mm canister equating to 235.62 mm) making the final stitched image representative of an area of 23,061.3 mm2. The interpolation was set to five, and the increments between half a coil pitch were specified at 20, giving a spatial resolution of 0.225 mm and 0.0563 mm in the vertical and horizontal directions, respectively. Positions were chosen for each scan so that

they were acquired one array coil above each other with no overlap. The impedance data for all three scans were vertically stitched together and axial channel C-scans of the vertical impedance component from the impedance vector are shown in Figure 8. One of the stress corrosion cracks in the centre of the far-right column is highlighted. To the right of each C-scan, the impedance plane Lissajous for the highlighted defect is also shown along a horizontal cursor passing through the maximum intensity of the defect indication in the C-scan. It can be seen, that the impedance plane response of the same defect for different frequencies varies drastically in amplitude and phase due to the differing interaction depth of the eddy currents with the defect [46].

**Figure 8.** Axial vertical impedance component C−scan images at 250, 300, 400, and 450 kHz on a dB scale alongside impedance plane plots of the response from the highlighted defect.

Additionally, Figure 8 also shows that at 250 kHz and 450 kHz, the impedance plane contains a large horizontal component and as such the resulting image contains a large amount of noise. In order to compensate for this effect, the impedance data at each

frequency were phase rotated so that the SNR of the highlighted defect was maximized – see Figure 9.

**Figure 9.** SNR vs. Angle of phase rotation for the axial dataset acquired at 250 kHz.

Figure 10 shows C-scan images of the optimised phase rotated axial data, while Table 2 denotes the SNR increases for both axial and transversal datasets at all frequencies recorded for the target defect. The increase in SNR for all defects is visually evident in Figure 10, and on average, the SNR was increased by 4.56 decibels for the targeted defect. This result illustrates the effectiveness that phase rotation can have on increasing the image performance of C-scans and the benefit of being able to flexibly perform such a task in post-processing.



To further enhance image quality and reveal more about the nature of the defect, a mixing of different datasets, as described in Section 2.4.1, was performed. The optimised transversal and axial datasets at 250 and 450 kHz were mixed together, as the dissimilar frequencies would produce differing eddy current penetration depths and thus be influenced in differing manners. Equation (4) mathematically describes the penetration depth of an eddy current for a given material, where *f* is the frequency of the voltage being excited in the array coils in hertz (Hz), *μ* is the magnetic permeability of the component under test in henries per meter (H/m), and *σ* is the electrical conductivity of the component under test in siemens per meter (S/m).

$$\delta = \frac{1}{\sqrt{\pi f \mu \sigma}}\tag{4}$$

**Figure 10.** Phase-rotated axial vertical impedance component C−scan images at 250,300,400 and 450 kHz on a dB scale alongside impedance plane plots of the response from the highlighted defect.

For stainless steel, with an electrical conductivity of 1.08 × 106 S/m, and a relative magnetic permeability of 1.0025, a frequency of 250 kHz would produce a penetration depth of 0.967 mm, while a frequency of 450 kHz would produce a penetration depth of 0.721 mm.

The resulting mixed C-scan image is shown in Figure 11. Table 3 documents the SNR of the highlighted defect. As is shown in Table 3, the SNR of the defect for the simple sum approximates to be the average across all four datasets that contributed to the mixed image, and as such it can be said the imaging performance has not been improved by this mixing methodology. Interestingly, this is a result that is also observed in ultrasound when fusing multi-modal Total Focused Method (TFM) images [49]. By contrast, the selective sum and average technique were able to boost the SNR by an average of 1.19 dB, demonstrating an increase in imaging performance.

**Figure 11.** Mixed vertical impedance component C−scan.

**Table 3.** Mixed Image SNR.


It is acknowledged that in this study, SNR is the only metric being used to evaluate the eddy current detection system. A better metric would be a physical parameter related to the geometry of the defect itself (i.e., crack extent, crack depth) and whether this is better reflected in the mixing of datasets. As reported in the literature, this is a highly complex inversion problem, with successful inversions demonstrated on only simple geometries [50–53] or overall dimensions such as the depth or extent on complex defect geometries [54,55]. In all these studies, the defects were manufactured to specified geometries before eddy current testing which is somewhat removed from a real inspection scenario where prior knowledge of the defect geometry is not known. In addition, the sizing algorithms used vary drastically from defect to defect making the inversion of defect size somewhat deterministic and not well suited to automated deployment and analysis with which this paper is concerned. While the current system and signal processing cannot currently invert physical defect size, it was shown on another sample and different probe that is better suited to low-frequency operation, that the system is able to detect embedded defects ~3 mm below the inspection surface.

To understand more about the physical geometry of the highlighted stress corrosion crack, a macrograph was taken at 96 times zoom and is shown in Figure 12. It can be seen that the defect under inspection is a multifaceted stress corrosion crack. Due to its multifaceted nature, the interaction with the induced eddy current will be highly complex and therefore inversion of the physical geometry would be highly challenging. It is expected that for a simple linear defect, such as a fatigue crack, mixing of datasets would lead to benefits in defect characterisation even if the SNR was adversely affected. This issue is subject to future work and will be investigated by the authors at a later date.

**Figure 12.** Photo of crack matrix and micrograph (**a**) Photo of crack matrix with the defect of interest

#### **4. Conclusions**

This paper demonstrates for the first time how eddy current inspection with full image post-processing functions can be robotically deployed, showing a significant step closer to Industry 4.0 applications. Variations in the lift-off of the eddy current array were compensated for by the use of a PI control system and a force–torque sensor ensuring excellent low-noise coupling throughout the inspection. Extensive software infrastructure was developed that allowed for the eddy current data to be post-processed to enhance the generated images and reveal more about the nature of the defects under inspection.

highlighted in a red circle. (**b**) Micrograph of the defect of interest at 96× zoom with desaturated background.

The capability of the eddy current inspection system was demonstrated by inspecting a nuclear canister with a matrix of 16 stress corrosion cracks. Three 180-degree scans were conducted, gathering axial and transversal datasets at four different frequencies simultaneously—250, 300, 400, and 450 kHz—detecting 15/16 stress corrosion cracks. In the resulting data, one defect was highlighted, and various post-processing techniques were employed to increase the image quality. It was shown that, by phase rotation alone, the SNR could be increased by an average of 4.56 decibels. Dataset mixing was also attempted, and it was shown that a selective sum and average could boost the SNR by an average of 1.19 decibels. The multifaceted nature of the stress corrosion crack under inspection created a complex eddy current interaction, making it difficult to invert the physical geometry of the crack. It is expected that for simpler defect geometries a benefit in defect characterisation would be observed through dataset mixing.

This work demonstrated the detection of defects in real-time via eddy current data and showed the ability to further post-process the acquired data to enhance image quality. The benefit of being able to post-process the acquired data in such a manner should not be understated, and it is hoped that similar studies such as this can be used to further develop the post-processing of eddy current data to the standard achieved in ultrasonic NDT.

In future work, the authors plan to improve and progress this study by performing eddy current characterisation on multi-angled known defects and comparing the results to simulated datasets; exploring the use of machine learning to automatically classify and characterise defects; and lastly, exploring the fusion of ultrasonic and eddy current datasets.

**Author Contributions:** Conceptualization, E.A.F., S.M., E.N., C.L., M.V., D.L. and E.M.; Data curation, E.A.F. and M.M.; Formal analysis, E.A.F., M.M. and S.M.; Funding acquisition, G.B., R.B., E.M., A.G., G.P. and C.N.M.; Investigation, E.A.F. and E.M.; Methodology, E.A.F. and C.L.; Project administration, E.A.F.; Resources, E.M.; Software, E.A.F., E.N., C.L., M.V. and D.L.; Supervision, G.B., R.B., E.M., A.G., G.P. and C.N.M.; Validation, E.A.F., C.L. and E.M.; Visualization, E.A.F. and E.M.; Writing – original draft, E.A.F., G.B., R.B., M.M., S.M., E.N., C.L., M.V., D.L., E.M., A.G., G.P. and C.N.M.; Writing – review & editing, E.A.F., G.B., R.B., M.M., S.M., E.N., C.L., M.V., D.L., E.M., A.G., G.P. and C.N.M.. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work is supported by the Research Centre for Non-Destructive Evaluation (RCNDE) on behalf of Sellafield Ltd. and NNL Ltd. in the UK under EPSRC Grant No. EP/L015587/1.

**Informed Consent Statement:** Not applicable.

**Acknowledgments:** The authors would like to thank Dirk Engelberg and his team at the Corrosion and Protection Centre within the Research Centre for Radwaste and Decommissioning at the University of Manchester for inducing the stress corrosion cracks on the canister samples.

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

#### **References**

