1. Introduction
Free-space optical (FSO) communication, known as optical wireless communication (OWC), has the advantages of a license-free spectrum, large bandwidth, flexible network and high data rate. However, FSO links are inevitably affected by atmospheric turbulence and adverse meteorological situations, including snow, fog, dust, etc. Radio frequency (RF) communication can provide a reliable link but is sensitive to rainy circumstances, especially for microwave and millimeter wave systems. To address these vulnerabilities, the past decade has witnessed growing interest in the integration of FSO link and RF link, forming hybrid FSO/RF heterogeneous networks. Such amalgamated systems aim to capitalize on the complementary strengths of both communication technologies, thereby enabling more efficient and reliable data transmission [
1,
2,
3,
4].
There are two modes for using FSO/RF links: hard-switching and soft-switching. In soft switching mode, both links are active simultaneously [
5,
6]. Only one link is active in hard switching mode [
7]. For mixed channels, a new throughput maximization algorithm was proposed in [
5] to optimize the bit rate, and the system performance of LDPC codes with regular or irregular structures was analyzed. In [
6], data were simultaneously transmitted over two links at the same rate, and space diversity technology was adopted to maximize spectrum utilization and reduce the influence of the turbulence channel. Through the analysis of outage probability, ref. [
7] studied the performance of a hybrid FSO/intelligent reflecting surface (IRS)-aided RF communication system based on hard switching. Experimental results showed that the outage probability increases with the increase in the switching threshold, and the increase in the signal-to-noise power ratio (SNR) suppressed outage performance. Additionally, deploying more elements in the IRS could result in SNR gain. When hard switching worked, ref. [
8] used a machine learning (ML) algorithm to predict the indicator of received signal strength (RSSI), which proved the reliability of hard switching. RSSI considered the values of the current state and the previous state, and the proper selection of threshold limits for the RSSI parameter was crucial: when the RSSI exceeded the threshold, the FSO link was chosen, whereas when it fell below the threshold, the RF link was selected.
Several researchers have explored channel prediction based on ML algorithms, including hybrid links, especially those combined with adaptive modulation (AM). In [
9], a proposal was made for a method of estimating SNR using artificial neural networks in AM and coding schemes. Power spectral density was used to classify SNR and played a role in adaptive coding and modulation. Once trained, it could determine the optimal adaptive coding and scheme at lower complexity, demonstrating its effectiveness in throughput performance. The possibility of using AM to select and switch modulation modes in hybrid link systems was proposed and verified by using intelligent power control and link switching in [
10]. It demonstrated that not only could the variation trend in RSSI be predicted, but also that power control could effectively reduce switching frequency, thereby enhancing the transmission quality of FSO link. In [
11], the ML algorithm was used to predict channel state information, in which the SNR of the next transmission channel was taken as the prediction target, and the past SNR with other relevant information was treated as the prediction basis. In [
12], the ML model was used to train the amplitude-frequency vector of data symbols with the goal of matching the SNR and achieving SNR estimation. The experiment showed that even with very low SNR, ML could estimate the SNR with very high accuracy, even reducing the mean square error to less than 0.01. Ref. [
13] proposed a hybrid FSO/RF system that manipulates adaptive switching techniques to form IM/DD and coherent heterodyne detection; finally, the expression of outage probability under various atmospheric conditions was obtained and was used as a criterion to compare single FSO systems with hybrid systems. Ref. [
14] proposed a link-switching mechanism based on ML in a hybrid FSO/RF system, the system utilized ML to predict the link margin and achieve link prediction based on weather conditions. As far as we know, the issue of ML-based modulation switching and link threshold for hybrid links considering weather data has not been addressed. This work is based on link thresholds and uses ML to select modulation for communication links based on current atmospheric conditions.
The rest of this paper is organized as follows:
Section 2 examines the proposed a hybrid FSO/RF model of an ML-based switching system and discusses the different atmospheric effects.
Section 3 discusses the ML model in determining link availability, deriving expressions for its spectral efficiency, link budget (LB), and bit error rate (BER).
Section 4 describes the simulation results, and useful concluding remarks are drawn in
Section 5.
3. Threshold Estimation and Data Set Generation by Machine Learning
According to the atmospheric characterization parameters given in the CCSDS141.1-R-1 red book, we gain the original data set
from the website of
https://rp5.ru/(accessed on 1 September 2020 to 1 December 2022), which involves rainfall rate, visibility, temperature, humidity, etc. The LB is used to plan the resource allocation for each modulation and to determine the working mode for hybrid FSO/RF links. The ML model is constructed to learn and predict from the data on rainfall and visibility on rainy and foggy days.
3.1. Construction of the Random Forest Algorithm Model
The random forest algorithm is an ensemble model suitable for classification problems, consisting of multiple decision trees. When training data are input into the model, a subset is randomly selected along with some of its feature attributes to build multiple small decision trees. When unknown data are input, predictions are made for each decision tree, and the final prediction is obtained through a voting process based on the predictions of the decision trees. In this experiment, 500 decision trees are selected, each with a maximum depth of 3.
The model works in two phases: the training part and the testing part. The training part includes defining the optimal range of multiple modulations based on a given target BER and
. Then, their corresponding LB threshold set
can be calculated to allocate the channel state set
. The FSO link has three modulation modes, including L-ary pulse-phase modulation (L-PPM), M-ary phase shift keying (M-PSK), and M-ary quadrature amplitude modulation (M-QAM). The RF link operates predominantly in two modulation modes: M-PSK and M-QAM. In Algorithm 1, the details of calculating link budget by switching modulation is provided.
Algorithm 1: Link budget switching scheme |
|
For the channel state of each modulation (
) and the proper target BER (
), set
is taken into account to construct a relational graph that delineates the correlation between the weather parameter and the average BER (
) across different modulations. The relational graph as shown in
Figure 2 is divided into some areas representing the optimum modulation range when
is smaller than
, and the
is calculated when
is equal to
. And we calculate the LB from arbitrary weather data. A comparison between
and
determines the channel state (
). If it is less than the minimum threshold (
), the communication is interrupted. If it is between
and
, the channel state of the current weather is
.
The second component focuses on constructing a random forest model that utilizes the weather set () and channel state set (). Within the training phase, multiple channel states are formulated as the output labels for the decision trees. The optimal channel state can be obtained by inputting data from real-time weather into the random forest model, but it requires in-depth analysis to determine whether of the channel state () under current weather data is less than . In cases where exceeds , the LB threshold is fine-tuned iteratively until it descends below . This iterative process is performed to produce the channel state congruent with the channel state set () established during the training phase.
3.2. Basic Parameters
3.2.1. Link Budget
LB refers to the remaining signal power or bandwidth in a communication link. Suppose
denotes the LB of the FSO link, and it can be expressed as [
14]
where
is the transmitter power,
is the receiver sensitivity,
is geometrical attenuation,
is system losses, and the atmospheric attenuation in the FSO link is denoted as
, which is a collective representation of
and
. Suppose
denotes the LB of the RF link, and it can be expressed as [
11]
where
is the effective isotropic radiated power,
indicates the receiver antenna gain,
represents the lowest normalized SNR,
R indicates the bit rate of the system, and
k is the Boltzmann constant. The receiver noise temperature is presented as
T,
represents path loss, and
is the atmospheric attenuation under the RF link, given by
and
.
3.2.2. Instantaneous BER
The average BER that reflects the performance of the system is the average value of the instantaneous BER over a time period. For the FSO link, the average BER (
) can be given by the instantaneous BER (
) and the PDF, and is denoted as
Similarly, the instantaneous BER of the RF link can be expressed as
to express
as the average BER of the RF link. Simultaneously, the instantaneous BER of modulation modes can be expressed as
BER of L-PPM:
where L is the symbol order,
.
BER of M-PSK:
where M stands for the modulation length of PSK.
BER of M-QAM: the instantaneous BER of M-QAM is shown as [
17]
3.3. Adaptive Modulation
The spectral efficiency of M-PSK is defined as the data rate transmitted within a given bandwidth, and can be given by [
18]
where
C represents the data rate used for transmission,
W shows the channel bandwidth, the modulation order is represented as
, and
is the probability of receiving SNR in the interval
, which can be expressed as
where
is the cumulative distribution function (CDF) of the turbulence-induced fading. Given
,
can be simplified to
And the spectral efficiency of the M-QAM is as follows [
19]:
The spectral efficiency of the L-PPM is as follows [
20], there is only one optical pulse among the N times slots:
where have only one optical pulse among the
L times slots.
5. Discussion
This paper uses a machine learning-based random forest algorithm to implement a soft-switching strategy in hybrid FSO/RF links and evaluate their performance. The random forest model is established according to the channel parameters, particularly focusing on rain and fog as the primary elements. Because the performance of the hybrid link is limited by severe weather conditions. With a weather data set split into a training part and a test part, the modulation adaptation accuracy performance was simulated. The results show that the training results of each group are consistent with the homologous testing results. The AI-based soft-switching strategy can enhance communication quality and reliability according to real-time environmental weather conditions. Furthermore, by using the trained link budget threshold, a suitable modulation type can be chosen to maximize the system efficiency with double links.
Compared to [
14], this work incorporates considerations for modulation. When determining the link quality model, real-time rain and fog data are used for training, in addition to taking into account the target BER. In order to align with the modulation, the LB is divided to ensure the stability and reliability of link performance. In order to maintain a balance in accuracy, this alignment may lead to a slight decrease in accuracy. Subsequently, we will continue to research this task, striving to achieve 100% prediction accuracy by conducting experimental testing with more external factors taken into account in some practical application scenarios. This will provide strong support for high-speed data transmission.