Signal processing for enhancing railway communication by integrating deep learning and adaptive equalization techniques

Abstract

With the increasing amount of data in railway communication system, the conventional wireless high-frequency communication technology cannot meet the requirements of modern communication and needs to be improved. In order to meet the requirements of high-speed signal processing, a high-speed communication signal processing method based on visible light is developed and studied. This method combines the adaptive equalization algorithm with deep learning and is applied to railway communication signal processing. In this research, the wavelength division multiplexing (WDM) and orthogonal frequency division multiplexing (OFDM) techniques are used, and fuzzy C equalization algorithm is used to softly divide the received signals, reduce signal distortion and interference suppression. The experimental results showed that increasing the step size could reduce the equalization effect, while increasing the modulation parameter will increase the bit error rate. Through deep learning to achieve channel equalization, visible light communication could effectively mitigate multi-path transmission and reflection interference, thereby reducing the bit error rate to the level of 0.0001. Under various signal-to-noise ratios, the system using the channel compensation method achieved the lowest bit error rate. This outcome was achieved by implementing hybrid modulation scheme, including Wavelength division multiplexing (WDM) and direct current-biased optical orthogonal frequency division multiplexing (DCO-OFDM) techniques. It has been proved that this method can effectively reduce the channel distortion when the receiver is moving. This study develops a dependable communication system, which enhances signal recovery, reduces interference, and improves the quality and transmission efficiency of railway communication. The system has practical application value in the field of railway communication signal processing.

1 Introduction

With the significant achievements of deep learning (DL) in the fields of natural language, speech recognition, computer vision, and wireless communication, DL, with its deep neural network structure and robust pattern recognition capability, can extract important features from complex RC signal data, thereby achieving more accurate and reliable signal processing. Due to the growing demand for data transmission, the traditional wireless RC communication technology has been limited, and an emerging high-frequency visible light communication (VLC) technology has attracted widespread attention from researchers [1]. High-frequency VLC technology offers not only high transmission rates, but also abundant spectrum resources. However, there are still technical barriers to its application in RC systems. Signals in the high-frequency band are susceptible to inter-symbol interference, background noise, and object reflections in railway environments, which significantly reduce the reliability and quality of signal transmission [2]. At the same time, the RC system has extremely high security requirements for data transmission, and it is necessary to ensure the confidentiality and integrity of the data. Although traditional adaptive equalization algorithms (AEA) have certain advantages in reducing signal distortion and interference, their effectiveness still needs to be improved when faced with high-frequency VLC and complex environmental noise [3,4]. Therefore, in order to improve the demand rate and reduce the computational complexity, CNN is innovatively combined with WDM-DCO-OFDM VLC offline system. Moreover, the fuzzy C-means clustering method is selected to partition the received signals to ensure better balancing effect. This study aims to contribute to the technological development of RC signal processing, improve the performance and effectiveness of RC signal processing, use DL algorithm to extract important RC signal features, solve various RC signal processing problems, promote processing technology, and provide new solutions for the safety and communication reliability of railway systems. The advantage of the proposed method lies in the combination of CNN to enhance the feature extraction and signal processing capabilities of the system, and the selection of fuzzy C-means clustering to effectively partition the received signal, ensuring a balanced effect in the signal processing process. This provides a novel and efficient solution to the RC signal processing problem in railway systems.

This study is divided into four parts. Section 1 provides an overview of current research on methods for processing VLC and RC signals. Section 2 introduces RC signal processing techniques based on DL and AEA. Section 2.1 is building VLC system based on improved CMA. Section 2.2 is design combining WDM and DCO-OFDM modulation technologies. Section 2.3 is intelligent spectrum design based on CNN. Section 3 introduces experimental results and analysis. Section 3.1 is analysis of simulation results of communication system based on blind equalization algorithm. Section 3.2 is intelligent spectrum analysis results. Section 4 is the conclusion of RC signal processing techniques based on DL and AEA.

2 Related works

VLC is a communication technology that utilizes the light and dark flickering signals emitted by visible light sources such as light emitting diode (LED) to realize information transmission. Due to the low cost and energy saving characteristics of LEDs, VLC has a wide range of applications [5,6]. Louro P et al. proposed an indoor navigation system based on VLC technology and applied the system for geolocation and navigation to solve automated logistics control problems. The communication network of this indoor navigation system used VLC transmitters with tri-color RGB white LEDs, and the uplink channel was established with a single optical signal [7]. Margariti K et al. applied VLC to a vehicle-to-vehicle communication system that combined discrete multi-tone modulation and organic light-emitting diode (OLED) transmitter to overcome the bandwidth limitations of organic devices. Simulation results showed that the vehicle-to-vehicle communication system utilizing OLEDs can achieve several Mbps of data communication, and the data rate can be more than 5 Mbps, and even if the distance reaches 40 m, the data rate of hundreds of kbps can be supported [8]. Although researchers have applied VLC technique in various applications, less research has been done in RC technique applications due to some technical difficulties faced in RC signal processing. Ma Z designed a reconfigurable intelligent surface method for interference suppression in railway wireless communication systems. It converted interference suppression into maximizing the receiving SINR problem and maximizing the expected receiving power, bypassing the channel estimation of interference links. The results showed that this method could ensure the maximum anti-interference gain [9]. Zhou T et al. built a new spatiotemporal channel prediction model that combined CNN and convolutional long short-term memory to optimize intelligent high-speed railways. The model evaluated the spatiotemporal computational complexity and prediction accuracy. The results showed that the model had certain applicability [10].

In order to minimize the effect of VLC on RC in high frequency bands, different researchers have proposed different methods for improvement. Zhang A et al. proposed a new sparse adaptive system identification method to solve the sparse adaptive channel reconstruction problem in time-varying cooperative communication networks. The method utilized the cost function of the LMS algorithm to incorporate a reweighting paradigm penalty for sparsity. Therefore, the weight factor was called the equilibrium parameter for reweighting paradigm adaptive sparse system identification. The experimental results showed that the simulation results were in good agreement with the theoretical analysis, indicating that the proposed algorithm had better convergence speed and better steady state behavior [11]. Yang et al. put forth a dual-mode adaptive switching blind equalization algorithm to address the issue of inter-code interference stemming from the bandwidth constraints of a high-power light source and a high-sensitivity detector, as well as the multipath effects observed in a seawater channel within a high-power underwater wireless optical communication system. The algorithm combined a judgment-directed least-mean-square mode and a variable-step-size normal-mode fractional-interval equalizer based on the attenuation characteristics and time-domain spreading characteristics of underwater wireless optical signals. Experimental results demonstrated that the research-designed dual-mode adaptive switching blind equalization algorithm could improve the performance of long-range UWOC systems [12]. Other researchers have proposed to improve the signal transmission by using OFDM technique, which is a multicarrier modulation technique that utilizes linear superposition in the frequency domain to enable parallel transmission of multiple signals. The introduction of OFDM technique has increased the capacity of the wireless communication system and hence it is widely used in the field of wireless communication. Sahoo P K et al. used hybrid modulation optical orthogonal frequency division multiplexing (OOFDM) technique and nonlinear compressive transform (NCT) to reduce the peak to average power ratio. Numerical simulations were performed to verify the effect on the system bit error rate (BER). The experimental results indicated that the low peak-to-average power ratio and RC signal of the hybrid modulation system were lower compared to the discrete Fourier transform precoding technique [13]. In addition, Wang D et al. constructed a data-driven fiber channel based on the bidirectional long and short-term memory in the DL algorithm, and also investigated the effects of dispersion, data mode, fiber length, fiber nonlinearity, pulse shaping and sampling rate on it. The experimental results showed that the data-driven fiber channel constructed based on the DL algorithm was insensitive to the data size and transmit power as well as independent of the fiber length, which dramatically reduced the computation time of the fiber optic system [14]. The methods proposed by the above researchers can be used as complementary techniques to existing simulation systems or as potential options for future simulation methods.

To enhance the quality and transmission rate of visible light in RC, the study employs a blend of wave division multiplexing and OFDM techniques to create a high-speed communication method in visible light. The DL algorithm extracts the key features of the RC signal and combines them with AEA for the purpose of precise signal recovery and the suppression of interference. The application of this method results in an improvement in the quality and transmission efficiency of the RC.

3 RC signal processing techniques based on DL and AEA

In response to the problems of complex computation and low demand rate in traditional RF wireless communication technology, an improved CMA algorithm is first designed for VLC systems. Then, WDM-DCO-OFDM technology is selected to improve the transmission rate of the system, and the signal is processed using the fuzzy C-equalization algorithm. Considering the low adaptability and high complexity of optical spectra, this study also uses CNN to automatically detect and extract spectral features.

3.1 Building VLC system based on improved CMA

VLC is affected by light and is relatively less affected by environmental noise. RF-based systems are more susceptible to electromagnetic interference and multipath attenuation, and VLC has better anti-interference performance. VLC does not penetrate walls during transmission and therefore has high security, making VLC more attractive in scenarios such as RC where confidentiality is required. RF spectrum resources are limited, while VLC uses the visible light band, resulting in relatively abundant RF resources. In high-density RC environments, VLC can better avoid spectrum congestion problems. In VLC systems, factors that affect the transmission characteristics of the channel include narrow LED modulation bandwidth, multipath due to reflected light, and frequency attenuation of the channel. These factors cause interference, increased RC signals, waveform distortion, and degradation of communication quality [15]. To solve these problems, the signal needs to be processed to compensate for the effects of interference. In this case, the principle of VLC system is shown in Fig 1.

[Figure omitted. See PDF.]

In Fig 1, the study utilizes a VLC system, which consists of three components: LED transmitter, LED receiver and visible light transmission. The transmitting end converts the electrical signal into an optical signal, which is emitted by the LED to the optical wireless channel after going through the modulation circuit and driver circuit. In the optical wireless channel, part of the optical signal is absorbed and the rest reaches the receiving end by scattering or reflection. After the data is processed such as preprocessing and coding modulation, the LEDs are driven to emit optical signals, which are focused by lenses to photodetectors and converted to electrical signals for further processing such as signal processing, demodulation and decoding, and finally the transmitted signals are recovered [16,17]. However, light refraction, reflection, and noise interference from mobile devices and indoor environments can cause severe inter-code interference.

To solve these problems, equalization techniques are used. However, this method can waste spectrum resources and reduce the effectiveness of system transmission. To correct and compensate the signal and reduce the inter-code interference, a module that can be varied according to the system conditions is incorporated into the communication system. This module is known as an equalizer, and an adaptive equalizer is capable of adapting channel characteristics based on training sequences to achieve channel estimation and equalization. The equalizer uses filters to filter out noise and improve the signal quality so that the adjudicator can correctly determine the signal [18]. Its equivalent model is shown in Fig 2.

[Figure omitted. See PDF.]

In Fig 2, {a_n} represents the input sequence, n(t) represents the additive Gaussian white noise, and represents the functions corresponding to the transmit filter, channel, receive filter, and equalizer, respectively. According to the equivalent model of the equalizer, the transmission characteristics of the channel are shown in Eq (1).

(1)

In Eq (1), H(w) represents the transmission estimate of the channel, where the equalizer’s transfer function G_E(w) needs to satisfy Eq (2).

(2)

In Eq (2), e^−jφ(w) represents the error signal. The adaptive equalizer generally contains a training phase and a usage phase, where the sender chooses the length of the training sequence and the receiver continuously adjusts the equalizer parameters according to the training sequence to compensate for signal impairments. However, care must be taken when designing the system to obtain the correct filter coefficients based on the training sequences even in the case of poor communication quality. The equalizer must be trained periodically and repeatedly to effectively eliminate inter-code interference, so the user data is divided into multiple time slots for transmission. Traditional adaptive equalization techniques have reduced the RC signal, but the algorithms require a large number of iterations to converge, tend to amplify the gradient noise, and occupy the channel bandwidth, reducing the RC channel’s adaptive capability. The improved CMA uses a priori information to produce filtering results that are close to the desired recovery results, which improves the convergence speed of the system and the tracking characteristics of the channel. Blind equalization is an adaptive equalization technique that continuously constructs a network with channel characteristics through the a priori information of the received sequence, effectively restores the signal and compensates for channel time variability, reduces mutual interference between symbols, and improves the system communication performance [19]. The process of computing balanced signals necessitates the initial collection of received signal data, followed by the determination of channel characteristics, the selection of suitable equalization algorithms, the implementation of equalization processing, and the assessment of the equalization effect. The CMA algorithm selected for the study has good performance and will not occupy bandwidth, and its principle is to use the a priori information of the received sequence as the cost function by adjusting the network parameters in order to approximate the transmitted sequence. The cost function is shown in Eq (3).

(3)

In Eq (3), y(n) represents the nth subcarrier signal at the receiver, E represents the likelihood function to compute the base coefficient vector, P represents the cost function in the CMA algorithm, α represents a constant, and α satisfies Eq (4).

(4)

In Eq (4), x(n) represents the transmit signal, and the iterative equation for the weight coefficients according to the gradient descent algorithm CMA is shown in Eq (5).

(5)

In Eq (5), λ represents the step size. The advantages of the CMA algorithm are that it does not require a very long training sequence, a priori information can be obtained through multiple means, and the algorithm is highly scalable. However, during normal communication, if a fault occurs, the training sequence must be sent to train the receiver before the start of the next communication. However, these characteristics do not meet the needs of modern digital communication systems for higher speed rates, higher capacity, higher validity and greater reliability. Therefore, there is a need to speed up signal processing and to provide new ideas for correcting signals using their statistical properties or other a priori information rather than using training sequences. The CMA-based blind equalization has certain advantages in handling constant modulus amplitude characteristic signals and severe multipath interference.

3.2 Design combining WDM and DCO-OFDM modulation technologies

OFDM is a type of multi-carrier modulation (MCM) that uses frequency division multiplexing to achieve parallel transmission of high-speed serial data. It has good resistance to multipath fading and can support multi-user access. OFDM has many advantages in high-speed data transmission, anti-interference capability, and spectrum utilization efficiency, and is widely used in various wireless communication systems. Meanwhile, OFDM systems can compensate for distortion in signal transmission by simple subcarrier equalization. However, conventional OFDM communication in VLC cannot be applied to intensity modulation because it is based on complex form and bipolar type signals. To employ intensity modulation in visible light systems, specific transformations, such as conjugate stacking and the addition of DC bias, are necessary to convert conventional complex form and bipolar type signals to a real form. In VLC system, high speed communication can be realized by combining WDM and DCO-OFDM modulation techniques. The WDM technique modulates the data stream to be transmitted onto optical signals with different frequencies for transmission, and the WDM technique can greatly increase the transmission capacity of the system [20]. The OFDM technique is a coding technique that converts high-speed serial data streams into multiple parallel data streams with the advantages of selective fading resistance and high utilization. The combination of these two techniques applied in VLC system can further improve the transmission rate of the system. In this VLC system, an unsupervised learning blind equalization method is designed and verified for achieving the equalization effect during high-speed communication. The design schematic of this system is shown in Fig 3.

[Figure omitted. See PDF.]

In Fig 3, the spectrum efficiency of VLC systems is increased by three times through DCO-OFDM. The system schematic includes coding, modulation and demodulation modules for data generation and OFDM, as well as signal processing modules. The signal processing module is mainly analyzed and processed on the computer, the waveform generator generates any waveform and inputs it into the computer to complete 16-QAM modulation. The complex number should be transformed into a time-angle processing signal, which can then be converted into multiple channels using IFFT for high-speed real number signals. Meanwhile, the equalization method module uses an unsupervised learning based equalizer to classify the data by fuzzy C-mean clustering method and applies a blind equalization algorithm to update the weight coefficients vector of the equalizer to achieve equalization [21]. The unsupervised learning based channel equalizer uses fuzzy C-mean clustering method to classify the data. The fuzzy C-means clustering process is illustrated in Fig 4.

[Figure omitted. See PDF.]

In Fig 4, the fuzzy C-means clustering algorithm needs to initialize the membership matrix, clearly separate the signals belonging to different clusters, and find and calculate the initial cluster center to achieve the goal of optimal selection. By iteratively updating and adjusting the clustering center and membership matrix, a soft division of the data is achieved and the centroid and variance of the clusters are output, thus minimizing the objective function. A reduction in the equalizer output value will result in an enhanced clustering effect, as indicated by a more compact constellation diagram of the signal. This, in turn, will facilitate more accurate signal judgment. In summary, the application of WDM and DCO-OFDM modulation techniques to VLC systems can improve the transmission rate and capacity. Meanwhile, the blind equalization method with unsupervised learning can achieve good equalization effect in high-speed communication.

3.3 Intelligent spectrum design based on CNN

Existing machine learning models for processing raw natural data require high-resolution spectral sampling due to the need to preserve as much of the original information as possible, which leads to a higher dimensionality of the raw data in the training set, making the mapping network structure complex, increasing algorithmic complexity, and decreasing real-time processing speed. Therefore, a novel approach is needed to build models with self-learning and evolutionary capabilities that can automatically detect and analyze desired features from images and further train the network structure. Such a model can replace human engineers for intelligent analysis and is highly versatile and adaptive, independent of the dimensionality of the training data. DL is a multilayer nonlinear deep neural network that proposes an automatic feature learning method in the process of model building, reducing the incompleteness caused by manually designed features.

The study applies the DL method to the measurement of optical communication parameters in spectral images, creating a method to detect optical communication parameters efficiently. Convolutional neural network (CNN) consists of two parts: feature extraction layer and feature mapping layer. The feature extraction layer automatically acquires feature vectors, including convolutional and pooling layers, where the input of each convolutional layer neuron is connected to the local receptive domain of the previous layer and extracts the features of that localization [22]. The mapping structure of the CNN mainly consists of two modules, convolution and pooling, in which the convolutional module extraction process is shown in Fig 5.

[Figure omitted. See PDF.]

In Fig 5, the 2×2 local in the upper left corner of the convolution kernel, i.e. (3,0,1,1), completes the convolution operation. Subsequently, the black square is shifted to the right, and the convolution kernel and (0,4,1,6) are used to perform the convolution operation, and so on. Subsampling from the convolutional feature map to the pooled feature map ensures that the processed image does not lose important features and is reduced in size. The information output from the mapping network contains a variety of performance features, and the mapping network is able to automatically learn the desired features in the spectral data through multiple iterations. The test set data to be analyzed is input into the trained mapping network, and the current input spectrum data is analyzed for specific performance through the network’s past learning experience. The CNN uses a locally connected and shared weights network structure, a design that reduces the number of weights, allowing the network structure to reduce the complexity of the model while retaining useful information [23].

The subsequent pooling layer is used to reduce the dimensionality of the data and facilitate classification. The feature mapping layer uses a traditional neural network structure with multiple fully connected hidden layers to classify the data using the features obtained from the feature extraction layer and provides the classification results through the output layer. Data analysis from an image perspective has a unique advantage because the raw data form two-dimensional spatial features in image space, i.e., the data points form local features with other data points in a small area around them. These local features can be used to distinguish between categories, e.g., by looking at the leaves of a plant it can be inferred that it belongs to a certain kind of plant rather than plastic [24,25]. The concept of CNNs is based on the recognition of these local features. In calculus, CNN can be expressed by the equation, see Eq (6).

(6)

In Eq (6), S(t) represents the signal received by the photodetector, x(t−a) represents the signal emitted after modulation, and w(a) represents the channel impulse response. The corresponding discrete form of Eq (6) is shown in Eq (7).

(7)

The matrix is used to represent Eq (7), the signal matrix A emitted after modulation is convolved with the channel impulse response matrix B, as shown in Eq (8).

(8)

In Eq (8), * represents the convolution operation, which can be expressed as a two-dimensional convolution when there are two variables in equation The form is shown in (9).

(9)

Eq (9) is computationally converted to Eq (10).

(10)

When in CNN, the convolution equation is changed, Eq (10) can be expressed as (11).

(11)

The CNN-based DL intelligent spectrogram analysis method includes the following steps: obtaining the set of spectral images to be analyzed, preprocessing the spectral images, training the CNN module, inputting the spectral images to be analyzed to the trained network for feature extraction and performance analysis, and outputting the results of the analysis. The analysis steps are shown in Fig 6.

[Figure omitted. See PDF.]

In Fig 6, when the spectrogram is analyzed, the spectral image under the desired analysis situation is first generated by the frequency domain measurement module, which is used to construct the training set of spectral data. Both the training set and the test set are combined from the images and the corresponding performance parameter label matrices. Moreover, each set of training set data includes the spectral images under different cases and their corresponding performance index parameters. The color optical spectrogram images are then converted into grayscale images, and the grayscale images are downsampled to improve the generalization ability, and the analysis structure is schematically shown in Fig 7.

[Figure omitted. See PDF.]

In Fig 7 spectrum image preprocessing, after converting the collected original color spectrum images to grayscale images, the original spectrum is downsampled to make the pixel size of the original spectrum reduced, and the processed training dataset is inputted to the spectrum map training dataset of the neural network. Finally, the CNN module is trained so that it can extract features from the input spectrum image and perform performance analysis. After the training is completed, the preprocessed spectrogram training set of the spectrogram to be analyzed is input into the established mapping structure. The details of the proposed model are shown in Table 1.

[Figure omitted. See PDF.]

4 Experimental results and analysis

In order to test the superiority of the proposed signal processing model, the study firstly conducts the simulation test of the communication system, so as to understand the theoretical feasibility of the system initially. Secondly, it analyzes the application effect of smart spectrum to verify the high performance of the RC signal processing system designed by the institute.

4.1 Analysis of simulation results of communication system based on Blind equalization algorithm

VPI software is called VPI TransmissionMaker, which is used for the construction and simulation simulation of visible light system, fiber optic communication and other systems of RC, as well as performance analysis. The experimental dataset is sourced from the mobile communication open data platform, which provides publicly available datasets on the field of mobile communication for research and analysis. 10872 RC related data are collected from the platform to form a dataset. To ensure that the training and testing of the model are balanced and representative, the experimental training and testing sets are divided in an 8:2 ratio. The dataset simulation system is based on VPI 9.0, and co-simulation with Matlab, in which VPI TransmissionMaker is used to realize the construction of the system platform and the transmission of signals. The signal processing is realized by calling the Matlab code, and the Cosimulation Interface is used for the modulator. The co-simulation interface uses Cosimulation Interface. The modulator is modulated by coherent detection. The broadband of the receiver is also 30 GHz, and the collected spectral data are sent to the spectrum analysis module for analysis. The experimental simulation parameters of the system are shown in Table 2.

[Figure omitted. See PDF.]

The parameters required for the experiment can be seen in Table 2, in which there are two photodetectors, two photodetectors are combined through a certain angle to form a hybrid optical receiver, and the hybrid receiver detects the received signals. To ascertain the impact of varying step sizes on the convergence performance of the CMA algorithm, the study employs four step sizes (λ = 0.008, λ = 0.006, λ = 0.004, and λ = 0.002) as experimental parameters. The number of iterations is utilized as the horizontal coordinate, while the mean-square error (MSE) between the labeled input and output values of the equalizer serves as the vertical coordinate. The experimental simulation results obtained are shown in Fig 8.

[Figure omitted. See PDF.]

In Fig 8, that the MSE is about 0.01, 0.02, 0.05 and 0.07 when the step size of 0.002, 0.04, 0.006 and 0.008 are chosen, respectively. As the step size increases, the algorithm’s convergence speed will be accelerated, and the mean-square error increases. Moreover, the equilibrium effect is significantly lower than that of the convergence effect with a smaller step size. With a step size of 0.002, the CMA algorithm reaches convergence around 1000 times, at which time the algorithm is more capable of tracking channel time variations. Choosing the appropriate step size value in the RC system can achieve a balance between convergence speed, accuracy, and stability. The optimal step size value of 0.002 can be effectively applied in actual RC systems, contributing to the improvement of system performance and the stability, reliability, and security of data transmission. Multiple different SNR values are used during the training process to generate training samples, covering a wide range of SNRs. When constructing the training dataset, the range of SNR values is set to 0dB~30dB. To ensure the robustness and generalization ability of the model, the training data is uniformly sampled throughout the entire SNR range. In addition to the step size thought, different modulation orders have an impact on the performance of the system. Four different modulation orders are selected, with the signal-to-noise ratio (SNR) as the horizontal coordinate and the BER as the vertical coordinate. The results are shown in Fig 9.

[Figure omitted. See PDF.]

In Fig 9, when the four modulation parameters are increased sequentially from OOK, 4PAM, 8PAM, and 16PAM, the RC signal of the system increases for the same SNR. However, as the SNR increases, the RC signal reaches a certain value and stops decreasing. When the RC signal is at the same value, the SNR is lower for OOK modulation. This indicates that the relationship between various modulation parameters and the SNR can help optimize the RC system design and improve the reliability and stability of data transmission. Under the same modulation format, the research designed DL-based blind equalization DL-CMA algorithm is compared with the conventional CMA algorithm, and the D-LSTM algorithm in the literature [25], and the variation of BER is shown in Fig 10.

[Figure omitted. See PDF.]

In Fig 10, the study proposes that the adaptive blind equalization method of DL optimizes the equalizer by iteratively calculating to find the extreme points of the error function. The method applies unsupervised learning to the channel equalization of VLC systems by equalizing the sampling sequence and outputting the final judgment. This effectively reduces the inter-coder interference caused by phenomena such as multipath transmission and reflections in VLC systems, and significantly reduces the RC signal to 0.0001 level. Although the processing speed is slow, it does not affect the timely transmission to achieve the maximum communication rate while realizing high-quality communication. This method realizes the self-learning and self-updating of the equalizer, which ensures the communication quality and reduces the system complexity. The module may be utilized as a software processing component of the channel equalizer or a channel equalization module of the simulation software. It can also be embedded into the signal receiver, thereby enabling the realization of intelligent signal equalization and optical performance compensation. This ultimately facilitates the automation and intelligence of the equalization function of signal reception. In addition to BER and MSE performance, the complexity of the algorithm is also an important measure for evaluating the performance of channel estimation algorithms. For RC, the fast or slow convergence speed of the algorithm affects its performance. The railway scenario, as a special case of data transmission, aims to analyze the data transmission performance in the railway environment. The railway system has unique characteristics, such as high-speed movement and signal crossing tunnels, and other special environments that may affect data transmission. Therefore, the number of iterations of this algorithm under different number of subcarriers is evaluated by simulation experiments and the results are shown in Fig 11.

[Figure omitted. See PDF.]

In Fig 11, the DL-based blind equalization DL-CMA algorithm requires fewer iterations to converge for different numbers of subcarriers. In addition, the number of iterations is inversely related to the SNR. When the SNR is large, there is less noise interference, and therefore the number of iterations required to reach the convergence condition is relatively small. Increasing the number of subcarriers in a single OFDM-IM symbol provides more historical channel information, which leads to a reduction in the number of iterations. Overall, the number of iterations of the BEM-EM channel estimation algorithm is within the acceptable range for HSR communications.

4.2 Intelligent spectrum analysis results

The labels of each set of test data are obtained by feeding the spectrum data to be tested into a mapping network that has been trained, and the results are validated. The bit error rate of the optical signal to noise ratio (OSNR) prediction system is a key performance parameter of optical networks, and its measurement and calibration can be achieved through interpolation. Wavelength is generally used to represent a specific frequency range of optical signals in optical communication systems, and wavelength estimation can determine the wavelength information of optical signals. Bandwidth recognition is usually used to determine the spectral characteristics of a signal, so that the system can configure and adjust parameters appropriately. To verify the advantages of the proposed DL-CMA method in spectrum analysis, a comparison is made between the multiple input multiple output (MIMO) equalization technique, recursive least squares (RLS) algorithm, and extreme learning machine (ELM) equalization technique. The results are shown in Fig 12.

[Figure omitted. See PDF.]

Fig 12(A) represents the recognition accuracy of the four algorithms on the spectral parameters when the performance parameter is OSNR. At this time, the highest recognition accuracy is DL-CMA algorithm with 98.41% recognition accuracy, and the lowest recognition accuracy is RLS with 67.76% recognition accuracy. Fig 12(B) graph represents the recognition accuracy of the four algorithms on the spectral parameters when the performance parameter is wavelength. At this time, the highest recognition accuracy is DL-CMA algorithm with 97.24% recognition accuracy, and the lowest recognition accuracy is ELM with 68.75% recognition accuracy. Fig 12(C) graph represents the recognition accuracy of the four algorithms for the spectrum parameters when the performance parameter is broadband. At this time, the highest recognition accuracy is the DL-CMA algorithm, with 97.87% recognition accuracy. The lowest recognition accuracy is the RLS, with 65.26% recognition accuracy. The research designed DL-CMA algorithm has the optimal recognition accuracy in OSNR, wavelength, and broadband performance parameters. In addition to the recognition accuracy of the spectral parameters, the processing speed of the spectrum analysis is also explored. The test time results of the four algorithms in the three performance parameters of OSNR, wavelength estimation, and bandwidth recognition are shown in Fig 13.

[Figure omitted. See PDF.]

In Fig 13(A), the test times of the four algorithms, RLS, ELM, MIMO, and DL-CMA, are 0.588s, 0.623s, 0.682s, and 0.418s, respectively. In Fig 13(B), the test times of the four algorithms, RLS, ELM, MIMO, and DL-CMA, are 0.513s, 0.536s, 0.627s, and 0.376s. Fig 13(C) shows that the test times of the four algorithms, RLS, ELM, MIMO, and DL-CMA, are 0.489s, 0.461s, 0.616s, and 0.358s, respectively. In the scenarios of OSNR, wavelength, and broadband, the DL-CMA algorithm has the shortest testing time. Compared to the other three algorithms, it has the highest efficiency.

To verify the effectiveness of the system designed based on the channel compensation method in the combination of WDM and DCO-OFDM modulation techniques, simulation experiments are conducted on Matlab software by selecting the modulation mode as 16 PAM modulation to simulate the change of channel compensation on the system under different mobile scenarios. The system with channel compensation, the system without channel compensation, and a CNN system combining independent convolution and hard thresholding from literature [24] are compared under mobile scenarios and the results are shown in Fig 14.

[Figure omitted. See PDF.]

The graph from Fig 14(A) represents the moving trajectory of the receiver and system 1 in Fig 14(B)–14(D) represents the system without using channel compensation. System 2 represents the CNN system combining independent convolution and hard thresholding from the literature [24] and system 3 represents the system using channel compensation in this study. In Fig 14, system 3 has the best BER performance for the three different SNR scenarios, which means that the equalization method proposed in this study has a lower RC signal. The system designed based on the channel compensation method in combination of WDM and DCO-OFDM modulation techniques effectively mitigates the distortion caused by the channel on the receiver side during its movement.

5 Conclusion

The study investigates the challenge of using AEA and DL to address the impact of RC signals on high-frequency band VLC. DL is utilized to extract valuable features from complex RC signal data, while AEA is employed to lessen signal distortion and suppress interference. The experiment determined that a larger step size expedited algorithm convergence. However, the increased mean square error yielded lower equalization effectiveness compared to smaller step sizes, which prioritized convergence. Additionally, under equivalent SNRs, the system’s RC signal amplified as the modulation parameter increased. After performing channel equalization using DL, the inter-coder interference caused by multipath transmission and reflections in VLC systems could be significantly reduced, leading to a noticeable decrease in the RC signal to the 0.0001 level. The study’s DL-CMA algorithm exhibited the highest recognition accuracy among three performance parameters: OSNR, wavelength, and broadband, achieving rates of 98.41%, 97.24%, and 97.87%, respectively. In terms of testing time, all four algorithms were tested for the three different performance parameter scenarios in under one second. Notably, the DL-CMA algorithm recorded the shortest testing time. When comparing the impact of channel compensation in communication systems, it was discovered that the proposed channel compensation-based method suggested by the study generated the lowest RC signals in the system. This system was developed via the combination of DCO-OFDM modulation and WDM techniques, successfully reducing channel-induced distortions during receiver movement. The findings were based on analysis of three distinct SNRs. The study demonstrates that the RC system is dependable, and integrating AEA and DL fusion into RC signal processing can enhance signal processing performance and effectiveness. However, the study did not consider railway communication signals in mountainous and high-altitude areas. Given the complex terrain and mountainous environment of such areas, it is possible that multipath propagation of signals may be more pronounced, resulting in more severe signal attenuation and reflection interference. In the future, it would be beneficial to collect railway communication data from mountainous and high-altitude areas in order to increase the diversity of the dataset. This would ensure the generalization ability of the proposed model in practical application scenarios in mountainous and high-altitude areas.

Supporting information

S1 Dataset.

https://doi.org/10.1371/journal.pone.0311897.s001

(DOC)

References

1. 1. Sindhuja R. and Shankar A R. "Effective PAPR reduction in SCFDM-based massive MIMO system using binary crow search algorithm for visible light communication towards 5G networks," IET Communications, vol. 14, no. 16, pp. 2780–2785, September, 2020,

* View Article

* Google Scholar

2. 2. Mu X., Yan Z., Yu Y., Yan H. and Han D. "A tunable self-mixing chaotic laser based on high frequency electro-optic modulation," Opto Laser Technol, vol. 127, no. 14, p. 106172, July, 2020,

* View Article

* Google Scholar

3. 3. Prodi N., Pellegatti M. and Visentin C. "Effects of type of early reflection, clarity of speech, reverberation and diffuse noise on the spatial perception of a speech source and its intelligibility," J. Acoust. Soc. Am, vol. 151, no. 5, pp. 3522–3534, May, 2022, pmid:35649908

* View Article

* PubMed/NCBI

* Google Scholar

4. 4. Hger C. and Pfister HD. "Physics-Based Deep Learning for Fiber-Optic Communication Systems," IEEE J SEL AREA COMM, vol. 99, no. 11, pp. 1–1, November, 2020,

* View Article

* Google Scholar

5. 5. Abrar S., Zerguine A. and Abed-Meraim K. "Adaptive algorithms for blind channel equalization in impulsive noise," IEEE Signal Process Mag, vol. 201, no. 5, p. 108626, May, 2022,

* View Article

* Google Scholar

6. 6. Zhang Y., Tian S. and Pan H. "Adaptive blind equalization of fast time-varying channel with frequency estimation in impulsive noise environment," IET COMMUN, vol. 15, no. 11, pp. 1507–1517, November, 2021,

* View Article

* Google Scholar

7. 7. Louro P., Vieira M. and Vieira M A. "Geolocalization and navigation by visible light communication to address automated logistics control," Opto Eng, vol. 61, no. 1, pp. 016104.1–016104.16, January, 2022,

* View Article

* Google Scholar

8. 8. Margariti K. and Kamalakis T. "Performance analysis of vehicle-to-vehicle communications based on organic light emitting diodes and discrete multitone modulation," Opto Eng, vol. 60, no. 4, pp. 046106.1–046106.16, April, 2021,

* View Article

* Google Scholar

9. 9. Ma Z., Wu Y., Xiao M., Liu G. and Zhang Z. "Interference suppression for railway wireless communication systems: A reconfigurable intelligent surface approach." TVT, vol. 70, no. 11, pp. 1593–11603, 2021,

* View Article

* Google Scholar

10. 10. Zhou T., Zhang H., Ai B., Xue C. and Liu L. "Deep-learning-based spatial–temporal channel prediction for smart high-speed railway communication networks." IEEE TWC, vol. 21, no. 7, pp. 5333–5345. 2022,

* View Article

* Google Scholar

11. 11. Zhang A., Liu P., Ning B. and Zhou Q. "Reweighted lp constraint LMS-based adaptive sparse channel estimation for cooperative communication system," IET Commun, vol. 14, no. 9, pp. 1384–1391, June, 2020,

* View Article

* Google Scholar

12. 12. Yang Y., Yang F., Liu W., Zhao P., Zhang J., He F. et al. "Performance analysis of dual-mode adaptive switching blind equalization algorithm in an underwater wireless optical communication system," Opt. Eng, vol. 60, no. 12, p. 126104.13–126104.13, December, 2021,

* View Article

* Google Scholar

13. 13. Sahoo P. K. "Hybrid mapped optical-OFDM using non-linear companding technique for indoor visible light communication application," IET Commun, vol. 14, no. 17, pp. 3073–3079, October, 2020,

* View Article

* Google Scholar

14. 14. Wang D., Song Y., Li J., Qin J. and Anthony C. Boucouvalas, "Data-driven Optical Fiber Channel Modeling: A Deep Learning Approach," J. Light. Technol, vol. 38, no. 17, pp. 4730–4743, May, 2020,

* View Article

* Google Scholar

15. 15. Tao Q., Li Y. and Zhong C. "A Novel Interference Cancellation Scheme for Bistatic Backscatter Communication Systems," IEEE Commun. Lett, vol. 25, no. 6, pp. 2014–2018, March, 2021,

* View Article

* Google Scholar

16. 16. Wei L. Y., Liu Y. and Chow C. W. "6.915-Gbit/s white-light phosphor laser diode-based DCO-OFDM visible light communication (VLC) system with functional transmission distance," Electron Lett, vol. 56, no. 18, pp. 945–947, July, 2020,

* View Article

* Google Scholar

17. 17. Choudhuri S., Adeniye S. and Sen A. "Distribution Alignment Using Complement Entropy Objective and Adaptive Consensus-Based Label Refinement For Partial Domain Adaptation," Artificial Intelligence and Applications, vol. 1, no. 1, pp. 43–51, January, 2021, https://doi.org/10.47852/bonviewAIA2202524

* View Article

* Google Scholar

18. 18. Jeeva M. P. A., Nagarajan T. and Vijayalakshmi P. "Adaptive multi-band filter structure-based far-end speech enhancement," IET Signal Processing, vol. 14, no. 5, pp. 288–299, May, 2020,

* View Article

* Google Scholar

19. 19. Nishimori K., Maruta K., Hiraguri T. and Shiomi H. "Fast-Converging Constant Modulus Algorithm with Variable Step Size for Multibeam Massive MIMO," IEICE Trans. Commun, vol. 105, no. 7, pp. 1–8, July, 2022,

* View Article

* Google Scholar

20. 20. Kodama T., Shimada K. and Matsumoto R. "Frequency-Packed Multiband-Coherent Transceiver With Symbol Rate-Adaptive Nyquist WDM Signals," IEEE Photonic Tech L, vol. 33, no. 21, pp. 1205–1208, November, 2021,

* View Article

* Google Scholar

21. 21. Giri M. K. and Majumder S. "Eigenvalue-based cooperative spectrum sensing using kernel fuzzy c-means clustering," Digit Signal Process, vol. 111, no. 2, p. 102996, February, 2021,

* View Article

* Google Scholar

22. 22. Guo L. "SAR image classification based on multi-feature fusion decision convolutional neural network," IET Image Processing, vol. 16, no. 1, pp. 1–10, January, 2021,

* View Article

* Google Scholar

23. 23. Yang R., Zha X. and Liu K. "A CNN model embedded with local feature knowledge and its application to time-varying signal classification," Neural Netw, vol. 142, no. 1, pp. 564–572, January, 2021, pmid:34343780

* View Article

* PubMed/NCBI

* Google Scholar

24. 24. Feng J., Chen J. and Sun Q. "Convolutional Neural Network Based on Bandwise-Independent Convolution and Hard Thresholding for Hyperspectral Band Selection," IEEE Trans Cybern, vol. 51, no. 9, pp. 4414–4428, September, 2021, pmid:32598287

* View Article

* PubMed/NCBI

* Google Scholar

25. 25. Sun T., Shao C., Liao H. and Ding S. and Xu X. "Research on adaptive local feature enhancement in convolutional neural networks," IET Image Processing, vol. 14, no. 8, pp. 4306–4315, February, 2021,

* View Article

* Google Scholar

Citation: Wang Y, Chang W, Li J, Yang C (2024) Signal processing for enhancing railway communication by integrating deep learning and adaptive equalization techniques. PLoS ONE 19(10): e0311897. https://doi.org/10.1371/journal.pone.0311897

About the Authors:

Yucai Wang

Roles: Formal analysis, Funding acquisition, Software

Affiliation: Department of Rail Transit, Shijiazhuang Institute of Railway Technology, Shijiazhuang, China

Wei Chang

Roles: Investigation, Methodology, Writing – original draft

E-mail: [email protected]

Affiliation: Department of Rail Transit, Shijiazhuang Institute of Railway Technology, Shijiazhuang, China

ORICD: https://orcid.org/0009-0006-1271-028X

Jingjiao Li

Roles: Methodology, Software, Writing – original draft

Affiliation: Department of Rail Transit, Shijiazhuang Institute of Railway Technology, Shijiazhuang, China

Cuilei Yang

Roles: Data curation, Software, Writing – original draft

Affiliation: Department of Rail Transit, Shijiazhuang Institute of Railway Technology, Shijiazhuang, China

References

1. Sindhuja R. and Shankar A R. "Effective PAPR reduction in SCFDM-based massive MIMO system using binary crow search algorithm for visible light communication towards 5G networks," IET Communications, vol. 14, no. 16, pp. 2780–2785, September, 2020,

2. Mu X., Yan Z., Yu Y., Yan H. and Han D. "A tunable self-mixing chaotic laser based on high frequency electro-optic modulation," Opto Laser Technol, vol. 127, no. 14, p. 106172, July, 2020,

3. Prodi N., Pellegatti M. and Visentin C. "Effects of type of early reflection, clarity of speech, reverberation and diffuse noise on the spatial perception of a speech source and its intelligibility," J. Acoust. Soc. Am, vol. 151, no. 5, pp. 3522–3534, May, 2022, pmid:35649908

4. Hger C. and Pfister HD. "Physics-Based Deep Learning for Fiber-Optic Communication Systems," IEEE J SEL AREA COMM, vol. 99, no. 11, pp. 1–1, November, 2020,

5. Abrar S., Zerguine A. and Abed-Meraim K. "Adaptive algorithms for blind channel equalization in impulsive noise," IEEE Signal Process Mag, vol. 201, no. 5, p. 108626, May, 2022,

6. Zhang Y., Tian S. and Pan H. "Adaptive blind equalization of fast time-varying channel with frequency estimation in impulsive noise environment," IET COMMUN, vol. 15, no. 11, pp. 1507–1517, November, 2021,

7. Louro P., Vieira M. and Vieira M A. "Geolocalization and navigation by visible light communication to address automated logistics control," Opto Eng, vol. 61, no. 1, pp. 016104.1–016104.16, January, 2022,

8. Margariti K. and Kamalakis T. "Performance analysis of vehicle-to-vehicle communications based on organic light emitting diodes and discrete multitone modulation," Opto Eng, vol. 60, no. 4, pp. 046106.1–046106.16, April, 2021,

9. Ma Z., Wu Y., Xiao M., Liu G. and Zhang Z. "Interference suppression for railway wireless communication systems: A reconfigurable intelligent surface approach." TVT, vol. 70, no. 11, pp. 1593–11603, 2021,

10. Zhou T., Zhang H., Ai B., Xue C. and Liu L. "Deep-learning-based spatial–temporal channel prediction for smart high-speed railway communication networks." IEEE TWC, vol. 21, no. 7, pp. 5333–5345. 2022,

11. Zhang A., Liu P., Ning B. and Zhou Q. "Reweighted lp constraint LMS-based adaptive sparse channel estimation for cooperative communication system," IET Commun, vol. 14, no. 9, pp. 1384–1391, June, 2020,

12. Yang Y., Yang F., Liu W., Zhao P., Zhang J., He F. et al. "Performance analysis of dual-mode adaptive switching blind equalization algorithm in an underwater wireless optical communication system," Opt. Eng, vol. 60, no. 12, p. 126104.13–126104.13, December, 2021,

13. Sahoo P. K. "Hybrid mapped optical-OFDM using non-linear companding technique for indoor visible light communication application," IET Commun, vol. 14, no. 17, pp. 3073–3079, October, 2020,

14. Wang D., Song Y., Li J., Qin J. and Anthony C. Boucouvalas, "Data-driven Optical Fiber Channel Modeling: A Deep Learning Approach," J. Light. Technol, vol. 38, no. 17, pp. 4730–4743, May, 2020,

15. Tao Q., Li Y. and Zhong C. "A Novel Interference Cancellation Scheme for Bistatic Backscatter Communication Systems," IEEE Commun. Lett, vol. 25, no. 6, pp. 2014–2018, March, 2021,

16. Wei L. Y., Liu Y. and Chow C. W. "6.915-Gbit/s white-light phosphor laser diode-based DCO-OFDM visible light communication (VLC) system with functional transmission distance," Electron Lett, vol. 56, no. 18, pp. 945–947, July, 2020,

17. Choudhuri S., Adeniye S. and Sen A. "Distribution Alignment Using Complement Entropy Objective and Adaptive Consensus-Based Label Refinement For Partial Domain Adaptation," Artificial Intelligence and Applications, vol. 1, no. 1, pp. 43–51, January, 2021, https://doi.org/10.47852/bonviewAIA2202524

18. Jeeva M. P. A., Nagarajan T. and Vijayalakshmi P. "Adaptive multi-band filter structure-based far-end speech enhancement," IET Signal Processing, vol. 14, no. 5, pp. 288–299, May, 2020,

19. Nishimori K., Maruta K., Hiraguri T. and Shiomi H. "Fast-Converging Constant Modulus Algorithm with Variable Step Size for Multibeam Massive MIMO," IEICE Trans. Commun, vol. 105, no. 7, pp. 1–8, July, 2022,

20. Kodama T., Shimada K. and Matsumoto R. "Frequency-Packed Multiband-Coherent Transceiver With Symbol Rate-Adaptive Nyquist WDM Signals," IEEE Photonic Tech L, vol. 33, no. 21, pp. 1205–1208, November, 2021,

21. Giri M. K. and Majumder S. "Eigenvalue-based cooperative spectrum sensing using kernel fuzzy c-means clustering," Digit Signal Process, vol. 111, no. 2, p. 102996, February, 2021,

22. Guo L. "SAR image classification based on multi-feature fusion decision convolutional neural network," IET Image Processing, vol. 16, no. 1, pp. 1–10, January, 2021,

23. Yang R., Zha X. and Liu K. "A CNN model embedded with local feature knowledge and its application to time-varying signal classification," Neural Netw, vol. 142, no. 1, pp. 564–572, January, 2021, pmid:34343780

24. Feng J., Chen J. and Sun Q. "Convolutional Neural Network Based on Bandwise-Independent Convolution and Hard Thresholding for Hyperspectral Band Selection," IEEE Trans Cybern, vol. 51, no. 9, pp. 4414–4428, September, 2021, pmid:32598287

25. Sun T., Shao C., Liao H. and Ding S. and Xu X. "Research on adaptive local feature enhancement in convolutional neural networks," IET Image Processing, vol. 14, no. 8, pp. 4306–4315, February, 2021,