Method for constructing fronthaul system architecture by using multiband auto-encoder

Abstract

The invention belongs to the technical field of optical wireless communication, and particularly relates to a method for constructing a fronthaul system architecture by using a multiband auto-encoder. The method comprises the following steps: designing a self-coding framework of multiband joint coding, and realizing a forward transmission network architecture supporting multi-user access; wherein a data-driven neural network channel model is used for learning deterministic damage and random characteristics existing in a physical channel; and on the basis of a data-driven channel model, a multi-band full-neural network transceiver is constructed, and damage from a plurality of transceiving systems and propagation links is jointly optimized by using a gradient descent algorithm. According to the invention, various damages in a multi-band system and a channel link in the forward transmission network can be optimized in an end-to-end joint manner, the global optimal communication performance is realized, and large-scale efficient deployment is supported. Experimental results show that 1.9 dB receiving optical power sensitivity gain can be realized. The method also has excellent robustness and stability, and can realize the average Q value gain of 1.41 dB in a turbulent channel.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical wireless communications, and in particular relates to a method for constructing a fronthaul system architecture using a multi-band autoencoder. Background Art

[0002] Optical wireless communications (OWC) have become an increasingly promising solution for the next generation of communications (6G) [1,2]. To support a wide range of future Internet applications, including mobile Internet, industrial Internet, augmented reality (AR), smart machines, and other large-scale device interconnection, researchers must move towards fronthaul solutions with high-capacity and high-bandwidth transmission [3-5]. Optical wireless communications have become an effective solution for fronthaul transmission due to their cost-effectiveness, easy deployment, high bandwidth, and enhanced security benefits [5,6].

[0003] In recent years, fronthaul solutions based on optical wireless transmission have gained an important position in the field of communication network architecture [6, 7]. However, the performance of existing optical wireless transmission is often limited by the combined impairments of the transceiver system and the channel link, including inter-symbol interference (ISI), signal-to-signal beat interference (SSBI), nonlinearity in optical communication systems, and atmospheric turbulence. Traditional modular digital signal processing (DSP) methods can specifically mitigate one or several types of impairments, but cannot jointly optimize all impairments, causing the system performance to fall into a local optimum.

[0004] The existing fronthaul solutions based on optical wireless communication are limited by the design of traditional modular DSP, which means that various types of system damage can only be optimized locally and it is impossible to achieve global optimization of communication performance. Summary of the invention

[0005] The purpose of the present invention is to provide a method for constructing a fronthaul system architecture using a multi-band autoencoder, so as to realize an end-to-end optimized fronthaul system supporting multiple access, so as to overcome the problem that the traditional modular design is difficult to jointly optimize various types of impairments in the data transmission link. The present invention realizes orthogonal multiplexing through frequency domain division, and in actual deployment, the transceiver only needs a neural network for joint encoding and decoding without an additional DSP, showing the potential to support large-scale and efficient deployment.

[0006] The method for constructing a fronthaul system architecture using a multi-band autoencoder provided by the present invention includes designing a new multi-band autoencoder (abbreviated as (MBAE)) to achieve end-to-end global optimization of the optical wireless fronthaul system; including using a data-driven neural network channel model to effectively learn the deterministic damage and random characteristics of the physical channel; and based on the data-driven channel model, constructing a multi-band full neural network transceiver, and using a gradient descent algorithm to jointly optimize the damage from multiple transceiver systems and propagation links. The specific steps are:

[0007] (1) Construct a data-driven neural network channel model to learn the deterministic impairments and random characteristics of the physical channel;

[0008] The data-driven neural network channel model mainly consists of two parts: an attention-based deterministic noise estimator (ADNE) and a generative adversarial network-based noise injector (GNI), as shown in Figure 1 The main framework of ADNE is a dual-branch channel simulator based on artificial neural network (ANN), which consists of a linear branch and a nonlinear branch, which are used to learn linear impairments and nonlinear impairments in optical wireless channels respectively. The linear branch consists of two hidden layers without any nonlinear activation function, which can be expressed as:

[0009]

[0010] In the formula, and Represent the weights and biases of the two hidden layers respectively. X represents the input of the channel model, and BN represents batch normalization, which is used to speed up the convergence of the model. In contrast, the nonlinear branch has only one hidden layer and uses Tanh as the nonlinear activation function. The transfer function of the nonlinear branch can be written as:

[0011]

[0012] in, and represents the weights and bias of the hidden layer of this branch. In addition, in order to effectively manage the weights of the linear and nonlinear branches, an attention mechanism is introduced to handle the changing nonlinear conditions in real systems. This mechanism allows these branches to adapt to the linearity and nonlinearity of the transmitted signal even under changing system conditions, such as different signal amplitudes. The weights of the attention mechanism can be expressed as:

[0013]

[0014] in, and Represents the weights and biases of the attention mechanism layer. linear and nonlinear Represent the weights of the linear branch and the nonlinear branch respectively. Therefore, the transfer function of the ADNE model can be expressed as follows:

[0015]

[0016] The loss function used to train the ADNE model is the mean squared error loss function (MSE).

[0017] Considering that ADNE tends to learn the deterministic response of the channel during training, once the training of ADNE is completed, it will become a deterministic mapper without any random features, which is inconsistent with the real channel. Therefore, we further propose the GNI model to add random time-varying features to the channel model to match the real channel. It is worth emphasizing that the random noise existing in the real system is time-independent, which makes it impossible to learn random noise itself. Therefore, the goal of GNI is to learn as many statistical features as possible from the noise samples and add random noise to the channel model according to the learned statistical distribution. When the training of ADNE is completed, the transmitted signal is first input into ADNE, and then the output signal of ADNE is subtracted from the real received signal to obtain the additive random noise sample for training GNI. Based on the generative adversarial network (GAN) that is good at learning the underlying data distribution, GNI is designed to learn the statistical distribution of random noise samples. Generative adversarial learning enables GNI to simulate random features that are very close to the real channel, and its training optimization objective can be expressed as:

[0018]

[0019] Where r is a random vector, which represents the input of the GNI generator. l is the label of GNI training, and p ε () represents the statistical distribution of random noise samples, p r () represents the distribution of the input random vector. In this work, we set r~N(0,1). Ω() and Θ() represent the outputs of the generator and discriminator of GNI, respectively. After completing the training of ADNE and GNI in sequence, the generators of ADNE and GNI are jointly deployed to simulate the real channel. The output of the entire channel model is finally expressed as:

[0020]

[0021] (2) Build a multi-band full neural network transceiver and iteratively update it using a gradient descent algorithm;

[0022] Based on the data-driven channel model, the present invention further constructs a full neural network receiver and transmitter to implement a multi-band autoencoder framework. Figure 1As shown, the forward transmission framework MBAE of the present invention is composed of three components: a transmitting neural network (TNN), a channel model and a receiving neural network (RNN). TNN is composed of multiple encoders and multiple convolutional neural networks (CNN) to generate multi-band signals. Correspondingly, RNN is composed of multiple decoders and CNNs to achieve decoding of corresponding frequency band signals. The neural networks at both ends of the transmitter and receiver are jointly optimized based on gradient descent to resist the joint damage existing in real optical wireless communication channels, including inter-code interference, nonlinearity and random noise.

[0023] For the i-th encoder, its output expression is as follows:

[0024] e i =ψ(v i ), (7)

[0025] In the formula, ψ() represents the forward propagation process of the neural network, v i is the input vector, e i is the corresponding output, and its dimension is recorded as E. In addition, in order to ensure the frequency domain orthogonality of the signals generated by different encoders, we use different CNNs to selectively filter the feature signals in the latent space generated by different encoders, thereby realizing frequency division multiplexing of multiple signals. For the i-th CNN, its convolution kernel expression can be written as:

[0026]

[0027] In the formula, represents the center frequency of the ith frequency band, q(t) represents the root cosine filter, and its roll-off factor is β. After each transmission signal passes through the corresponding Encoder and CNN in turn, we add each signal together to obtain the final multi-band transmission signal, which can be expressed as:

[0028]

[0029] Corresponding to the transmitting neural network, the receiving neural network consists of multiple CNNs and decoders to filter and decode the signals of each sub-band respectively. At the receiving end, the received signal is first extracted by CNN. For the i-th signal, the signal after CNN can be expressed as:

[0030]

[0031] Among them, a(t) represents the received signal, o i (t) represents the signal of the i-th frequency band.

[0032] After CNN filtering, the signal is sent to the decoder that integrates equalization and decoding functions. In order to reduce the inter-symbol interference suffered by the signal, each decoder input includes a waveform of adjacent symbols, and then the output layer uses a softmax activation function to calculate the probability corresponding to each predicted symbol. In the entire MBAE network training, cross entropy is used as the loss function of the neural network, which can be expressed as:

[0033]

[0034] Where N is the total number of frequency bands, M is the batch size, k is the middle moment of the encoder input time slice data, and p i,j represents the output probability of the jth batch of the decoder corresponding to the i-th frequency band.

[0035] The technical features and performance advantages of the present invention are:

[0036] The multi-band autoencoder framework (MBAE method) proposed in the present invention can end-to-end jointly optimize various types of damages including inter-code interference, beat frequency interference, nonlinearity, atmospheric turbulence, etc. caused by multi-band systems and channel links in the fronthaul network, achieve global optimal communication performance and support large-scale efficient deployment, and solve the problem that traditional modular design can only achieve local performance optimization and low deployment efficiency. Experimental results show that compared with traditional multi-band carrier-free phase amplitude modulation, the present invention can achieve a 1.9dB received optical power sensitivity gain and a 214mV dynamic adjustment range gain without relying on additional DSP. In addition, the present invention shows excellent robustness and stability, and can achieve an average Q value gain of 1.41dB in turbulent channels, and supports a maximum link budget of 44.7dB. This shows that compared with traditional digital signal processing methods, MBAE can achieve substantial performance enhancement and highlight its potential to provide higher link budgets and more robust performance in actual deployment scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the architecture and signal flow of MBAE of the present invention.

[0038] Figure 2 The performance comparison of the MBAE method of the present invention with other traditional methods under simulated channels. Among them, a is the Q factor performance of the three frequency bands of each method under different channel factors factor, b is the average Q factor performance of the three frequency bands of each method under different channel factors factor, c is the spectrum of the transmitted signal and the received signal when the channel factor is 55, and d is the spectrum of the transmitted signal and the received signal when the channel factor is 15.

[0039] Figure 3The data processing flow and experimental device diagram for the experiment of the present invention. Among them, a is the digital signal processing flow at the transmitting end, b is the end-to-end optimization process of the MBAE of the present invention, c is the digital signal processing flow at the receiving end, and d is a schematic diagram of the experimental device.

[0040] Figure 4 The schematic diagram of the effect of the channel model of the MBAE method of the present invention on modeling a real optical wireless channel. Among them, a is a schematic diagram of the frequency response of the real channel and the learning curve of the deterministic noise estimator, b is the estimation result of the channel frequency response by the channel model, and c is the simulation result of the statistical characteristics of random noise.

[0041] Figure 5 The performance comparison diagram of the MBAE method of the present invention and other traditional methods under different voltage peak-to-peak values ​​is shown in FIG. a is the bit error rate performance of the three frequency bands of different methods, b is the Q factor performance of the three frequency bands of different methods, and c is the average Q factor performance of the three frequency bands of different methods.

[0042] Figure 6 The performance comparison diagram of the MBAE method of the present invention and other traditional methods under different received optical powers and the supported link budget results are shown in Figure 1. Among them, a is the Q factor performance of various methods under different received optical powers, and b is the average Q factor performance of various methods under different receiving link budgets.

[0043] Figure 7 The continuous test results of the bit error rate of the MBAE method of the present invention and other traditional methods in a turbulent channel and the statistical distribution results of the bit error rate are shown in Figure 1. Among them, a is the bit error rate performance of different methods under the condition of random turbulence, and b is the corresponding statistical distribution diagram of the bit error rate performance. DETAILED DESCRIPTION

[0044] The present invention is further described below in conjunction with specific simulation and experimental methods and results.

[0045] (1) Simulation settings and experimental results

[0046] In order to preliminarily verify the performance of the proposed multi-band autoencoder framework, we constructed a theoretical channel model as a simulation platform for performance evaluation. The simulation preliminarily verifies the MBAE model through the theoretical channel, which only considers nonlinear distortion, frequency fading, and additive noise. It can be mathematically expressed as:

[0047]

[0048] Where x and y represent the transmitted signal and the received signal. g() represents nonlinear distortion, and f represents frequency. represents negative exponential frequency fading, and n is the additive Gaussian white noise of the channel. In this study, fs The value of is 0.1302. The signal-to-noise ratio of factor and Gaussian white noise are set according to specific requirements. The expression of the nonlinear function can be expressed as:

[0049]

[0050] Among them, the values ​​of b1, b2, b3 and b4 are set to 1.259, 0.01373, -0.4576 and 0.8165 respectively.

[0051] The present invention explores the influence of different channel frequency responses on MBAE performance in a simulated channel. The use of MBAE is divided into two steps, training and deployment. The training phase is further divided into the training of the channel model and the training of the transceiver. Therefore, before the formal measurement, it is first necessary to use the simulated channel to collect pilot data to train the channel model, and then train the transceivers of multiple MBAEs, and the trained channel model will be frozen at this time. After the training is completed, MBAE can be used for deployment to generate a transmission signal and for decoding the received signal at the receiving end. Figure 2 (a) shows a schematic diagram of the Q factor performance of the MBAE method of the present invention and other traditional digital signal processing methods under different channel factors. At this time, the signal-to-noise ratio of the channel is fixed at 10dB. For carrier-free phase amplitude (CAP) modulation, the performance difference between its three frequency bands is always significant under the influence of spectrum fading. In contrast, for MBAE, when the factor exceeds 35, the performance of its three frequency bands remains highly consistent, and the difference is negligible. Only when it is lower than 35, the performance of MBAE begins to show a significant gap. And under different channel responses, MBAE always obtains better average performance than CAP, such as Figure 2 As shown in (b), this highlights the adaptability of MBAE to different channel conditions. Figure 2 (c) and (d) show the spectrum of the transmitted and received signals when the factor is 55 and 15, respectively. As we can see, when the channel spectrum is not severe, MBAE can effectively balance the performance of the signals in its three frequency bands. However, when the signal bandwidth is significantly limited, MBAE will sense the fading of the channel through the channel model and appropriately adjust the performance of each frequency band to obtain the global optimal average performance, so that it achieves an average Q factor performance that is significantly better than the traditional algorithm under various channel conditions, such as Figure 2 (b) as shown.

[0052] (2) Experimental platform and experimental results

[0053] After conducting basic simulations, we further verified the performance in a real optical wireless communication system. Figure 3The data processing flow and experimental setup diagram of this experiment are shown in Figure 2. The digital signal processing flow at the transmitter is as follows: Figure 3 (a). The symbols to be transmitted are initially converted to one-hot encoding and then subjected to a time sliding window operation to match the encoder input size. Subsequently, the symbols of the three bands are input into different sub-transmit neural networks (STNs) for multi-band joint encoding. The outputs of the three STNs are summed using an adder to produce the final transmitted signal. To ensure compatibility with the sampling rate of the arbitrary waveform generator (AWG), the signal is additionally resampled before being input into the AWG. Figure 3 (d) shows the devices and channels included in the system used. At the transmitting end, the transmitted data is first uploaded to the AWG (M8195A) to complete the digital-to-analog conversion. Then, the generated analog signal passes through an electronic amplifier (EA, H301-1110, 10GHz) and two attenuators (ATTs, YM11520, 18GHz) to amplify the signal and adjust the power of the transmitted signal. Subsequently, a Mach-Zehnder modulator (MZM, T.MXH1.5-20PD-ADC-LV, 27GHz) is used to modulate the signal to a laser with a wavelength of 1550nm. Finally, a collimating lens is used to focus the emitted light. In addition, an erbium-doped fiber amplifier (EDFA, EDFA-DC-30-S-FA) and an optical attenuator (OA, M-13 / 15-40-FA) are used. The channel consists of 3 meters of free space, and the air temperature is controlled at 25 degrees Celsius. The turbulence phase plate (TPP, LS-100 rotating stage) was located 2.8 m away from the receiving end to simulate atmospheric turbulence.

[0054] At the receiving end, the incident light is first focused by another collimating lens, and then the power of the received light is controlled and measured by an OA (OAT-M-13 / 15-40-FA). Subsequently, another EDFA (EDFA-DC-10-S-FA) is used to amplify the power of the received light, and then a photodetector (PD, Agilent 11982A, 11GHz electrical bandwidth and 15GHz optical bandwidth) is used to convert the signal from the optical domain to the electrical domain. The resulting electrical signal is sampled by an oscilloscope (OSC, DSAX92004A). After completing the analog-to-digital conversion, the received data is stored for subsequent offline digital signal processing and decoding performance analysis.

[0055] MBAE’s end-to-end optimization process is as follows Figure 3As shown in (b), it includes the training of the channel model and the transceiver model. In the channel model training process, we first collect the transmitted data, receive the data through the real channel, and build a data set. The training of the channel model is divided into the training of ADNE and the training of GNI. Based on the trained channel model, each STN and SRN are trained simultaneously through end-to-end optimization. At this time, the channel model is frozen and embedded in the gradient transfer between the transceiver.

[0056] Before evaluating the communication performance of the proposed multi-band autoencoder framework, the performance of the channel models based on ADNE and GNI are first investigated. Figure 4 (a) shows the response of the real system, with a 3dB bandwidth of about 6.59GHz. The frequency response of the entire system is mainly limited by the bandwidth of the low-cost EA and PD, and drops rapidly around 10G. Figure 4 (b) is the estimation result of the channel response by ADNE, which shows that ADNE can effectively learn the deterministic response of the system, but it will inevitably not match the response of the real system due to the influence of random noise in the system. The random noise in the system may come from the shot noise of the light source, propagation link and device under spontaneous radiation and stimulated radiation. In order to make the channel model have random characteristics similar to the real system, GNI is used to learn the statistical characteristics of random noise samples and used as a random noise source in the end-to-end optimization stage. Figure 4 (c) shows the simulation results of the statistical distribution of random noise. In the initial stage of training, there is a difference between the distribution of GNI output and the label. However, after completing GNI training, their distributions are perfectly aligned and show very similar variances. This ensures that our channel model can have largely similar random characteristics to the real channel, which can be fed back into the training of TNN and RNN to achieve end-to-end optimization.

[0057] Figure 5(a) shows the bit error rate performance of the MBAE method of the present invention and the traditional algorithm as the voltage peak-to-peak value (Vpp) changes when the received optical power (ROP) is set to -24dBm. The bit error performance of all methods shows a V-shaped trend. This is because in the low Vpp region, the signal performance is limited by the low signal-to-noise ratio, while in the high Vpp region, the signal performance is limited by the nonlinearity of the system. For CAP modulation, the complex channel response leads to significant performance differences between its three bands. Although the neural network-based post-equalization can improve the performance in the nonlinear region, it does not fundamentally solve the problem of extremely low performance in the third band, which results in the average performance of CAP failing to exceed the 7% hard decision forward error correction code (HD-FEC) threshold. In contrast, MBAE maintains a more balanced performance on the three bands, and its Q factor and average Q factor performance are significantly better than the baseline, achieving a dynamic adjustment range gain of 214mV at the 7% HD-FEC threshold. It is worth noting that in the nonlinear region, the performance of the three bands of MBAE begins to diverge. This divergence is mainly due to the degradation of the peak energy of the three bands in the nonlinear region. However, since the third band occupies most of the energy, it shows more robustness to energy loss, thus achieving better performance.

[0058] Figure 6 (a) and Figure 6 (b) shows the Q factor performance of different schemes with varying received optical power (ROP) and varying link budget, respectively. As ROP decreases and link budget increases, the performance of the three sub-bands and the average performance between different schemes will decrease. However, MBAE always maintains the highest average performance, achieving a 1.9dB receiver sensitivity gain at a 7% HD-FEC threshold and a transmission rate of 36Gbps, and can ultimately support a link budget of 44.7dB. This shows that MBAE achieves substantial performance enhancement compared to the baseline and highlights its potential to provide higher link budgets in real deployment scenarios.

[0059] Figure 7 The bit error rate performance and its statistical distribution of the three frequency bands using different methods under random turbulence implementation are shown. Figure 7 As shown in (a), under different turbulence conditions, the bit error rate performance of each frequency band of different algorithms will fluctuate accordingly. In order to quantitatively analyze the performance of various algorithms, we conducted a statistical analysis of the bit error rate, such as Figure 7(b), where the upper and lower edges represent the range within 1.5 interquartile ranges (IQRs). The boxes represent data in the range of 25% to 75% of the total data bit error rate, and the red cross symbols represent outliers. For CAP modulation, the performance fluctuation is smallest in the first and third bands, while the second band shows significant variability and the third band always performs the worst. In contrast, MBAE shows greater performance fluctuations in all three bands. This variability may be due to the different degrees of convergence of the neural network-based decoding process, which is greatly affected by changes in signal quality, especially when the hyperparameters of all neural networks are completely fixed. We based on Figure 7 The average bit error rate calculation of the method further statistically calculated the average Q factor of various methods under random turbulence conditions. The average Q factors of the three sub-bands of the final two baseline schemes were 7.39dB and 7.45dB respectively, while the average Q factor of the three sub-bands of MBAE was 8.86dB. Compared with CAP modulation using a combination of neural network-based equalization and decision feedback equalization, MBAE has an average Q factor gain of 1.41dB. This shows that MBAE can achieve significant performance improvements even under dynamic conditions, further emphasizing its potential for deployment in real-world scenarios.

[0060] References:

[0061] 1.N.Chi, Y.Zhou, Y.Wei, and F.Hu, "Visible Light Communication in 6G: Advances, Challenges, and Prospects," IEEE Vehicular Technology Magazine, vol.15, no.4, pp.93-102, 2020.

[0062] 2.X.You et al., "Towards 6G wireless communication networks: vision, enabling technologies, and new paradigm shifts," Science China-InformationSciences, vol.64, no.1, Jan, 2021.

[0063] 3.H.Tataria et al.,“6G wireless systems:Vision,requirements,challenges,insights,and opportunities,”Proceedings of the IEEE,vol.109,no.7,pp.1166-1199,2021.

[0064] 4.W.Ejaz et al.,“A comprehensive survey on resource allocation forCRAN in5G and beyond networks,”Journal of Network and Computer Applications,vol.160,pp.102638,2020.

[0065] 5.A.Fayad,T.Cinkler,J.Rak,and M.Jha,“Design of cost-efficientopticalfronthaul for 5G / 6G networks:An optimization perspective,”Sensors,vol.22,no.23,pp.9394,2022.

[0066] 6.H.-B.Jeon et al.,“Free-space optical communications for 6G wirelessnetworks:Challenges,opportunities,and prototype validation,”IEEECommunications Magazine,vol.61,no.4,pp.116-121,2023.

[0067] 7.L.-Y.Wei et al.,“2.333-Tbit / s bi-directional optical mobilenetworks usingoptical wireless communication(OWC),”Optics Communications,vol.475,pp.126187,2020。

Claims

1. A method for constructing a forward transmission system architecture using a multi-band autoencoder, characterized in that: The method includes designing a multi-band autoencoder, abbreviated as MBAE, to achieve end-to-end global optimization of optical wireless fronthaul systems; constructing a data-driven neural network channel model to learn the deterministic impairments and random characteristics of the physical channel; and constructing a multi-band full neural network transceiver based on the data-driven channel model, using the gradient descent algorithm to jointly optimize the impairments from multiple transceiver systems and propagation links; the specific steps are: (1) Construct a data-driven neural network channel model to learn the deterministic impairments and random characteristics of the physical channel; The data-driven neural network channel model consists of two parts: an attention-based deterministic noise estimator (ADNE) and a generative adversarial network-based noise injector (GNI); wherein: The main framework of the noise estimator (ADNE) is a dual-branch channel simulator based on an artificial neural network (ANN), which consists of a linear branch and a nonlinear branch, respectively used to learn linear impairments and nonlinear impairments in optical wireless channels; wherein the linear branch consists of two hidden layers, expressed as: In the formula, and Represent the weights and biases of the two hidden layers respectively; X represents the input of the channel model, BN represents batch normalization, which is used to speed up the convergence of the model; the nonlinear branch has only one hidden layer, and Tanh is used as the nonlinear activation function; the transfer function of the nonlinear branch is written as: in, and represents the weights and bias of the hidden layer of this branch; in addition, in order to effectively manage the weights of linear and nonlinear branches, an attention mechanism is introduced to deal with the changing nonlinear conditions in the real system; this mechanism allows these branches to adapt to the linearity and nonlinearity of the transmitted signal even under changing system conditions; the weight of the attention mechanism is expressed as: in, and represents the weight and bias of the attention mechanism layer, λ linear and λ nonlinear Represent the weights of the linear branch and the nonlinear branch respectively, so the transfer function of the ADNE model is expressed as follows: The loss function used to train the ADNE model is the mean square error loss function (MSE); The noise injector (GNI) adds random time-varying characteristics to the channel model to match the real channel. The noise injector (GNI) is used to learn as many statistical features as possible from the noise samples and add random noise to the channel model according to the learned statistical distribution. When the training of ADNE is completed, the transmitted signal is first input into ADNE, and then the output signal of ADNE is subtracted from the real received signal to obtain the additive random noise sample for training GNI. Based on the generative adversarial network (GAN) that is good at learning the underlying data distribution, GNI is used to learn the statistical distribution of random noise samples. Generative adversarial learning enables GNI to simulate random features that are very close to the real channel, and its training optimization goal is expressed as: Where r is a random vector, which represents the input of the GNI generator; l is the label of GNI training, and p ε (·) represents the statistical distribution of random noise samples, p r (·) represents the distribution of the input random vector; where r is set to N(0,1); Ω(·) and Θ(·) represent the outputs of the generator and discriminator of GNI, respectively; after completing the training of ADNE and GNI in sequence, the generators of ADNE and GNI are jointly deployed to simulate the real channel; the output of the entire channel model is finally expressed as: (2) Build a multi-band full neural network transceiver and iteratively update it using a gradient descent algorithm; Based on the data-driven channel model, a full neural network receiver and transmitter are constructed to realize a multi-band autoencoder (MBAE); the MBAE consists of the following three components: a transmitting neural network (TNN), a channel model and a receiving neural network (RNN); the TNN consists of multiple encoders and multiple convolutional neural networks (CNNs) to generate multi-band signals; correspondingly, the RNN consists of multiple decoders and multiple convolutional neural networks (CNNs) to realize the decoding of the corresponding frequency band signals; the neural networks at both ends of the transmitter and receiver are jointly optimized based on the gradient descent method to resist the joint impairments existing in real optical wireless communication channels, including inter-symbol crosstalk, nonlinearity and random noise.

2. The method according to claim 1, characterized in that In step (2), for the i-th encoder, its output expression is as follows: yes i =ψ(v i ), (7) In the formula, ψ(·) represents the forward propagation process of the neural network, v i is the input vector, e i is the corresponding output, and its dimension is recorded as E. In addition, in order to ensure the frequency domain orthogonality of the signals generated by different encoders, different CNNs are used to selectively filter the feature signals in the latent space generated by different encoders, thereby realizing the frequency division multiplexing of multiple signals. For the i-th CNN, its convolution kernel expression is written as: In the formula, represents the center frequency of the i-th frequency band, q(t) represents the root cosine filter, and its roll-off factor is β; after each transmission signal passes through the corresponding Encoder and CNN in turn, the signals are added together to obtain the final multi-band transmission signal, which is expressed as: Corresponding to the transmitting neural network, the receiving neural network consists of multiple CNNs and decoders to filter and decode the signals of each sub-band respectively; At the receiving end, the received signal is first extracted through CNN. For the i-th signal, the signal after CNN is expressed as: Among them, a(t) represents the received signal, o i (t) represents the signal of the i-th frequency band; After CNN filtering, the signal is sent to a decoder that integrates equalization and decoding functions. In order to reduce the inter-symbol interference suffered by the signal, each decoder input includes a waveform of adjacent symbols, and then the output layer uses a softmax activation function to calculate the probability corresponding to each predicted symbol. In the entire MBAE network training, cross entropy is used as the loss function of the neural network, which is expressed as: Where N represents the total number of frequency bands, M represents the batch size, k represents the middle moment of the encoder input time slice data, and p i,j Represents the output probability of the jth batch of the decoder corresponding to the i-th frequency band.

Images