Nonlinear blind equalization method of VLC system based on variational auto-encoder

Abstract

The invention belongs to the technical field of visible light communication, and relates to a nonlinear blind equalization method of a VLC (Visible Light Communication) system based on a variational auto-encoder, which comprises the following steps: inputting a received signal into a trained blind equalization network model to obtain an equalized signal; the training process of the blind equalization network model comprises the following steps: pre-processing a sending signal and a receiving signal of the VLC to obtain a pre-processed sending signal Tx and a pre-processed receiving signal Rx; inputting the Rx into an equalizer network to obtain the prior probability distribution of the potential variable signal Tx * and the posterior probability distribution of the given received signal Rx; inputting the prior probability distribution and the posterior probability distribution into a channel estimation network to obtain likelihood probability distribution and an equalization signal; calculating a loss function value according to the prior probability distribution, the posterior probability distribution and the likelihood probability distribution, and updating parameters of the model according to the loss function value until a stop condition is reached; according to the method, multiple loss functions are fused, and the problem of a local optimal solution caused by single loss is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of visible light communication, and relates to a non-linear blind equalization method for a VLC system based on a variational autoencoder. Background Art

[0002] Visible Light Communication (VLC) uses the visible light spectrum to transmit information and is an emerging optical wireless communication technology. This technology mainly modulates optical signals through light-emitting diodes (LEDs) for data transmission. Due to its advantages such as high cost-effectiveness, strong security, no electromagnetic pollution, and no need for spectrum authorization, VLC technology has become a hot topic in the scientific research field and a technology with great development potential. The research on VLC has been continuously deepened, mainly focusing on aspects such as optical materials, advanced modulation techniques, signal equalization processing, and multiplexing techniques, especially in the field of equalization technology. However, when signals are transmitted in a VLC communication system, they will be affected by severe linear and non-linear effects. The non-linear effects in visible light will dominate and seriously affect the transmission performance of the system.

[0003] The rapid development of machine learning has brought about the diverse application development of equalization technology. Traditional neural networks have shown high equalization performance in dealing with non-linear impairments, but most of them are based on the implementation of non-blind method equalizers. Currently, a Blind Equalizer (BE) can perform channel equalization without a training sequence. This method provides an effective practical way to reduce signal distortion without consuming additional bandwidth. This means that during signal transmission, the blind equalization technology can avoid using training sequences, thus not occupying valuable communication bandwidth, effectively improving the communication capacity of optical fiber communication systems, and saving communication resources.

[0004] In recent years, some methods have been proposed, such as blind equalization algorithms based on the Constant Modulus Algorithm (CMA) and Least Mean Square (LMS). Most of these methods estimate the channel by statistical signal characteristics to achieve blind equalization, but such methods have high computational complexity, poor equalization effect, and poor robustness. Summary of the Invention

[0005] To solve the above-mentioned problems of the prior art, the present invention adopts a non-linear blind equalization method for a VLC system based on a variational autoencoder, including: obtaining the received signal at the receiving end of the VLC system, inputting the received signal into a trained blind equalization network model to obtain an equalized signal; the blind equalization network model includes: an equalizer network and a channel estimation network; the training process of the blind equalization network model includes:

[0006] S1. Obtain the transmission signal of the transmitter of the VLC system and its corresponding received signal of the receiver; preprocess the transmission signal and its corresponding received signal of the receiver to obtain training samples; the training samples include: the preprocessed transmission signal Tx and its corresponding received signal Rx of the receiver.

[0007] S2. Input the received signal Rx into the equalizer network to obtain the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution given the received signal Rx.

[0008] S3. Input the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution given the received signal Rx into the channel estimation network to obtain the likelihood probability distribution of the received signal Rx given the latent variable signal Tx * and the equalized signal Rx * ;

[0009] S4. Calculate the loss function value according to the likelihood probability of the received signal Rx given the latent variable signal Tx * , the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution given the received signal Rx, update the parameters of the blind equalization network model according to the loss function value, and complete the training of the blind equalization network model when the pre-set training stop condition is reached.

[0010] Beneficial effects:

[0011] The mean square loss function of the present invention ensures that the prior probability distribution of the latent variable signal is close to the preset ideal distribution, measures the difference between the prior distribution and the channel data distribution through KL divergence, and prevents overly discrete or unreasonable signal values; the singleton loss function measures the error between the posterior probability distribution predicted by the equalizer network and the real data, ensures that the equalizer network can extract effective features from the received signal to compensate for signal distortion, and can dynamically adjust the output signal to match the expected latent variable distribution during the training process, which enables better performance even under unseen channel conditions; the reconstruction loss function measures the accuracy of reconstructing the received signal from the latent variable, evaluates the modeling ability of the channel estimation network for the channel response model, and different from the singleton loss function, verifies the rationality of the equalized signal from another perspective; by fusing the above three loss functions, the problem of local optimal solutions that may be caused by a single loss is avoided, the equalization effect is improved, and at the same time, the "adversarial" mechanism between the equalizer network and the channel estimator network improves the robustness and generalization ability of the model, that is, the equalizer network attempts to recover the original signal, while the channel estimator verifies the rationality of the recovered signal. Description of the Drawings

[0012] Figure 1The architecture diagram of the PAM-VLC system provided by the embodiments of the present invention;

[0013] Figure 2 The flowchart of a non-linear blind equalization method for a VLC system based on a variational autoencoder provided by the embodiments of the present invention. Detailed implementation manners

[0014] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0015] As Figure 1 described, the PAM-VLC system architecture includes a transmitter and a receiver; the transmitter includes: an arbitrary waveform generator, two electronic amplifiers, a bias unit, an LED, and an optical filtering component; the receiver includes: a PIN photodetector, a digital storage oscilloscope, and an offline signal processing module. The working process is as follows: the original binary data to be transmitted, that is, the transmission signal, is mapped (for example, using PAM-8 mapping) into PAM symbols, and after upsampling, upconversion, and pulse shaping, it is sent to an arbitrary waveform generator (AWG); the first electronic amplifier further enhances the PAM signal generated by the AWG, and the bias unit applies a DC bias voltage to the signal enhanced by the first electronic amplifier to ensure that the signal output by the bias unit can activate the LED and make it reach the operating voltage threshold; the LED then converts the signal output by the bias unit into an optical signal, and the optical signal is transmitted to the receiver after necessary optical processing through an optical filtering component (i.e., a grating and filter system); the processed optical signal is received by the PIN photodetector at the end of the transmission process, the second electronic amplifier amplifies the signal received by the PIN photodetector, the digital storage oscilloscope samples and records the amplified signal, and the offline signal processing module analyzes and processes the signal recorded by the digital storage oscilloscope to obtain an equalized signal, and the equalized signal is converted into the original binary bits through PAM mapping to complete the signal compensation.

[0016] Based on the above PAM-VLC system, as Figure 2 shown, the present invention adopts a non-linear blind equalization method for a VLC system based on a variational autoencoder, including the following steps: obtaining the received signal at the receiver end of the VLC system, inputting the received signal into a trained blind equalization network model to obtain an equalized signal; the blind equalization network model includes: an equalizer network and a channel estimation network; the training process of the blind equalization network model includes:

[0017] S1. Obtain the transmission signal of the transmitter of the VLC system and its corresponding received signal of the receiver; preprocess the transmission signal and its corresponding received signal of the receiver to obtain training samples; the training samples include: the preprocessed transmission signal Tx and its corresponding received signal Rx of the receiver.

[0018] Preprocessing the transmission signal and its corresponding received signal of the receiver includes: setting the sliding window size tap = n, performing data segmentation on the transmission signal and its corresponding received signal of the receiver according to the set sliding window size n, and dividing the segmented transmission signal Tx and its corresponding received signal Rx of the receiver into training samples and validation samples according to a ratio of 7:3.

[0019] S2. Input the received signal Rx of the training samples into the equalizer network to obtain the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution given the received signal Rx.

[0020] The equalizer network is a deep learning network model. In this embodiment, the entity extraction neural network (EXNN) model is selected; the EXNN model includes an input layer, a hidden layer, and an output layer; among them, the input layer is used to receive the received signal Rx, the hidden layer is used to extract the features of the received signal Rx, each hidden layer includes a cascaded convolutional layer and an activation function layer, the convolutional layer extracts the linear and non-linear impairments of the received signal Rx by sampling convolutional kernels, the activation function layer uses the Relu function as the activation function to capture and model non-linear features, and the output layer includes two softmax layers, and the two softmax layers respectively output the prior probability distribution P * (Tx θ ) and the posterior probability distribution given the received signal Rx * .

[0021] The prior probability distribution describes the statistical description of the possible states of the latent variable signal in the case where the probability distribution of the received signal of the receiver is not given. This prior probability should be as close as possible to the set prior probability, so calculate the mean square loss function of the prior probability distribution and the probability distribution of the received signal Rx where D KL () represents calculating the Kullback-Leibler divergence, λ is a regularization parameter, θ is the parameter of the equalizer network, is the parameter of the channel estimation network, is the probability distribution of the received signal Rx. In this embodiment, in the VLC system of PAM8, the distribution of Tx is observed to be closer to a uniform distribution. Therefore, the probability distribution of the received signal Rx is set to a uniform distribution. Through the mean square loss function L prior can effectively adapt to and predict the latent variable signal data, especially in the case of complex or incompletely observable signal conditions, and then constrain the generation of the equalized signal.

[0022] The posterior probability distribution reflects that the equalizer network adjusts the output signal through the specific information of the received signal Rx at the receiving end to compensate for signal distortion. The posterior probability distribution predicted by the equalizer network should be as close as possible to the predicted prior probability distribution of the latent variable signal Tx * of, then use the Kullback-Leibler divergence to calculate the posterior probability distribution singleton loss function of the latent variable signal Tx * when the received signal Rx is given Or use the cross-entropy calculation method to calculate the error for the data of multi-class label classification.

[0023] S3. Input the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution when the received signal Rx is given into the channel estimation network, and obtain the likelihood probability distribution of the received signal Rx * when the latent variable signal Tx is given * ;

[0024] The channel estimation network is the same EXNN as the equalizer network, including an input layer, a hidden layer, and an output layer; among them, the input layer is used to receive the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution when the received signal Rx is given, and the hidden layer is used to extract the features of the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution when the received signal Rx is given. The output layer includes two softmax layers, and the two softmax layers respectively output the likelihood probability distribution P * of the received signal Rx when the latent variable signal Tx is given θ (Rx|Tx * ) and the equalized signal Rx * ;

[0025] The main task of the channel estimation network is to recover the received signal Rx from the prior probability distribution and posterior probability distribution output by the equalization network * , and at the same time measure from another angle the various distortions brought by the channel response when the signal passes through the communication channel, such as noise, attenuation, and interference, etc.

[0026] Specifically, the calculation of the reconstruction loss function can be expressed as The reconstruction loss function is expressed in the posterior probability distribution Next, calculate the expectation E of the logarithm of the likelihood probability, such that the latent variable signal Tx * can recover the original signal Rx as accurately as possible * .

[0027] S4. According to the given latent variable signal Tx * When the likelihood probability of the received signal Rx, the prior probability distribution of the latent variable signal Tx * and the posterior probability distribution when the received signal Rx is given, calculate the loss function value, and update the parameters of the blind equalization network model according to the loss function value. When the preset training stop condition is reached, the training of the blind equalization network model is completed.

[0028] The loss function Loss is as follows:

[0029] Loss = αL prior + βL sig + γL Rec + δL mse

[0030] Among them, α, β, γ, and δ are weights used to balance the influence of different losses on the total training process.

[0031] L mse is the loss between the calculated equalized signal and the actual signal data sent by the corresponding transmitting end of the signal. Usually, MSE can be used as the loss metric:

[0032]

[0033] Among them, N represents the length of the signal. Usually, this loss function is used to quantify the error between the model and the target signal. The latent variable signal Tx * can be obtained by randomly sampling its prior probability distribution and the posterior probability distribution when the received signal Rx is given.

[0034] The process of determining whether the preset training stop condition is reached includes: determining whether the preset maximum number of training times M has been reached. If so, the preset training stop condition is reached; otherwise, the validation samples are input into the blind equalization network model for training. If the accuracy of training the blind equalization network model has not improved for m consecutive times, the preset training stop condition is reached; among them, M is generally set to 600, and m is generally set to 20.

[0035] Optimizers for backpropagation updates usually choose the Adam optimizer. In this embodiment, the Adam optimizer performs well in dealing with coefficient gradients and non-stationary targets. At the same time, for the parameters θ of the equalizer network and the channel estimation network, they are updated separately. During the backpropagation and parameter update process, the loss is calculated through the chain rule to update θ and the two network model parameters respectively. After each iteration, it gradually approaches the optimal solution. The model weights are updated through the optimizer for the next model training.

[0036] The above-mentioned embodiments have further elaborated on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above-mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A nonlinear blind equalization method for a VLC system based on a variational autoencoder, characterized in that: include: Get the received signal at the receiving end of the VLC system, input the received signal into the trained blind equalization network model, and obtain the equalized signal Rx * ; The blind equalization network model includes: an equalizer network and a channel estimation network; the training process of the blind equalization network model includes: S1, obtaining a transmission signal of a transmitting end of a VLC system and a receiving signal of a corresponding receiving end; preprocessing the transmission signal and the receiving signal of the corresponding receiving end to obtain a training sample; the training sample includes: a preprocessed transmission signal Tx and a receiving signal Rx of the corresponding receiving end; S2. Input the received signal Rx into the equalizer network to obtain the latent variable signal Tx * The prior probability distribution of and the posterior probability distribution when a given received signal Rx is given; S3, the potential variable signal Tx * The prior probability distribution of the received signal Rx and the posterior probability distribution are input into the channel estimation network to obtain the given latent variable signal Tx * The likelihood probability distribution of the received signal Rx and the equalized signal Rx * ; S4, according to the given potential variable signal Tx * The likelihood probability of the received signal Rx and the latent variable signal Tx * The loss function value is calculated based on the prior probability distribution of the received signal Rx and the posterior probability distribution when the received signal Rx is given, and the parameters of the blind equalization network model are updated according to the loss function value. When the preset training stop condition is reached, the training of the blind equalization network model is completed.

2. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 1, characterized in that: The equalizer network is EXNN; EXNN includes an input layer, a hidden layer and an output layer; wherein the input layer is used to receive the received signal Rx, the hidden layer is used to extract the features of the received signal Rx, and the output layer includes two softmax layers, which respectively output the latent variable signal Tx according to the features extracted by the hidden layer * The prior probability distribution of and the posterior probability distribution of a given received signal Rx; where EXNN is an entity extraction neural network.

3. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 1, characterized in that: The channel estimation network is the same as the equalizer network EXNN; EXNN includes an input layer, a hidden layer and an output layer; wherein the input layer is used to receive the latent variable signal Tx * The prior probability distribution and the posterior probability distribution of the given received signal Rx, the hidden layer is used to extract the latent variable signal Tx * The prior probability distribution and the characteristics of the posterior probability distribution when the received signal Rx is given, the output layer includes two softmax layers, and the two softmax layers output the given latent variable signal Tx according to the features extracted by the hidden layer * The likelihood probability distribution of the received signal Rx and the equalized signal Rx * ; Among them, EXNN is the entity extraction neural network.

4. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 1, characterized in that: The loss function Loss is: Loss=αL prior +βL sig +γL Rec +δL mse Among them, α, β, γ, δ are weights, L prior is the latent variable signal Tx * The mean square loss function of the prior probability distribution and the probability distribution of the received signal Rx, L sig is the latent variable signal Tx * The singleton loss function of the posterior probability distribution given the received signal Rx, L Rec is the reconstruction loss function of the received signal, L mse is the error loss function between the latent variable signal and the signal actually sent by the sender.

5. The nonlinear blind equalization method of a VLC system based on a variational autoencoder according to claim 4, characterized in that: Loss function L prior The calculation formula is: Among them, D KL () indicates the calculation of Kullback-Leibler divergence, λ is the adjustment parameter, P θ (Tx * ) is the latent variable signal Tx * The prior probability distribution of is the probability distribution of the received signal Rx.

6. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 4, characterized in that: Loss function L sig The calculation formula is: Among them, D KL () indicates the calculation of Kullback-Leibler divergence, is the posterior probability distribution of the given received signal Rx, P θ (Tx * ) is the latent variable signal Tx * The prior probability distribution of .

7. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 4, characterized in that: The calculation formula of the reconstruction loss function is: in, is the posterior probability distribution of the given received signal Rx, P θ (Rx|Tx * ) is a given latent variable signal Tx * The likelihood probability distribution of the received signal Rx when .

8. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 4, characterized in that: Loss Function Among them, the latent variable signal Tx * It is obtained by randomly sampling its prior probability distribution and the posterior probability distribution when a given received signal Rx is given, where N represents the length of the signal and i represents the index of the data point of the signal.

9. The nonlinear blind equalization method for a VLC system based on a variational autoencoder according to claim 1, characterized in that: Preprocessing the transmission signal and the received signal at the corresponding receiving end includes: setting the sliding window size, performing data segmentation on the transmission signal and the received signal at the corresponding receiving end according to the set sliding window size, and dividing the segmented transmission signal Tx and the received signal Rx at the corresponding receiving end into training samples and verification samples.

10. The nonlinear blind equalization method of a VLC system based on a variational autoencoder according to claim 9, characterized in that: The process of determining whether a preset training stop condition is reached includes: determining whether a preset maximum number of training times has been reached, if so, the preset training stop condition is reached; otherwise, inputting the verification sample into the blind equalization network model for training, if the accuracy of the blind equalization network model has not been improved after multiple consecutive trainings, the preset training stop condition is reached.

Images