Adaptive channel response matching neural network equalization method and system

Abstract

The invention relates to the technical field of optical fiber communication, in particular to an adaptive channel response matching neural network equalization method and system. The method comprises the following steps: acquiring signal data, preprocessing the signal data, and dividing the signal data into a training set and a test set; designing an ACRM-NNE model, and obtaining an ACRM-NNE training model through training of the training set; performance is tested through a test set, and the ACRM-NNE training model is used as an adaptive channel response matching neural network equalizer; and matching a neural network equalizer through adaptive channel response to carry out equalization. According to the adaptive channel response matching neural network equalizer, bandwidth limited signals can be effectively processed, the adaptive channel response matching neural network equalizer is suitable for equalization of pulse amplitude modulation signals with inter-symbol interference, different influence factors can be effectively considered by a multi-head attention layer, and the time sequence processing capacity is greatly improved; the cascaded structure of the encoder and the decoder in the ACRM-NNE model can replace the operation mode of partial response equalization and MLSE decoding, and the digital signal processing flow is simplified.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber communication, and more specifically, to an adaptive channel response matching neural network equalization method and system. Background Art

[0002] In recent years, with the development of new applications such as the Internet of Things and cloud computing, network traffic has grown exponentially. These new applications have brought about the demand for bandwidth and computing resources, which has promoted the development of short-distance (<80 km) optical fiber communication systems in data centers. At the same time, the growing data center business has also prompted operators, enterprises and service providers to provide Ethernet solutions for high-speed network communications. At present, the interconnection rate of data centers is transitioning from 400Gb / s to 800Gb / s. Whether it is 400Gb / s or 800Gb / s, intensity modulation direct detection has been widely used due to its advantages such as low complexity, low cost and energy saving. The currently deployed 400Gbit / s Ethernet (400GbE) adopts a 4-channel 53Gbaud four-level pulse amplitude modulation (PAM-4) transmission solution. For the next generation of "beyond 400GbE" Ethernet (such as 800GbE or 1.6TbE), the PAM-4 modulation format is still a promising solution in the future due to its low complexity and easy upgrade from the existing 400GbE. In order to obtain greater transmission capacity to meet the growing traffic demand, it is necessary to increase the number of wavelength division multiplexing and transmit PAM-4 signals with higher baud rates. However, due to the limited bandwidth of the transceiver devices, increasing the number of wavelength division multiplexing or transmitting at a higher baud rate will cause severe ISI in the PAM-4 signal, damaging the communication quality.

[0003] The prior art discloses a QPSK modulation super-Nyquist transmission method and system based on neural network equalization, which introduces convolutional coding and neural network equalization into the super-Nyquist transmission process. At the transmitter end, the data information is first channel encoded, and then QPSK modulation mapping is performed to divide the original information into I and Q channels. The two channels of data are simultaneously subjected to FTN pulse shaping and sent to the channel for transmission; at the receiver end, the two channels of signals transmitted through the channel are simultaneously matched filtered to restore the waveforms of the two channels of data originally transmitted, and FTN sampling is performed after the waveform is restored. FTN sampling is performed according to the shaping pulse interval to obtain the sampling value; then, the sampled signal is sent to the neural network equalizer for equalization to obtain the sample value after equalization; demapping is performed, and the output obtained is a soft output, and finally soft-decision Viterbi decoding is performed to restore the original data. However, this method also requires the use of an MLSE decoder to restore the ISI introduced by the partial response equalizer, which ultimately leads to a complex digital signal processing process. In addition, the conventional neural network equalization scheme lacks the ability to pay attention to timing information, so the ability to equalize ISI in processing time series tasks needs to be improved. Summary of the invention

[0004] The purpose of the present invention is to disclose an adaptive channel response matching neural network equalization method and system with better signal equalization effect.

[0005] In order to achieve the above object, the present invention provides an adaptive channel response matching neural network equalization method, comprising:

[0006] S1: Acquire signal data, and divide the signal data into training set and test set after preprocessing;

[0007] S2: Design an ACRM-NNE model, train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; the ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: a multi-head attention layer, a feedforward network layer, a residual, and normalization;

[0008] S3: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to step S2. If the bit error rate is greater than or equal to the set value, use the ACRM-NNE training model as an adaptive channel response matching neural network equalizer.

[0009] S4: Equalization via adaptive channel response matching neural network equalizer.

[0010] Furthermore, in step S1, it includes:

[0011] Build an optical fiber communication system, which includes a transmitter and a receiver. In the transmitter, the transmission data is mapped to a PAM-4 electrical signal, which is converted into an optical signal through an arbitrary waveform transmitter and a Mach-Zehnder modulator after upsampling and pulse shaping. In the receiver, the optical signal is converted into an electrical signal by a photodetector, digitized by an analog-to-digital converter, and then resampled, matched filtered, and synchronized to obtain a received signal.

[0012] Data collection: Save the synchronized signal from the receiving end as data to be equalized;

[0013] Data preprocessing: perform data preprocessing on the data to be equalized obtained during data collection, and form an array of length 2L+1 with the signal at the current moment and the signals at L moments before and after the current moment as the processed data. The processed training data is divided into a training set and a test set at a ratio of 70%:30%, with the training set accounting for 70% and the test set accounting for 30%.

[0014] Furthermore, in data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end so that the bandwidth is limited, and severe ISI is introduced into the communication signal; the signal with severe ISI introduced into the receiving end is saved as data to be equalized.

[0015] Furthermore, in step S2, the ACRM-NNE model includes:

[0016] The loss function loss1 is designed in the encoder to make the encoder equivalent to the functions of full response equalization and post-filtering. The loss function loss2 is designed in the decoder to output an equalized signal without ISI. The sum of the two losses is used as the total loss function of the ACRM-NNE model.

[0017] Further, in step S2, the ACRM-NNE model is trained by the training set, and obtaining the ACRM-NNE training model includes: initializing the ACRM-NNE model using the Kaiming initialization method, and assigning values ​​to the parameters required in the training process; calculating the total loss, that is, the mean square error, which is specifically calculated as follows:

[0018]

[0019] in and A are the encoder output and its label, y * and y are the encoder output and its label respectively, and the Adam optimizer is used to update the ACRM-NNE model parameters through the back-propagation algorithm until the loss function converges and the training is stopped to obtain the ACRM-NNE model.

[0020] Further, in step S2, the ACRM-NNE model includes: a linear layer converts the input signal into a value vector, which is input into N1 encoders for processing. A multi-head attention layer is used in the encoder to capture the dependencies between different positions, and then a residual is introduced into the normalization layer to avoid degradation during network training. A nonlinear activation function is introduced into the feedforward network, so that the ACRM-NNE model can handle nonlinear problems and improve fitting capabilities; loss1 is calculated after the signal passes through the encoder, and the signal is input into N2 decoders, in which a multi-head attention layer is also used to process the signal; the output of the decoder extracts features through a linear layer to obtain the output y of the entire model. * , and add the calculated loss2 to loss1 as the total loss.

[0021] Furthermore, in step S2, the multi-head attention layer includes:

[0022]

[0023] in

[0024] Where Z is the value vector output by the multi-head attention layer, (·) T is the transpose of the matrix, Q, K, V are the query vector, key vector and value vector respectively, W Q , W K , W V are the query weight matrix, key weight matrix, and value weight matrix respectively, X is the value vector input to the multi-head attention layer, and d k is the dimension of the key vector. The query vector, key vector and value vector in the above formula are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix, which respectively represent the characteristics of the current value vector, the influence of other value vectors on the current value vector and the linear change of the input value vector. The inner product of the query vector at the current moment and the key vector at other moments represents the influence of other value vectors on the current value vector, divided by Ensure that the gradient of the training process is stable, and normalize it through the softmax function and multiply it with the corresponding value vector to obtain the value vector processed by the attention layer.

[0025] Furthermore, in step S2, a channel parameter γ is introduced into the encoding device, and the channel parameter can be adaptively updated as the network parameters change; at this time, the encoder label is:

[0026] y A (n) = y(n) + γ·y(n-1).

[0027] Further, in step S4, equalization by adaptive channel response matching neural network equalizer includes:

[0028] The impulse response of the channel is hch (n), the received signal can be expressed as in is the convolution symbol, N(n) is the additive noise, and y(n) is the transmitted signal; the impulse response of the adaptive channel response matching neural network equalizer is where h FR (n), h PF (n) and h MLSE (n) are the impulse responses of full response equalization, post-filtering and MLSE respectively; the signal equalization process can be expressed as:

[0029]

[0030] in is the convolution symbol, y * (n) is the equalized signal, and x(n) is the signal to be equalized.

[0031] In addition, the present invention also provides an adaptive channel response matching neural network equalization system, comprising:

[0032] Data module: obtain signal data, pre-process the signal data and divide it into training set and test set;

[0033] Training module: Design an ACRM-NNE model, train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; Train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: multi-head attention layer, feedforward network layer, residual and normalization;

[0034] Testing module: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to the execution training module. If the bit error rate is greater than or equal to the set value, the ACRM-NNE training model is used as an adaptive channel response matching neural network equalizer.

[0035] Equalization module: Equalization is performed through adaptive channel response matching neural network equalizer.

[0036] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0037] The present invention trains the ACRM-NNE model multiple times to obtain an adaptive channel response matching neural network equalizer, which can effectively process bandwidth-limited signals and is suitable for pulse amplitude modulation signal equalization with inter-symbol crosstalk. The multi-head attention layer in the ACRM-NNE model can effectively take into account different influencing factors, greatly improving the processing capability of time series; the encoder and decoder structure in the ACRM-NNE model can replace the operating mode of partial response equalization and MLSE decoding, and simplify the digital signal processing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the adaptive channel response matching neural network equalization method described in Example 1;

[0039] Figure 2 This is a schematic diagram of the optical fiber communication system described in Example 2;

[0040] Figure 3 This is a schematic diagram of the ACRM-NNE model structure described in Example 2;

[0041] Figure 4 This is a block diagram of the adaptive channel response matching neural network equalization system described in Example 3; DETAILED DESCRIPTION

[0042] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0043] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0044] Embodiment 1:

[0045] This embodiment provides Figure 1 The adaptive channel response matching neural network equalization method shown includes:

[0046] S1: Acquire signal data, and divide the signal data into training set and test set after preprocessing;

[0047] S2: Design an ACRM-NNE model, train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; the ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: a multi-head attention layer, a feedforward network layer, a residual, and normalization;

[0048] S3: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to step S2. If the bit error rate is greater than or equal to the set value, use the ACRM-NNE training model as an adaptive channel response matching neural network equalizer.

[0049] S4: Equalization via adaptive channel response matching neural network equalizer.

[0050] In this embodiment, the adaptive channel response matching neural network equalizer obtained by training the ACRM-NNE model multiple times can effectively process bandwidth-limited signals and is suitable for pulse amplitude modulation signal equalization with inter-symbol interference. The multi-head attention layer in the ACRM-NNE model can effectively take into account different influencing factors and greatly improve the processing capability of time series. The encoder and decoder structure in the ACRM-NNE model can replace the operating mode of partial response equalization and MLSE decoding and simplify the digital signal processing process.

[0051] Embodiment 2:

[0052] This embodiment further discloses on the basis of the first embodiment:

[0053] Furthermore, in step S1, it includes:

[0054] Build as Figure 2 The optical fiber communication system shown in the figure includes a transmitting end and a receiving end; in the transmitting end, the transmission data is mapped into a PAM-4 electrical signal, which is converted into an optical signal through an arbitrary waveform transmitter and a Mach-Zehnder modulator after upsampling and pulse shaping; in the receiving end, the optical signal is converted into an electrical signal by a photodetector, digitized by an analog-to-digital converter, and obtained through resampling, matched filtering and synchronization.

[0055] Data collection: Save the synchronized signal from the receiving end as data to be equalized;

[0056] Data preprocessing: perform data preprocessing on the data to be equalized obtained during data collection, and form an array of length 2L+1 with the signal at the current moment and the signals at L moments before and after the current moment as the processed data. The processed training data is divided into a training set and a test set at a ratio of 70%:30%, with the training set accounting for 70% and the test set accounting for 30%.

[0057] Furthermore, in data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end so that the bandwidth is limited, and severe ISI is introduced into the communication signal; the signal with severe ISI introduced into the receiving end is saved as data to be equalized.

[0058] Furthermore, in step S2, the ACRM-NNE model includes:

[0059] The loss function loss1 is designed in the encoder to make the encoder equivalent to the functions of full response equalization and post-filtering. The loss function loss2 is designed in the decoder to output an equalized signal without ISI. The sum of the two losses is used as the total loss function of the ACRM-NNE model.

[0060] Further, in step S2, the ACRM-NNE model is trained by the training set, and obtaining the ACRM-NNE training model includes: initializing the ACRM-NNE model using the Kaiming initialization method, and assigning values ​​to the parameters required in the training process; calculating the total loss, that is, the mean square error, which is specifically calculated as follows:

[0061]

[0062] in and A are the encoder output and its label, y * and y are the encoder output and its label respectively, and the Adam optimizer is used to update the ACRM-NNE model parameters through the back-propagation algorithm until the loss function converges and the training is stopped to obtain the ACRM-NNE model.

[0063] Further, in step S2, if Figure 3 The ACRM-NNE model shown in the figure includes: the linear layer converts the input signal into a value vector, which is input into N1 encoders for processing. The multi-head attention layer is used in the encoder to capture the dependencies between different positions, and then the residual is introduced into the normalization layer to avoid degradation during network training. A nonlinear activation function is introduced into the feedforward network, so that the ACRM-NNE model can handle nonlinear problems and improve fitting capabilities; after the signal passes through the encoder, loss1 is calculated, and the signal is input into N2 decoders, where the multi-head attention layer is also used to process the signal; the output of the decoder extracts features through the linear layer to obtain the output y of the entire model * , and add the calculated loss2 to loss1 as the total loss.

[0064] Furthermore, in step S2, the multi-head attention layer includes:

[0065]

[0066] in

[0067] Where Z is the value vector output by the multi-head attention layer, (·) T is the transpose of the matrix, Q, K, V are the query vector, key vector and value vector respectively, W Q , W K , W Vare the query weight matrix, key weight matrix, and value weight matrix respectively, X is the value vector input to the multi-head attention layer, and d k is the dimension of the key vector. The query vector, key vector and value vector in the above formula are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix, which respectively represent the characteristics of the current value vector, the influence of other value vectors on the current value vector and the linear change of the input value vector. The inner product of the query vector at the current moment and the key vector at other moments represents the influence of other value vectors on the current value vector, divided by Ensure that the gradient of the training process is stable, and normalize it through the softmax function and multiply it with the corresponding value vector to obtain the value vector processed by the attention layer.

[0068] Furthermore, in step S2, a channel parameter γ is introduced into the encoding device, and the channel parameter can be adaptively updated as the network parameters change; at this time, the encoder label is:

[0069] y A (n) = y(n) + γ·y(n-1).

[0070] Further, in step S4, equalization by adaptive channel response matching neural network equalizer includes:

[0071] The impulse response of the channel is h ch (n), the received signal can be expressed as in is the convolution symbol, N(n) is the additive noise, and y(n) is the transmitted signal; the impulse response of the adaptive channel response matching neural network equalizer is where h FR (n), h PF (n) and h MLSE (n) are the impulse responses of full response equalization, post-filtering and MLSE respectively; the signal equalization process can be expressed as:

[0072]

[0073]

[0074] in is the convolution symbol, y * (n) is the equalized signal, and x(n) is the signal to be equalized.

[0075] In this embodiment, the adaptive channel response matching neural network equalizer obtained by training the ACRM-NNE model multiple times can effectively process bandwidth-limited signals and is suitable for pulse amplitude modulation signal equalization with inter-symbol interference. The multi-head attention layer in the ACRM-NNE model can effectively take into account different influencing factors and greatly improve the processing capability of time series. The encoder and decoder structure in the ACRM-NNE model can replace the operating mode of partial response equalization and MLSE decoding and simplify the digital signal processing process.

[0076] Embodiment three:

[0077] This embodiment also provides Figure 4 An adaptive channel response matching neural network equalization system is shown, comprising:

[0078] Data module: obtain signal data, pre-process the signal data and divide it into training set and test set;

[0079] Training module: Design an ACRM-NNE model, train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; Train the ACRM-NNE model with the training set, and obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: multi-head attention layer, feedforward network layer, residual and normalization;

[0080] Testing module: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to the execution training module. If the bit error rate is greater than or equal to the set value, the ACRM-NNE training model is used as an adaptive channel response matching neural network equalizer.

[0081] Equalization module: Equalization is performed through adaptive channel response matching neural network equalizer.

[0082] In this embodiment, the adaptive channel response matching neural network equalizer obtained by training the ACRM-NNE model multiple times can effectively process bandwidth-limited signals and is suitable for pulse amplitude modulation signal equalization with inter-symbol interference. The multi-head attention layer in the ACRM-NNE model can effectively take into account different influencing factors and greatly improve the processing capability of time series. The encoder and decoder structure in the ACRM-NNE model can replace the operating mode of partial response equalization and MLSE decoding and simplify the digital signal processing process.

[0083] Embodiment 4:

[0084] This embodiment provides a specific embodiment of an ACRM-NNE-based balancing solution, including:

[0085] Step 1: Dataset creation. The optical signal is generated and transmitted in the system, converted into an electrical signal by a photodetector, and then used as the training set and test set of ACRM-NNE after digital signal processing and data preprocessing. The original PAM-4 signal sent by the transmitter is used as the label of the training set and test set;

[0086] Step 2: Design ACRM-NNE, train ACRM-NNE using the training set obtained in step 1, and save the training model;

[0087] Step 3: Test ACRM-NNE. ACRM-NNE loads the training model in step 2, and uses the model to equalize the test set data obtained in step 1, then calculates its bit error rate to obtain the performance of ACRM-NNE.

[0088] Step 1: Create a data set. The specific steps are as follows:

[0089] Build a fiber-optic communication system. The communication system consists of two parts, the transmitter and the receiver. In the transmitter, the transmission data is mapped to a PAM-4 electrical signal, which is converted into an optical signal through an arbitrary waveform transmitter and a Mach-Zehnder modulator after upsampling and pulse shaping. At the receiver, the optical signal is converted into an electrical signal by a photodetector, digitized by an analog-to-digital converter, and then resampled, matched filtered, and synchronized to obtain the signal to be equalized.

[0090] Data collection. First, by increasing the baud rate of the PAM-4 signal to exceed the device bandwidth and limit the system bandwidth, severe ISI is introduced into the communication signal; second, the PAM-4 data obtained after signal mapping is saved as a label; finally, the signal after synchronization at the receiving end is saved as data to be equalized;

[0091] Data preprocessing. The data to be equalized obtained in data collection is preprocessed as follows. First, the signal at the current moment and the signals at L moments before and after the current moment are combined into an array of length 2L+1 as the model input at the current moment; finally, the processed training data is divided into a training set and a test set in a ratio of 70%:30%.

[0092] Step 2: Design ACRM-NNE. The specific steps are as follows:

[0093] Model design. The network structure of ACRM-NNE is as follows: First, the signal features are extracted through a linear layer, and the number of neural units is 128×1; then the features are successively passed to three encoders, in which a multi-head attention layer with 8 attention heads is used to capture the influence relationship between different moments, and the fully connected layer + ReLU activation function is used in the feedforward network to extract features; finally, the linear layer is used to extract features to obtain the first loss calculation. The number of neural units is 32×1. Subsequently, the signal features are extracted again through the linear layer, and the number of neural units is 64×1. The features are successively passed to three decoders, in which the multi-head attention layer with 8 attention heads is also used, and the feedforward network uses the fully connected layer + ReLU activation function to extract features. Finally, the linear layer is used to extract features to obtain the output of the model, and the number of neural units is 32×1.

[0094] Training scheme design. The ACRM-NNE model uses Kaiming initialization to initialize the parameters in the model; the Adam optimizer is used as the gradient descent algorithm with an initial learning rate of 0.001; MSE is used as the loss function to evaluate the difference between the predicted value and the label to reflect the performance of the model during training, and the model parameters are updated through the back-propagation algorithm until the loss function converges.

[0095] In order to better implement this embodiment, in the above method, step 3 tests ACRM-NNE, and the specific steps are as follows:

[0096] ACRM-NNE loads the training parameters from step 2.

[0097] Use the test set preprocessed in step 1 to test ACRM-NNE, equalize the signal with ISI and calculate its bit error rate. If its bit error rate performance is better than the traditional solution, it means that the ACRM-NNE equalization scheme has a better processing effect on the signal ISI problem, and its codec structure simplifies the digital signal processing process; if its bit error rate performance is poor, it is necessary to return to step 2, redesign the ACRM-NNE model, and retrain until the bit error rate performance is better than the traditional solution.

[0098] In this embodiment, the adaptive channel response matching neural network equalizer obtained by training the ACRM-NNE model multiple times can effectively process bandwidth-limited signals and is suitable for pulse amplitude modulation signal equalization with inter-symbol interference. The multi-head attention layer in the ACRM-NNE model can effectively take into account different influencing factors and greatly improve the processing capability of time series. The encoder and decoder structure in the ACRM-NNE model can replace the operating mode of partial response equalization and MLSE decoding and simplify the digital signal processing process.

[0099] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. Adaptive channel response matching neural network equalization method, characterized in that: include: S1: Acquire signal data, and divide the signal data into training set and test set after preprocessing; S2: Design an ACRM-NNE model, train the ACRM-NNE model using the training set, and obtain an ACRM-NNE training model; The ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: multi-head attention layers, feed-forward network layers, residuals, and normalization; S3: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to step S2. If the bit error rate is greater than or equal to the set value, use the ACRM-NNE training model as an adaptive channel response matching neural network equalizer. S4: Equalization via adaptive channel response matching neural network equalizer.

2. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that: In step S1, it includes: Build an optical fiber communication system, which includes a transmitter and a receiver. In the transmitter, the transmission data is mapped to a PAM-4 electrical signal, which is converted into an optical signal through an arbitrary waveform transmitter and a Mach-Zehnder modulator after upsampling and pulse shaping. In the receiver, the optical signal is converted into an electrical signal by a photodetector, digitized by an analog-to-digital converter, and then resampled, matched filtered, and synchronized to obtain a received signal. Data collection: Save the synchronized signal from the receiving end as data to be equalized; Data preprocessing: perform data preprocessing on the data to be equalized obtained during data collection, and form an array of length 2L+1 with the signal at the current moment and the signals at L moments before and after the current moment as the processed data. The processed training data is divided into a training set and a test set at a ratio of 70%:30%, with the training set accounting for 70% and the test set accounting for 30%.

3. The adaptive channel response matching neural network equalization method according to claim 2, characterized in that: In data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end, which limits the bandwidth and introduces severe ISI to the communication signal; the signal with severe ISI introduced at the receiving end is saved as data to be equalized.

4. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that: In step S2, the ACRM-NNE model includes: The loss function loss1 is designed in the encoder to make the encoder equivalent to the functions of full response equalization and post-filtering. The loss function loss2 is designed in the decoder to output an equalized signal without ISI. The sum of the two losses is used as the total loss function of the ACRM-NNE model.

5. The adaptive channel response matching neural network equalization method according to claim 4, characterized in that: In step S2, the ACRM-NNE model is trained by the training set, and the ACRM-NNE training model is obtained, including: initializing the ACRM-NNE model using the Kaiming initialization method, and assigning values ​​to the parameters required in the training process; calculating the total loss, that is, the mean square error, which is specifically calculated as follows: in and A are the encoder output and its label, y * and y are the encoder output and its label respectively, and the Adam optimizer is used to update the ACRM-NNE model parameters through the back-propagation algorithm until the loss function converges and the training is stopped to obtain the ACRM-NNE model.

6. The adaptive channel response matching neural network equalization method according to claim 5, characterized in that: In step S2, the ACRM-NNE model includes: the linear layer converts the input signal into a value vector, which is input into N1 encoders for processing. The multi-head attention layer is used in the encoder to capture the dependencies between different positions, and then the residual is introduced into the normalization layer to avoid degradation during network training. A nonlinear activation function is introduced into the feedforward network, so that the ACRM-NNE model can handle nonlinear problems and improve fitting capabilities; after the signal passes through the encoder, loss1 is calculated, and the signal is input into N2 decoders, in which the multi-head attention layer is also used to process the signal; the output of the decoder is extracted through the linear layer to obtain the output y of the entire model * , and add the calculated loss2 to loss1 as the total loss.

7. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that: In step S2, the multi-head attention layer includes: in Where Z is the value vector output by the multi-head attention layer, (·) T is the transpose of the matrix, Q, K, V are the query vector, key vector and value vector respectively, W Q , W K , W V are the query weight matrix, key weight matrix, and value weight matrix respectively, X is the value vector input to the multi-head attention layer, and d k is the dimension of the key vector. The query vector, key vector and value vector in the above formula are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix, which respectively represent the characteristics of the current value vector, the influence of other value vectors on the current value vector and the linear change of the input value vector. The inner product of the query vector at the current moment and the key vector at other moments represents the influence of other value vectors on the current value vector, divided by Ensure that the gradient of the training process is stable, and normalize it through the softmax function and multiply it with the corresponding value vector to obtain the value vector processed by the attention layer.

8. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that: In step S2, the channel parameter γ is introduced into the encoding device, and the channel parameter can be adaptively updated as the network parameters change; at this time, the encoder label is: and A (n)=y(n)+γ·y(n-1)。 9. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that: In step S4, equalization by adaptive channel response matching neural network equalizer includes: The impulse response of the channel is h ch (n), the received signal can be expressed as in is the convolution symbol, N(n) is the additive noise, y(n) is the transmitted signal; the impulse response of the adaptive channel response matching neural network equalizer is where h FR (n), h PF (n) and h MLSE (n) are the impulse responses of full response equalization, post-filtering and MLSE respectively; the signal equalization process can be expressed as: in is the convolution symbol, y * (n) is the equalized signal, and x(n) is the signal to be equalized.

10. Adaptive channel response matching neural network equalization system, characterized in that: include: Data module: obtain signal data, pre-process the signal data and divide it into training set and test set; Training module: design the ACRM-NNE model, train the ACRM-NNE model through the training set, and obtain the ACRM-NNE training model; Train the ACRM-NNE model through the training set to obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascade of multiple encoders and decoders, including: multi-head attention layers, feed-forward network layers, residuals, and normalization; Testing module: Test the performance of the ACRM-NNE training model through the test set to obtain the bit error rate. If the bit error rate is lower than the set value, return to the execution training module. If the bit error rate is greater than or equal to the set value, the ACRM-NNE training model is used as an adaptive channel response matching neural network equalizer. Equalization module: Equalization is performed through adaptive channel response matching neural network equalizer.

Images