Adaptive channel response matching neural network equalization method and system

Abstract

This invention relates to the field of optical fiber communication technology, and more specifically, to an adaptive channel response matching neural network equalization method and system. The method includes: acquiring signal data; preprocessing the signal data and dividing it into a training set and a test set; designing an ACRM-NNE model; training the model using the training set to obtain an ACRM-NNE training model; testing the performance using the test set; and using the ACRM-NNE training model as an adaptive channel response matching neural network equalizer; and performing equalization using the adaptive channel response matching neural network equalizer. The adaptive channel response matching neural network equalizer of this invention can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. The multi-head attention layer can effectively consider different influencing factors, greatly improving the processing capability for time series signals. The cascaded encoder and decoder structure in the ACRM-NNE model can replace the operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow.

Description

Technical Field

[0001] This 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 Technology

[0002] In recent years, with the development of new applications such as the Internet of Things (IoT) and cloud computing, network traffic has grown exponentially. These new applications have brought about demands for bandwidth and computing resources, driving the development of short-distance (<80 km) fiber optic communication systems for data centers. At the same time, the ever-growing data center business has prompted operators, enterprises, and service providers to offer Ethernet solutions for high-speed network communication. Currently, data center interconnection rates are transitioning from 400Gb / s to 800Gb / s. Whether at 400Gb / s or 800Gb / s, intensity modulation direct detection is widely used due to its advantages such as low complexity, low cost, and energy efficiency. The currently deployed 400Gbit / s Ethernet (400GbE) uses a 4-channel 53Gbaud four-level pulse amplitude modulation (PAM-4) transmission scheme. For next-generation Ethernet "beyond 400GbE" (such as 800GbE or 1.6TbE), the PAM-4 modulation format remains a promising solution due to its low complexity and ease of upgrading from existing 400GbE infrastructure. To achieve greater transmission capacity to meet the ever-increasing traffic demands, it is necessary to increase the number of wavelength division multiplexing (WDM) signals and transmit PAM-4 signals at higher baud rates. However, due to the bandwidth limitations of transceiver devices, increasing the number of WDM signals or transmitting at higher baud rates will lead to severe inter-satellite interference (ISI) in the PAM-4 signals, thus impairing communication quality.

[0003] Existing technology discloses a QPSK modulation super Nyquist transmission method and system based on neural network equalization. This method introduces convolutional coding and neural network equalization into the super Nyquist transmission process. At the transmitter, data is first channel-coded and then QPSK modulated, splitting the original information into I-paths and Q-paths. Both data paths are simultaneously subjected to FTN pulse shaping and sent to the channel for transmission. At the receiver, the two signals transmitted through the channel are simultaneously matched-filtered to recover the waveforms of the original two transmitted data paths. After waveform recovery, FTN sampling is performed according to the shaping pulse interval to obtain sampled values. These sampled signals are then fed into a neural network equalizer for equalization, obtaining equalized sample values. Demapping is then performed, resulting in a soft output. Finally, soft-decision Viterbi decoding is performed to recover the original data. However, this method still requires the use of an MLSE decoder to restore the ISI introduced by the partial response equalizer, ultimately complicating the digital signal processing flow. Furthermore, conventional neural network equalization schemes lack the ability to address time-series information; therefore, the ability to equalize ISI needs improvement in time-series processing tasks. Summary of the Invention

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

[0005] To achieve the above objectives, the present invention provides an adaptive channel response matching neural network equalization method, comprising:

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

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

[0008] S3: Test the performance of the ACRM-NNE training model using the test set and 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 is achieved through an adaptive channel response matching neural network equalizer.

[0010] Further, step S1 includes:

[0011] An optical fiber communication system is constructed, which includes a transmitter and a receiver. At the transmitter, the transmitted data is mapped into a PAM-4 electrical signal, which is then upsampled and pulse shaped before being converted into an optical signal by an arbitrary waveform transmitter and a Mach-Zehnder modulator. 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 received signal.

[0012] Data collection involves saving the synchronized signal from the receiving end as the data to be equalized.

[0013] Data preprocessing involves preprocessing the data to be balanced obtained from data collection. The signal at the current time is combined with the signals at each of the previous and next L times to form an array of length 2L+1, which is used as the processed data. The processed training data is then divided into a training set and a test set in a 70%:30% ratio, with the training set accounting for 70% and the test set accounting for 30%.

[0014] Furthermore, during data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end, thus limiting the bandwidth and introducing severe ISI into the communication signal. The signal with severe ISI introduced at the receiving end is saved as data to be equalized.

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

[0016] In the encoder, a loss function loss1 is designed to make the encoder equivalent to the functions of full response equalization and post-filtering; in the decoder, a loss function loss2 is designed 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, training the ACRM-NNE model using the training set to obtain the ACRM-NNE training model includes: initializing the ACRM-NNE model using the Kaiming initialization method and assigning values ​​to the parameters required during training; calculating the total loss, i.e., the mean squared error, specifically as follows:

[0018]

[0019] in and y A These are the encoder output and its label, y * Let y be the encoder output and its label, respectively. The ACRM-NNE model parameters are updated using the Adam optimizer through backpropagation algorithm until the loss function converges, at which point training stops and the ACRM-NNE model is obtained.

[0020] Further, in step S2, the ACRM-NNE model includes: a linear layer converting the input signal into a value vector, which is then fed into N1 encoders for processing. The encoders use multi-head attention layers to capture dependencies between different positions, followed by a normalization layer to introduce residuals, preventing degradation during network training. A non-linear activation function is also introduced into the feedforward network, enabling the ACRM-NNE model to handle non-linear problems and improve its fitting ability. After the signal passes through the encoder, loss1 is calculated, and the signal is input into N2 decoders, which also use multi-head attention layers to process the signal. The decoder output is processed through a linear layer to extract features, obtaining the output y of the entire model. * The calculated loss2 is added to loss1 to obtain the total loss.

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

[0022]

[0023] in

[0024] Where Z is the value vector of the multi-head attention layer output, (·) T Let Q be the transpose of the matrix, K, and V be the query vector, key vector, and value vector, respectively, and W be the value vector. Q W K W V These 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 Let be the dimension of the key vector. In the above formula, the query vector, key vector, and value vector are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix. These represent the features 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, respectively. The inner product of the query vector at the current time step and the key vector at other time steps represents the influence of other value vectors on the current value vector. Dividing by... To ensure gradient stability during training, the gradient is normalized using the softmax function and then multiplied by 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, which can adaptively update 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, the equalization process using an adaptive channel response matching neural network equalizer includes:

[0028] The impulse response of the channel is hch If (n), then the received signal can be expressed as in Let y(n) be the convolution symbol, N(n) be the additive noise, and y(n) be 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) represent the impulse responses of full-response equalization, post-filtering, and MLSE, respectively; the signal equalization process can be expressed as:

[0029]

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

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

[0032] Data module: Acquires signal data, preprocesses the signal data, and divides it into training and test sets;

[0033] Training module: Design the ACRM-NNE model, train the ACRM-NNE model using the training set to obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascaded multi-encoder and decoder, including: multi-head attention layer, feedforward network layer, residual and normalization.

[0034] Test module: Test the performance of the ACRM-NNE training model through the test set and obtain the bit error rate. If the bit error rate is lower than the set value, return to execute the 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 using an adaptive channel response matching neural network equalizer.

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

[0037] This invention obtains an adaptive channel response matching neural network equalizer by training the ACRM-NNE model multiple times. This equalizer can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. 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 signals. The encoder and decoder structure in the ACRM-NNE model can replace the operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow. Attached Figure Description

[0038] Figure 1 The flowchart is shown in Embodiment 1, illustrating the adaptive channel response matching neural network equalization method.

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

[0042] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Example 1:

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

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

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

[0048] S3: Test the performance of the ACRM-NNE training model using the test set and 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 is achieved through an adaptive channel response matching neural network equalizer.

[0050] This embodiment obtains an adaptive channel response matching neural network equalizer by training the ACRM-NNE model multiple times. This equalizer can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. 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 operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow.

[0051] Example 2:

[0052] This embodiment further discloses information based on Embodiment 1:

[0053] Further, step S1 includes:

[0054] Building such Figure 2 The optical fiber communication system shown includes a transmitter and a receiver. At the transmitter, the transmitted data is mapped into a PAM-4 electrical signal, which is then upsampled and pulse shaped before being converted into an optical signal by an arbitrary waveform transmitter and a Mach-Zehnder modulator. 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 received signal.

[0055] Data collection involves saving the synchronized signal from the receiving end as the data to be equalized.

[0056] Data preprocessing involves preprocessing the data to be balanced obtained from data collection. The signal at the current time is combined with the signals at each of the previous and next L times to form an array of length 2L+1, which is used as the processed data. The processed training data is then divided into a training set and a test set in a 70%:30% ratio, with the training set accounting for 70% and the test set accounting for 30%.

[0057] Furthermore, during data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end, thus limiting the bandwidth and introducing severe ISI into the communication signal. The signal with severe ISI introduced at the receiving end is saved as data to be equalized.

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

[0059] In the encoder, a loss function loss1 is designed to make the encoder equivalent to the functions of full response equalization and post-filtering; in the decoder, a loss function loss2 is designed 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, training the ACRM-NNE model using the training set to obtain the ACRM-NNE training model includes: initializing the ACRM-NNE model using the Kaiming initialization method and assigning values ​​to the parameters required during training; calculating the total loss, i.e., the mean squared error, specifically as follows:

[0061]

[0062] in and y A These are the encoder output and its label, y * Let y be the encoder output and its label, respectively. The ACRM-NNE model parameters are updated using the Adam optimizer through backpropagation algorithm until the loss function converges, at which point training stops and the ACRM-NNE model is obtained.

[0063] Further, in step S2, as Figure 3 The ACRM-NNE model shown includes: a linear layer converts the input signal into a value vector, which is then fed into N1 encoders for processing. The encoders use multi-head attention layers to capture dependencies between different locations. Residuals are then introduced into the normalization layer to prevent degradation during network training. A non-linear activation function is introduced into the feedforward network, enabling the ACRM-NNE model to handle non-linear problems and improve its fitting ability. After the signal passes through the encoder, loss1 is calculated, and the signal is then input into N2 decoders, which also use multi-head attention layers to process the signal. The decoder outputs are processed through a linear layer to extract features, yielding the overall model output y. * The calculated loss2 is added to loss1 to obtain the total loss.

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

[0065]

[0066] in

[0067] Where Z is the value vector of the multi-head attention layer output, (·) T Let Q be the transpose of the matrix, K, and V be the query vector, key vector, and value vector, respectively, and W be the value vector. Q W K W VThese 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 Let be the dimension of the key vector. In the above formula, the query vector, key vector, and value vector are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix. These represent the features 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, respectively. The inner product of the query vector at the current time step and the key vector at other time steps represents the influence of other value vectors on the current value vector. Dividing by... To ensure gradient stability during training, the gradient is normalized using the softmax function and then multiplied by 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, which can adaptively update 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, the equalization process using an adaptive channel response matching neural network equalizer includes:

[0071] The impulse response of the channel is h ch If (n), then the received signal can be expressed as in Let y(n) be the convolution symbol, N(n) be the additive noise, and y(n) be 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) represent the impulse responses of full-response equalization, post-filtering, and MLSE, respectively; the signal equalization process can be expressed as:

[0072]

[0073]

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

[0075] This embodiment obtains an adaptive channel response matching neural network equalizer by training the ACRM-NNE model multiple times. This equalizer can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. 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 operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow.

[0076] Example 3:

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

[0078] Data module: Acquires signal data, preprocesses the signal data, and divides it into training and test sets;

[0079] Training module: Design the ACRM-NNE model, train the ACRM-NNE model using the training set to obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascaded multi-encoder and decoder, including: multi-head attention layer, feedforward network layer, residual and normalization.

[0080] Test module: Test the performance of the ACRM-NNE training model through the test set and obtain the bit error rate. If the bit error rate is lower than the set value, return to execute the 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 using an adaptive channel response matching neural network equalizer.

[0082] This embodiment obtains an adaptive channel response matching neural network equalizer by training the ACRM-NNE model multiple times. This equalizer can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. 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 operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow.

[0083] Example 4:

[0084] This embodiment provides a specific implementation of an equalization scheme based on ACRM-NNE, including:

[0085] Step 1: Dataset Creation. Optical signals are generated and transmitted within the system, converted into electrical signals by a photodetector, and then processed through digital signal processing and data preprocessing to serve as the training and test sets for ACRM-NNE. The raw PAM-4 signals transmitted from the transmitter are used as labels for both the training and test sets.

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

[0087] Step 3: Test ACRM-NNE. ACRM-NNE loads the trained model from Step 2 and uses the model to balance the test set data obtained in Step 1. Then, it calculates the bit error rate to obtain the performance of ACRM-NNE.

[0088] Step 1 involves creating the dataset, and the specific steps are as follows:

[0089] A fiber optic communication system is constructed. The system consists of two parts: a transmitter and a receiver. At the transmitter, transmitted data is mapped to a PAM-4 electrical signal, which, after upsampling and pulse shaping, is converted into an optical signal by an arbitrary waveform transmitter and a Mach-Zehnder modulator. 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 tags. Finally, the signal synchronized at the receiving end is saved as data to be equalized.

[0091] Data preprocessing. The data to be balanced obtained from the data collection will be preprocessed as follows: First, the signal at the current time and the signals at each of the L time points before and after the current time will be combined to form an array of length 2L+1, which will be used as the model input at the current time. Finally, the processed training data will be divided into training set and test set in a 70%:30% ratio.

[0092] Step two involves designing ACRM-NNE, with the specific steps as follows:

[0093] Model Design. The ACRM-NNE network structure is as follows: First, signal features are extracted through linear layers, with 128×1 neurons; then, the features are sequentially fed into three encoders, where a multi-head attention layer with eight attention heads captures the influence relationships between different time points, and features are extracted in the feedforward network using a fully connected layer + ReLU activation function; finally, features are extracted using linear layers to obtain the first loss calculation. The number of neurons is 32×1; then, a linear layer is used to extract signal features again, with a number of neurons of 64×1; the features are then fed into three decoders, including a multi-head attention layer with eight attention heads, and a feedforward network that uses a fully connected layer with a ReLU activation function to extract features; finally, a linear layer is used to extract features to obtain the model output, with a number of neurons of 32×1.

[0094] Training scheme design. The ACRM-NNE model uses the Kaiming initialization method 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; the MSE is used as the loss function to evaluate the difference between the predicted value and the label to reflect the model's performance during training, and the model parameters are updated through the backpropagation algorithm until the loss function converges.

[0095] To better implement this embodiment, in the above method, step three involves testing ACRM-NNE, and the specific steps are as follows:

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

[0097] The ACRM-NNE is tested using the pre-processed test set from step one. Signals with ISI are equalized and their bit error rate is calculated. If its bit error rate performance is better than that of the traditional scheme, it indicates that the ACRM-NNE equalization scheme is better at handling the ISI problem of signals, and its codec structure simplifies the digital signal processing flow. If its bit error rate performance is poor, it is necessary to return to step two, redesign the ACRM-NNE model, and retrain it until the bit error rate performance is better than that of the traditional scheme.

[0098] This embodiment obtains an adaptive channel response matching neural network equalizer by training the ACRM-NNE model multiple times. This equalizer can effectively handle bandwidth-limited signals and is suitable for equalizing pulse amplitude modulation signals with inter-symbol interference. 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 operation mode of partial response equalization and MLSE decoding, simplifying the digital signal processing flow.

[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 implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An adaptive channel response matching neural network equalization method, characterized in that, include: S1: Acquire signal data, preprocess the signal data and divide it into training set and test set; S2: Design the ACRM-NNE model, train the ACRM-NNE model using the training set, and obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascaded multi-encoder and decoder, including: a multi-head attention layer, a feedforward network layer, residuals, and normalization. The ACRM-NNE model includes: In the encoder, a loss function loss1 is designed to make the encoder equivalent to the functions of full response equalization and post-filtering; in the decoder, a loss function loss2 is designed to output an equalized signal without ISI; the sum of the two losses is used as the total loss function L of the ACRM-NNE model. The linear layer converts the input signal into a value vector, which is then processed by N1 encoders. The encoders use multi-head attention layers to capture dependencies between different locations. Residuals are then introduced into the normalization layer to prevent degradation during network training. A non-linear activation function is introduced into the feedforward network, enabling the ACRM-NNE model to handle non-linear problems and improve its fitting ability. After passing through the encoders, the signal's loss1 is calculated, and the signal is then input into N2 decoders, which also use multi-head attention layers for signal processing. The decoder outputs are processed through a linear layer to extract features, yielding the overall model output y. * The calculated loss2 is added to loss1 to obtain the total loss; S3: Test the performance of the ACRM-NNE training model using the test set and 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 is achieved through an adaptive channel response matching neural network equalizer.

2. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that, Step S1 includes: An optical fiber communication system is constructed, which includes a transmitter and a receiver. At the transmitter, the transmitted data is mapped into a PAM-4 electrical signal, which is then upsampled and pulse shaped before being converted into an optical signal by an arbitrary waveform transmitter and a Mach-Zehnder modulator. 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 received signal. Data collection involves saving the synchronized signal from the receiving end as the data to be equalized. Data preprocessing involves preprocessing the data to be balanced obtained from data collection. The signal at the current time is combined with the signals at each of the previous and next L times to form an array of length 2L+1, which is used as the processed data. The processed training data is then divided into a training set and a test set in a 70%:30% ratio, 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, During data collection, the baud rate of the PAM-4 signal is increased to exceed the bandwidth of the receiving end, thus limiting the bandwidth and introducing severe ISI into the communication signal. The signal with severe ISI 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, training the ACRM-NNE model using the training set to obtain the ACRM-NNE training model includes: initializing the ACRM-NNE model using the Kaiming initialization method and assigning values ​​to the parameters required during training; calculating the total loss, i.e., the mean squared error, specifically as follows: in and y A These are the encoder output and its label, y * Let y be the encoder output and its label, respectively. The ACRM-NNE model parameters are updated using the Adam optimizer through backpropagation algorithm until the loss function converges, at which point training stops and the ACRM-NNE model is obtained.

5. 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 of the multi-head attention layer output, (·) T Let Q be the transpose of the matrix, K, and V be the query vector, key vector, and value vector, respectively, and W be the value vector. Q W K W V These 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 Let be the dimension of the key vector. In the above formula, the query vector, key vector, and value vector are obtained by multiplying the value vector of the input attention layer by the corresponding weight matrix. These represent the features 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, respectively. The inner product of the query vector at the current time step and the key vector at other time steps represents the influence of other value vectors on the current value vector. Dividing by... To ensure gradient stability during training, the gradient is normalized using the softmax function and then multiplied by the corresponding value vector to obtain the value vector processed by the attention layer.

6. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that, In step S2, a channel parameter γ is introduced into the encoding device. This 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)。 7. The adaptive channel response matching neural network equalization method according to claim 1, characterized in that, In step S4, equalization is performed using an adaptive channel response matching neural network equalizer, including: The impulse response of the channel is h ch If (n), then the received signal can be expressed as in Let y(n) be the convolution symbol, N(n) be the additive noise, and y(n) be 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) represent the impulse responses of full-response equalization, post-filtering, and MLSE, respectively; the signal equalization process can be expressed as: in The symbol for convolution is y. * x(n) is the equalized signal, and x(n) is the signal to be equalized.

8. An adaptive channel response matching neural network equalization system, characterized in that, include: Data module: Acquires signal data, preprocesses the signal data, and divides it into training and test sets; Training module: Design the ACRM-NNE model, train the ACRM-NNE model using the training set, and obtain the ACRM-NNE training model; The ACRM-NNE model is trained using the training set to obtain the ACRM-NNE training model; The ACRM-NNE model consists of a cascaded multi-encoder and decoder, including: a multi-head attention layer, a feedforward network layer, residuals, and normalization. The ACRM-NNE model includes: In the encoder, a loss function loss1 is designed to make the encoder equivalent to the functions of full response equalization and post-filtering; in the decoder, a loss function loss2 is designed to output an equalized signal without ISI; the sum of the two losses is used as the total loss function L of the ACRM-NNE model. The linear layer converts the input signal into a value vector, which is then processed by N1 encoders. The encoders use multi-head attention layers to capture dependencies between different locations. Residuals are then introduced into the normalization layer to prevent degradation during network training. A non-linear activation function is introduced into the feedforward network, enabling the ACRM-NNE model to handle non-linear problems and improve its fitting ability. After passing through the encoders, the signal's loss1 is calculated, and the signal is then input into N2 decoders, which also use multi-head attention layers for signal processing. The decoder outputs are processed through a linear layer to extract features, yielding the overall model output y. * The calculated loss2 is added to loss1 to obtain the total loss. Test module: Test the performance of the ACRM-NNE training model through the test set and obtain the bit error rate. If the bit error rate is lower than the set value, return to execute the 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 using an adaptive channel response matching neural network equalizer.

Images