Nonlinear impairment compensation system and method based on neural network equalizer

Abstract

The application relates to a nonlinear impairment compensation system and method based on a neural network equalizer, which comprises a front DSP processing module, a preprocessing module, an equalizer input signal sequence obtaining module, a grouping interception module, a signal coding module, a signal corresponding module, a disorder module, a segmentation module, a multi-symbol output neural network equalizer module, and a signal decoding module. Compared with the prior art, the application has the advantages of compensating for nonlinear loss, reducing the calculation complexity, and being highly universal.

Description

Technical Field

[0001] This invention relates to nonlinear compensation in the field of optical fiber communication, and in particular to a nonlinear damage compensation system and method based on a neural network equalizer. Background Technology

[0002] With the increasing demand for bandwidth and the continuous development of deep learning, machine learning-based equalization techniques have become a research hotspot in the field of optical communication. Artificial Neural Networks (ANNs) show promise as a feasible method to solve the nonlinear impairment problem in optical fiber communication systems.

[0003] Improving the structure of neural network equalizers can further enhance their performance in compensating for nonlinear impairments. Adding hidden layers enhances the nonlinear characteristics of neural networks, convolutional layers improve feature extraction capabilities, and recurrent structures endow neural networks with memory. Both Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks contribute to nonlinear equalization, but LSTMs outperform ordinary RNNs in long sequence equalization. Gateway Recurrent Units (GRUs) combine the forgetting and selective memory operations of LSTMs, achieving performance similar to LSTMs with approximately a 25% reduction in complexity.

[0004] However, existing neural network equalizers based on the aforementioned networks only output single symbols and mostly employ the sliding window method, where these windows are the inputs to the neural network model. Since adjacent windows contain a large amount of identical data, computing these similar inputs increases complexity.

[0005] Therefore, how to achieve nonlinear loss compensation in optical fiber communication through neural network equalizers and reduce computational complexity has become a technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a nonlinear damage compensation system and method based on a neural network equalizer.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A nonlinear damage compensation system based on a neural network equalizer, the system comprising:

[0009] The front-end DSP processing module is used to process the fiber optic signal R received at the receiving end. xAfter preprocessing, the equalizer input signal sequence y is obtained;

[0010] The grouping and truncation module is used to group the equalizer input signal sequence y into a set G of equalizer input signal sequence segments, where the window sliding distance during each truncation is the number of output symbols n. symbols ;

[0011] The signal encoding module is used to encode the transmitted signal sequence x. Each pair of transmitted signal points corresponds to an encoding tag, resulting in the encoded transmitted signal sequence x. c ;

[0012] The signal correspondence module is used to encode the transmitted signal sequence x. c Each signal point in the set is mapped one-to-one with the set G of equalizer input signal sequence segments to obtain the set P.

[0013] The shuffling module is used to shuffle set P to obtain a shuffled set P. shuffled ;

[0014] The partitioning module is used to partition the disordered set P. shuffled The data is segmented to obtain the training set P. train With test set P test ;

[0015] Multi-symbol output neural network equalizer module, used to use training set P train Perform training and backpropagate the loss function using the test set P. test Perform the test and use the test set P test The second stage of training is performed and the loss function is backpropagated to obtain better neural network parameters, and the input signal sequence y of the equalizer is equalized.

[0016] The signal decoding module is used to decode the output value of the equalizer and restore it to a bit sequence to compensate for nonlinear damage.

[0017] Furthermore, the set G of equalizer input signal sequence segments is represented as:

[0018]

[0019]

[0020]

[0021] Among them, g k n represents the sequence segment of the input signal to each equalizer. windows For each group of signal sequence segments g k The length of , where n is the length of the signal sequence x transmitted at the end.

[0022] Furthermore, the multi-symbol output neural network equalizer module includes:

[0023] The LSTM layer is used to receive training sequences and train the neural network.

[0024] The fully connected layer is used to receive the training results and obtain the output value of the multi-symbol output neural network equalizer module.

[0025] The loss function submodule is used to reflect the difference between the equalization output value and the actual reference value, and to measure the equalization effect.

[0026] Furthermore, the input of the multi-symbol output neural network equalizer module includes the signal sequence to be equalized and the reference signal sequence, and the output includes the equalization result signal sequence and the loss value.

[0027] Furthermore, the loss function is expressed as:

[0028]

[0029] Where p represents the true distribution of each symbol in the information source, and q represents the non-true distribution obtained by classification decision after training the neural network.

[0030] A method for using the aforementioned nonlinear damage compensation system based on a neural network equalizer, the method comprising the following steps:

[0031] 1) The signal sequence is sent and transmitted through the channel. After passing through the front-end DSP processing module, the equalizer input signal sequence y is obtained.

[0032] 2) Use the grouping and truncation module to group the equalizer input signal sequence y to obtain the equalizer input signal sequence segment set G;

[0033] 3) Use the signal encoding module to encode every two signal points in the transmitting signal sequence x into a tag, thus obtaining the encoded transmitting signal sequence x. c ;

[0034] 4) Use the signal mapping module to encode the transmitted signal sequence x. c Each signal point in the set is mapped one-to-one with the set G of equalizer input signal sequence segments to obtain the set P.

[0035] 5) Use the shuffling module to shuffle set P, resulting in a shuffled set P. shuffled ;

[0036] 6) Use the partitioning module to process the disordered set P shuffled The data is segmented to obtain the training set P. train With test set P test ;

[0037] 7) Using training set P train The multi-symbol output neural network equalizer module is trained, the equalizer output value is compared with the actual reference value and the loss function is calculated, and the loss function is fed back to update the neural network parameters.

[0038] 8) Use test set P test Test the multi-symbol output neural network equalizer module, compare the equalizer output value with the actual reference value and calculate the loss function, compare the loss function value with the value in step 7), until the bit error rate converges to the ideal value;

[0039] 9) Use test set P test The multi-symbol output neural network equalizer module is trained using the training data in the second stage. The equalizer output value is compared with the actual reference value and the loss function is calculated. The loss function is then fed back to update the neural network parameters, and finally the trained neural network is obtained.

[0040] 10) Using the trained neural network, the actual transmitted signal is processed by the front-end DSP processing module and then input into the multi-symbol output neural network equalizer module to obtain the equalization result. The signal is then decoded by the signal decoding module to restore the bit sequence and compensate for nonlinear damage.

[0041] Further, step 3) specifically involves: the signal encoding module encoding each pair of transmitting signal points (x) in the transmitting signal sequence x. i x i+1 ) corresponds to an encoded label x ck This yields an encoded version of the transmitted signal sequence x, with a length equal to half the length of the transmitted signal sequence x. c .

[0042] Furthermore, the decoding in step 10) specifically involves the signal decoding module demapping each signal point in the equalized signal sequence into a combination of two symbols, and restoring the demapped signal sequence into a bit sequence.

[0043] An electronic device includes a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the program to implement the method described above.

[0044] A computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described above.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] I. This invention achieves effective compensation for nonlinear impairments in optical fiber communication through an improved multi-symbol output neural network equalizer, overcomes the problem of inter-symbol interference, effectively solves the problem of high computational complexity of single-symbol output neural networks, and reduces computational complexity.

[0047] Second, the data preprocessing operations used in this invention, such as grouping and signal alignment, can also be used for data preprocessing of other commonly used equalization algorithms, such as the Constant Modulus Algorithm (CMA), the Cascaded Multi-Modulus Algorithm (CMMA), and the Volterra nonlinear equalization algorithm. In terms of application scenarios, this invention is applicable to nonlinear equalization in various scenarios such as long-distance single-carrier coherent optical transmission, short-distance intensity modulation direct detection (IM-DD) systems, and fiber-to-wireless (ROF) systems, and has strong versatility. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0049] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0050] Figure 3 Schematic diagram of the "data preprocessing" section of an improved multi-symbol output neural network equalizer;

[0051] Figure 4 This is a schematic diagram of the LSTM network cell structure;

[0052] Figure 5 This is a schematic diagram of a fully connected layer network structure.

[0053] The labels in the diagram indicate:

[0054] 1. Pre-processing DSP module; 2. Packet truncation module; 3. Signal encoding module; 4. Signal mapping module; 5. Out-of-order processing module; 6. Segmentation module; 7. Multi-symbol output neural network equalizer module; 8. Signal decoding module; 9. Transmitter signal sequence x; 10. Channel and pre-processing DSP operation; 11. Equalizer input signal sequence y; 12. Number of packet truncation windows n windows 13. Set G of equalizer input signal sequence segments; 14. Encoding operation; 15. Encoded transmitter signal sequence x. c 16. Signal alignment operation; 17. Set P; 18. Disordering operation; 19. Disordered set P shuffled 20. Segmentation operation; 21. Training set P train 22. Test set P test23. The previous cell output information h t-1 24. Current cell input information x t 25. Cell state c at the previous moment t-1 26. Sigmoid activation function; 27. Forget gate; 28. Sigmoid activation function; 29. ​​Input gate; 30. Tanh activation function; 31. Candidate cell state. 32. Current cell state c t 33. Sigmoid activation function; 34. Output gate; 35. Tanh activation function; 36. Output h at the current time step. t 37. Input layer, 38. Hidden layer, 39. Output layer. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0056] like Figure 1 As shown, a nonlinear damage compensation system based on a neural network equalizer is disclosed, the system comprising:

[0057] The front-end DSP processing module 1 is used to process the fiber optic signal R received at the receiving end. x After preprocessing, the equalizer input signal sequence y11 is obtained;

[0058] Grouping and truncation module 2 is used to group the equalizer input signal sequence y 11 to obtain a set of equalizer input signal sequence segments G 13, where the window sliding distance during each truncation is the number of output symbols n. symbols ;

[0059] The set of equalizer input signal sequence segments G13 is represented as follows:

[0060]

[0061]

[0062]

[0063] Among them, g k n represents the sequence segment of the input signal to each equalizer. windows For each group of signal sequence segments g k The length of , where n is the length of the transmitted signal sequence x 9;

[0064] Signal encoding module 3 is used to encode the transmitting signal sequence x9, with each pair of transmitting signal points corresponding to an encoding tag, to obtain the encoded transmitting signal sequence x9. c 15;

[0065] Signal correspondence module 4 is used to encode the transmitted signal sequence x. c Each signal point in 15 is matched one-to-one with the set of equalizer input signal sequence segments G 13 to obtain set P 17;

[0066] The shuffling module 5 is used to shuffle set P17 to obtain a shuffled set P. shuffled 19;

[0067] Segmentation module 6 is used to partition the disordered set P shuffled 19 is segmented to obtain the training set P. train 21 and test set P test twenty two;

[0068] Multi-symbol output neural network equalizer module 7, used to use training set P train 21. Train the program and backpropagate the loss function using the test set P. test 22. Conduct tests using test set P. test 22. Perform the second stage of training and backpropagate the loss function to obtain better neural network parameters, and perform equalization on the input signal sequence y11 of the equalizer;

[0069] The multi-symbol output neural network equalizer module 7 includes:

[0070] The LSTM layer is used to receive training sequences and train the neural network.

[0071] The fully connected layer is used to receive the training results and obtain the output value of the multi-symbol output neural network equalizer module 7;

[0072] The loss function submodule is used to reflect the difference between the equalization output value and the actual reference value, and to measure the equalization effect;

[0073] The input of the multi-symbol output neural network equalizer module 7 includes the signal sequence to be equalized and the reference signal sequence, and the output includes the equalization result signal sequence and the loss value.

[0074] The loss function is expressed as:

[0075]

[0076] Where p represents the true distribution of each symbol in the source, and q represents the non-true distribution obtained by the classification decision after the neural network is trained. By minimizing the cross-entropy loss, the similarity between the neural network's decision symbols and the symbols sent by the source can be improved.

[0077] Signal decoding module 8 is used to decode the output value of the equalizer and restore it to a bit sequence to compensate for nonlinear damage.

[0078] like Figure 2 As shown, a method for using the aforementioned nonlinear damage compensation system based on a neural network equalizer includes the following steps:

[0079] 1) The signal sequence is sent and transmitted through the channel. After passing through the front-end DSP processing module 1, the equalizer input signal sequence y11 is obtained.

[0080] 2) Use the grouping and truncation module 2 to group the equalizer input signal sequence y 11 to obtain the equalizer input signal sequence segment set G 13;

[0081] 3) Using signal encoding module 3, encode every two signal points in the transmitting signal sequence x9 into a tag, obtaining the encoded transmitting signal sequence x9. c 15;

[0082] 4) Use signal mapping module 4 to encode the transmitted signal sequence x. c Each signal point in 15 is matched one-to-one with the set of equalizer input signal sequence segments G 13 to obtain set P17;

[0083] 5) Use the shuffling module 5 to shuffle set P17, resulting in a shuffled set P. shuffled 19;

[0084] 6) Use the partitioning module 6 to partition the disordered set P shuffled 19 is segmented to obtain the training set P. train 21 and test set P test twenty two;

[0085] 7) Using training set P train 21. Train the multi-symbol output neural network equalizer module 7, compare the equalizer output value with the actual reference value and calculate the loss function, and feed it back to update the neural network parameters;

[0086] 8) Use test set P test 22 Test the multi-symbol output neural network equalizer module 7, compare the equalizer output value with the actual reference value and calculate the loss function, compare the loss function value with that in step 7), until the bit error rate converges to the ideal value;

[0087] 9) Use test set P test 22 is used as training data to train the multi-symbol output neural network equalizer module 7 in the second stage. The equalizer output value is compared with the actual reference value and the loss function is calculated. The loss function is then fed back to update the neural network parameters, and finally the trained neural network is obtained.

[0088] 10) Using the trained neural network, the actual transmitted signal is processed by the front-end DSP processing module 1 and then input into the multi-symbol output neural network equalizer module 7 to obtain the equalization result. The signal is then decoded by the signal decoding module 8 to restore the bit sequence and compensate for nonlinear damage.

[0089] Steps 1) to 6) are data preprocessing, steps 7) to 9) are training, and step 10) is the use of the multi-symbol output neural network equalizer module 7. The number of output symbols in the multi-symbol output neural network equalizer module 7 can be reasonably selected according to specific circumstances. When cross-entropy is used as the loss function for equalization and demapping, a linear increase in the number of output symbols leads to an exponential increase in the output layer length. Furthermore, the closer the output symbol is to the edge of the sequence, the fewer its corresponding effective taps. Therefore, it is necessary to reasonably select the number of output symbols. This embodiment uses a multi-symbol output neural network equalizer module 7 with two output symbols as an example. The type of neural network used in the multi-symbol output neural network equalizer module 7 can be selected from various options depending on the specific circumstances, such as Convolutional Neural Network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) network, etc. This embodiment uses LSTM as an example, and its specific working principle is as follows:

[0090] Data preprocessing, specifically as follows:

[0091] like Figure 3 As shown, the receiving end receives the fiber optic signal R. x After passing through the pre-DSP processing module 1, channel and pre-DSP operations are performed to obtain the equalizer input signal sequence y11. The group borrowing module 2 then sets the group truncation window number n. windows 12. Group the equalizer input signal sequence y 11, that is, after completing the synchronization operation, group each pair of transmitter signal points (x) i x i+1 ) and the signal point pairs (y11) in the equalizer input signal sequence y11 i y i+1Corresponding to each other, and then extending to the left and right respectively at the signal point pair (n windows (2-1) signal points are used to obtain a signal of length n. windows equalizer input sequence fragment g k The above operation is performed sequentially on each pair of signal points. The window sliding distance during each capture is equal to the number of output symbols, i.e., n. symbols =2, which yields the set of equalizer input signal sequence segments C13:

[0092]

[0093]

[0094]

[0095] The transmitting signal sequence x9 has a length of n. The transmitting signal sequence x9 is encoded using encoding operation 14. Each pair of transmitting signal points (x9...) i x i+1 ) corresponds to a coded label x ck (For PAM4 signals, two symbols correspond to 16 combinations, i.e., 00 corresponds to 0, 01 corresponds to 1, ..., 33 corresponds to 15), resulting in the encoded transmitter signal sequence x. c 15. Its length is n / 2. Perform signal alignment operation 16, and align the encoded signal points x. ck The set of equalizer input signal sequence fragments g after grouping and truncation k By performing a one-to-one correspondence, we obtain set P17. In the scrambling module 5, we shuffle each element in set P17 using scrambling operation 18 to obtain a scrambled set P. shuffled 19.

[0096] The partitioning module has 6 pairs of disordered sets P. shuffled 19. Perform segmentation operation 20 to obtain training set P train 21 and test set P test twenty two.

[0097] The training is as follows:

[0098] The loss function used in the training phase to calculate the difference between the equalizer output value and the reference value can be selected according to specific circumstances. When LSTM is used for both equalization and demapping, the task can be transformed into a multi-class classification problem, and cross-entropy can be chosen as the loss function; when LSTM is only used for equalization, the task can be transformed into a fitting problem, and the mean squared error function can be chosen as the loss function. This embodiment uses the cross-entropy function as an example.

[0099] like Figure 4As shown, the training sequence is passed through an LSTM layer with 101 hidden neurons. This LSTM layer includes a forget gate 27, an input gate 29, and an output gate 34. The previous cell output information h... t-1 23 and current cell input information x t 24 In the LSTM network, the cell state c from the previous time step is first determined by the sigmoid activation function 28 through the forget gate 27. t-1 The forgetting level is 25. The output f of forgetting gate 27. t for:

[0100] f t =σ(W f ·[h t-1 x t ]+b f )

[0101] Where σ is the sigmoid activation function 28, W f b is the weight of the forget gate. f Offset the forget gate.

[0102] Input gate 29 determines the new information i that needs to be added to the historical cell state via sigmoid function 28. t , through tan h Activation function 30 generates candidate cell states 31, and update the cell state to obtain the current cell state c. t 32. The formulas for each step are as follows:

[0103] i t =σ(W i ·[h t-1 x t ]+b i )

[0104]

[0105]

[0106] Output gate 34 determines the output h at the current moment. t 36, its formula is:

[0107] o t =σ(W o ·[h t-1 x t ]+b o )

[0108] h t =o t *tanh(c t )

[0109] In an LSTM network, the weights, such as W... f W i W c W0, etc., and each bias b f b i b c b o These are all trainable variables.

[0110] The training set P obtained after data preprocessing train 21 and test set P test 22 is used for LSTM equalizer training, and the entire training process is divided into two stages.

[0111] In the first stage, the training set P is obtained. train 21 is used for training. The training set P is used for training. train The signal sequence fragment set in 21 is input into the neural network. After passing through the LSTM layer, the data is flattened and sent to the fully connected layer.

[0112] like Figure 5 As shown, the fully connected layer includes an input layer 37, a hidden layer 38, and an output layer 39. The hidden layer has 101 neurons, the hidden layer 38 has a length of 84, and the output layer 39 has a length of 16. The obtained output value... Compared with the reference value (training set P) train The transmitting end encoded signal x in 21 train The comparison is performed using the cross-entropy formula shown below to calculate the loss function:

[0113]

[0114] The calculated loss function is fed back into the neural network to update its parameters. Given a test period c, for the training set P train For all reference values ​​and equalizer input sequence pairs in step 21, repeat the previous training process. For every c repetitions, calculate the average loss value. Simultaneously using test set P test Repeat the above steps for all reference values ​​and input sequence pairs in step 22 to obtain the average loss value. (Only calculate the loss value, do not update the weights). Iterate continuously, comparing... and The decline, if In the decline If the performance stops decreasing and the gap widens, it indicates that the training is heading towards overfitting, and the performance in practical applications has reached convergence. Adjust the parameters of the neural network and observe the comparison. and The equalizer is trained without severe overfitting until the bit error rate performance (or loss value) roughly converges to the ideal position. At this point, the first stage of training is complete.

[0115] To fully utilize the data in the training sequence for the second stage of training, the neural network parameters, which were determined in the first stage, are used with the test set P. test 22 is fed as training data into the multi-symbol output neural network equalizer module 7 for training. Finally, the trained neural network parameters are obtained.

[0116] The usage of the multi-symbol output neural network equalizer module 7 is as follows:

[0117] The signal sequence to be equalized is preprocessed to obtain a set of input signal sequence segments for the equalizer, which is then input into a trained neural network to obtain the output value of the equalizer. The output value is demapped by the signal decoding module 8, that is, each signal point in the equalized signal sequence is demapped into a combination of two symbols (e.g., for the PAM4 signal, 0 is demapped to 00, 1 is demapped to 01, ..., 15 is demapped to 33), and the demapped signal sequence is restored to the bit sequence.

[0118] This multi-symbol output neural network can achieve certain gains compared to traditional nonlinear compensation algorithms, effectively reduce its computational complexity compared to single-symbol output neural networks, and avoid overfitting problems during training. It also has a good compensation effect on nonlinear damage in optical fiber communication systems.

[0119] An electronic device includes a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the program to implement the method described above.

[0120] A computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described above.

[0121] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A nonlinear damage compensation system based on a neural network equalizer, characterized in that, The system includes: The front-end DSP processing module (1) is used to process the optical fiber signal received at the receiving end. The equalizer input signal sequence is obtained after preprocessing. (11); The grouping and truncation module (2) is used to extract the input signal sequence of the equalizer. (11) Group the signals to obtain the set of equalizer input signal sequence segments G(13), where the window sliding distance during each truncation is equal to the number of output symbols. ; The signal encoding module (3) is used to encode the signal sequence from the transmitting end. (9) Encode the signal by assigning a coding tag to every two transmitting signal points to obtain the encoded transmitting signal sequence. (15); The signal correspondence module (4) is used to encode the signal sequence of the transmitting end. Each signal point in (15) is mapped one-to-one with the set of equalizer input signal sequence segments G(13) to obtain the set (17); The disordered module (5) is used to sort the set P (17) Shuffle the set to obtain a disordered set. (19); The segmentation module (6) is used to segment the disordered set. (19) Perform segmentation to obtain the training set. (21) with the test set (twenty two); Multi-symbol output neural network equalizer module (7), used to use the training set (21) Perform training and backpropagate the loss function using the test set. (22) Conduct tests and use the test set. (22) Perform the second stage of training and backpropagate the loss function to obtain better neural network parameters, and adjust the input signal sequence of the equalizer. (11) To achieve equilibrium; The signal decoding module (8) is used to decode the output value of the equalizer and restore it to a bit sequence to compensate for nonlinear damage; The set of equalizer input signal sequence segments G(13) is represented as: in, This represents a segment of the input signal sequence for each equalizer group. For each group of signal sequence segments Length, n For the transmitting end signal sequence (9) length; The loss function is expressed as: in, p This represents the true distribution of symbols in the information source. q This represents the non-true distribution derived from the classification decisions made after training the neural network. x i This is the signal point at the transmitting end.

2. The nonlinear damage compensation system based on a neural network equalizer according to claim 1, characterized in that, The multi-symbol output neural network equalizer module (7) includes: The LSTM layer is used to receive training sequences and train the neural network. The fully connected layer is used to receive the training results and obtain the output value of the multi-symbol output neural network equalizer module (7); The loss function submodule is used to reflect the difference between the equalization output value and the actual reference value, and to measure the equalization effect.

3. The nonlinear damage compensation system based on a neural network equalizer according to claim 2, characterized in that, The input of the multi-symbol output neural network equalizer module (7) includes the signal sequence to be equalized and the reference signal sequence, and the output includes the equalization result signal sequence and the loss value.

4. A method for using the nonlinear damage compensation system based on a neural network equalizer as described in claim 1, characterized in that, The method includes the following steps: 1) The signal sequence is sent and transmitted through the channel. After passing through the front-end DSP processing module (1), the equalizer input signal sequence is obtained. (11); 2) Use the grouping and truncation module (2) to process the input signal sequence of the equalizer. (11) Group the signals to obtain the set of equalizer input signal sequence segments G (13). 3) Use the signal encoding module (3) to encode the signal sequence at the transmitting end. In (9), every two signal points are encoded into a tag, resulting in the encoded transmitter signal sequence. (15); 4) Use the signal correspondence module (4) to encode the transmitter signal sequence. Each signal point in (15) is mapped one-to-one with the set of equalizer input signal sequence segments G(13) to obtain the set (17); 5) Use the disordered module (5) on the set (17) Shuffle the set to obtain a disordered set. (19); 6) Use the partitioning module (6) on the disordered set (19) Perform segmentation to obtain the training set. (21) with the test set (twenty two); 7) Using the training set (21) Train the multi-symbol output neural network equalizer module (7), compare the equalizer output value with the actual reference value and calculate the loss function, and feed it back to update the neural network parameters; 8) Use the test set (22) Test the multi-symbol output neural network equalizer module (7), compare the equalizer output value with the actual reference value and calculate the loss function, compare the loss function value with that in step 7), until the bit error rate converges to the ideal value; 9) Use the test set (22) The multi-symbol output neural network equalizer module (7) is trained as training data in the second stage. The equalizer output value is compared with the actual reference value and the loss function is calculated. The loss function is then fed back to update the neural network parameters, and finally the trained neural network is obtained. 10) Using the trained neural network, the actual transmitted signal is processed by the front-end DSP processing module (1) and then input into the multi-symbol output neural network equalizer module (7) to obtain the equalization result. The signal is then decoded by the signal decoding module (8) to restore the bit sequence and compensate for nonlinear damage.

5. The method according to claim 4, characterized in that, Step 3) specifically involves the signal encoding module (3) encoding the signal sequence from the transmitting end. (9) Each pair of transmitting end signal points Corresponding to an encoded tag The length of the transmitted signal sequence is obtained. (9) is half the length of the encoded transmitter signal sequence. x c (15).

6. The method according to claim 4, characterized in that, The decoding in step 10) specifically involves the signal decoding module (8) demapping each signal point in the equalized signal sequence into a combination of two symbols, and restoring the demapped signal sequence into a bit sequence.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 4 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 4 to 6.

Images