Signal processing method and device, equipment, storage medium and computer program product

Abstract

The invention relates to the technical field of network communication, and discloses a signal processing method and device, equipment, a storage medium and a computer program product, and the method comprises the steps: obtaining an equalized signal with nonlinear damage, carrying out the normalization processing of the equalized signal, and generating a target input signal; the target input signal is input into the improved multi-task neural network for nonlinear compensation, a compensated signal is obtained, the multi-task neural network determines the number of neurons in each layer through a step-by-step traversal method and Bayesian optimization, and optimization training is carried out according to an optimization target constructed by calculation complexity. Through a weight sharing mechanism of multi-task learning, joint processing of multiple symbols is realized, redundant calculation operation in a traditional single-task neural network is reduced, the calculation complexity is remarkably reduced while the compensation precision is ensured, and the real-time processing requirement in a high-data-rate transmission scene is met.

Description

Technical Field

[0001] The present application relates to the field of network communication technology, and in particular to a signal processing method, apparatus, device, storage medium, and computer program product. Background Art

[0002] With the improvement of computing power and the development of artificial intelligence technology, machine learning, especially deep learning, has the ability to extract nonlinear features from complex data. Therefore, it is widely used in nonlinearity compensation (NLC) in optical communications. However, existing single-task neural network models require a large amount of sample data and high computing power to process high-speed, large-scale data streams in nonlinear compensation. They cannot reduce computational complexity while ensuring compensation accuracy, resulting in high cost and energy consumption in practical applications. Summary of the Invention

[0003] The main purpose of this application is to provide a signal processing method, device, equipment, storage medium and computer program product, aiming to solve the technical problems that the existing single-task neural network model requires a large amount of sample data and high computing power in nonlinear compensation, and faces high cost and energy consumption problems.

[0004] To achieve the above objectives, the present application proposes a signal processing method, which includes:

[0005] Acquiring an equalized signal with nonlinear impairment, and performing normalization processing on the equalized signal to generate a target input signal;

[0006] The target input signal is input into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal method and Bayesian optimization, and performs optimization training according to the optimization target constructed according to computational complexity.

[0007] Optionally, before the step of inputting the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, the method further includes:

[0008] Obtaining an initial multi-task neural network, wherein the initial multi-task neural network includes an input layer, an intermediate layer, and an output layer;

[0009] Determine the first number of neurons in the optimal configuration of the output layer in the initial multi-task neural network by using a step-by-step traversal method starting from the number of neurons that meets the minimum task requirements and gradually increasing the configuration;

[0010] Determine the number of second neurons in the input layer and the number of third neurons in the middle layer of the initial multi-task neural network by constructing a Gaussian process regression model based on Bayesian optimization;

[0011] Optimizing the network structure of the initial multi-task neural network according to the first number of neurons, the second number of neurons, and the third number of neurons to obtain a target multi-task neural network;

[0012] An optimization target is formulated according to the signal recovery loss and the computational complexity loss, and the target multi-task neural network is trained based on the optimization target to obtain an improved multi-task neural network.

[0013] Optionally, the step of determining the number of second neurons in the input layer and the number of third neurons in the intermediate layer in the initial multi-task neural network by constructing a Gaussian process regression model based on Bayesian optimization includes:

[0014] Generating initial hyperparameter combinations by random sampling, and training and evaluating the performance of the multi-task neural network corresponding to each of the initial hyperparameter combinations to obtain initial function values, wherein the initial hyperparameter combinations include different combinations of the number of neurons in the input layer and the intermediate layer;

[0015] Constructing a Gaussian process regression model based on the initial hyperparameter combination and the corresponding initial function value;

[0016] Selecting currently unevaluated hyperparameter combinations in sequence according to a preset acquisition function, evaluating the objective function values ​​of the selected hyperparameter combinations, and updating the Gaussian process regression model until a preset iteration termination condition is met;

[0017] After the Gaussian process regression model is updated, the number of neurons in the input layer and the middle layer that minimizes the objective function value are selected from all evaluated hyperparameter combinations as the second number of neurons in the input layer and the third number of neurons in the middle layer.

[0018] Optionally, the step of formulating an optimization target according to the signal recovery loss and the computational complexity loss, and training the target multi-task neural network based on the optimization target to obtain an improved multi-task neural network includes:

[0019] The computational complexity is obtained by using the number of multiplication operations generated by the connection of neurons in each layer of the target multi-task neural network as a quantitative indicator of computational complexity loss;

[0020] The mean square error between the output signal of the target multi-task neural network and the original undamaged signal is used as the signal recovery loss;

[0021] constructing an optimization target according to the computational complexity, the signal recovery loss, and a hyperparameter between the signal recovery accuracy and the computational complexity;

[0022] The target multi-task neural network is trained using the optimization objective to obtain an improved multi-task neural network.

[0023] Optionally, the step of obtaining an equalized signal with nonlinear impairment and performing normalization processing on the equalized signal to generate a target input signal includes:

[0024] Converting the optical signal transmitted in the link into an electrical signal through a coherent receiver, and sampling the electrical signal based on a digital sampling oscilloscope to obtain a target signal;

[0025] Performing digital signal processing on the target signal to generate an equalized signal containing only nonlinear impairments, wherein the digital signal processing includes dispersion compensation, frequency offset estimation, and adaptive equalization correction.

[0026] Optionally, after the step of inputting the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, the method further includes:

[0027] Passing the compensated signal to a carrier phase recovery module to extract phase information;

[0028] The compensated signal is demodulated into digital data for output through a decision module according to the phase information.

[0029] In addition, to achieve the above-mentioned purpose, the present application also proposes a signal processing device, which includes:

[0030] A signal preprocessing module, configured to obtain an equalized signal with nonlinear impairments and perform normalization processing on the equalized signal to generate a target input signal;

[0031] A signal compensation module is used to input the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training based on the optimization target constructed according to computational complexity.

[0032] In addition, to achieve the above-mentioned purpose, the present application also proposes a signal processing device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the signal processing method described above.

[0033] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the steps of the signal processing method described above are implemented.

[0034] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the signal processing method described above are implemented.

[0035] This application discloses obtaining a post-equalization signal with nonlinear damage, normalizing the post-equalization signal to generate a target input signal; inputting the target input signal into an improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training based on an optimization target constructed based on computational complexity. Through the weight sharing mechanism of multi-task learning, the joint processing of multiple symbols is achieved, reducing redundant computational operations in traditional single-task neural networks. While ensuring compensation accuracy, the computational complexity is significantly reduced, meeting the real-time processing requirements in high data rate transmission scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0038] Figure 1 This is a flow chart of the first embodiment of the signal processing method of the present application;

[0039] Figure 2 This is a schematic diagram of the network structure based on the single-task feedforward neural network balancing solution;

[0040] Figure 3 This is a schematic diagram of the network structure based on the multi-task feedforward neural network balancing solution;

[0041] Figure 4 This is a schematic diagram of the structure of a coherent optical communication system based on nonlinear compensation of a multi-task neural network in this application;

[0042] Figure 5This is a flow chart of a second embodiment of the signal processing method of the present application;

[0043] Figure 6 This is a schematic diagram of the module structure of the signal processing device according to an embodiment of the present application;

[0044] Figure 7 Schematic diagram of the device structure of the hardware operating environment involved in the signal processing method in the embodiment of the present application.

[0045] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0047] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0048] The main solution of the embodiment of the present application is: obtaining an equalized signal with nonlinear damage, and normalizing the equalized signal to generate a target input signal; inputting the target input signal into an improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training according to the optimization target constructed based on computational complexity.

[0049] With the rapid development of network applications such as generative artificial intelligence and short video streaming, global data traffic has shown explosive growth, especially placing higher demands on high-speed, low-latency optical communication systems. To meet these challenges, coherent optical communication technology has become one of the key technologies to meet the growing demand for network bandwidth. Current long-distance digital coherent systems have effectively solved the problem of linear channel distortion through advanced digital signal processing (DSP) technology and achieved remarkable results. However, with the increase in system speed and transmission distance, the impact of nonlinear effects in optical communication systems has become increasingly significant, becoming a major bottleneck for further improving system capacity.

[0050] In optical fiber transmission, nonlinear signal impairments are primarily caused by self-interaction, cross-interaction, and their interaction with dispersion within the optical fiber. Although scholars have proposed a variety of nonlinear compensation techniques over the past few decades, computational complexity (CC) remains a significant challenge for these technologies as system complexity increases. Traditional digital back propagation (DBP) technology is widely used for nonlinear compensation, but the amount of computation required increases significantly with increasing signal transmission distance, making it difficult to achieve efficient processing in real-time applications. Furthermore, despite the proposal of simplified algorithms such as filtered DBP and perturbation back propagation, these methods have not been able to significantly reduce computational complexity, and their efficiency and stability remain problematic when processing large-scale data.

[0051] With the improvement of computing power and the development of artificial intelligence technology, nonlinear compensation methods based on machine learning have gradually become an important direction for solving nonlinear effects. Machine learning, especially deep learning, has the ability to extract nonlinear features in complex data, and is therefore widely used in nonlinear compensation in optical communications. Through methods such as neural networks (NN), machine learning can directly model the nonlinear effects in optical fibers by learning from large amounts of transmission data, without relying on traditional physical models such as the nonlinear Schrödinger equation. The advantage of these methods is that they can provide more accurate results than traditional compensation technologies when nonlinear damage is more complex.

[0052] However, despite the strong potential of machine learning in nonlinear compensation, existing neural network models often require significant computational resources, especially when processing high-speed, large-scale data streams. Training deep neural networks requires vast amounts of sample data and high computational power, which makes their practical application costly and energy-intensive. Therefore, reducing the computational complexity of machine learning models while maintaining their efficiency and accuracy in nonlinear compensation remains a hot topic and a challenge in current research.

[0053] This application provides a neural network nonlinear compensation method based on multi-task learning (MTL), which aims to reduce the computational complexity of nonlinear compensation in coherent optical communication systems while maintaining the system's performance indicators. Using the method of the present invention, the nonlinear compensation of coherent optical communication systems can maintain bit error rate (Q factor) performance comparable to traditional solutions while significantly reducing complexity. In particular, under the system configuration after symbol number optimization, the computational complexity is significantly reduced, while achieving the needs of real-time processing and meeting the application requirements of high-speed and long-distance transmission scenarios.

[0054] It should be noted that the execution subject of this embodiment can be a computing service device with signal processing, network communication, and program execution functions, such as a computer, or an electronic device capable of implementing the above functions. This embodiment and the following embodiments are described below using a coherent optical communication system as an example.

[0055] Based on this, the embodiment of the present application provides a signal processing method, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the signal processing method of the present application.

[0056] In this embodiment, the signal processing method includes:

[0057] Step S10: obtaining an equalized signal with nonlinear impairment, and performing normalization processing on the equalized signal to generate a target input signal.

[0058] It should be noted that even after equalization of a signal with nonlinear impairments in a coherent optical communication system, nonlinear distortion remains after linear distortion compensation, such as dispersion compensation, frequency offset estimation and compensation, and adaptive equalization. This nonlinear distortion arises from distortion caused by nonlinear effects such as the Inker effect and four-wave mixing during fiber transmission. The target input signal is a normalized signal adapted to the input structure of a multi-task neural network.

[0059] It should be understood that the nonlinear compensation scheme of the present application focuses on processing the above-mentioned equalized signal, which is in sharp contrast to the traditional compensation method that directly targets the unequalized signal. It normalizes the equalized signal from which linear damage has been eliminated, focuses on nonlinear damage modeling, simplifies the network input dimension, and reduces computational complexity.

[0060] Understandably, to adapt to the input requirements of a neural network, signal characteristics such as amplitude, phase, or power can be scaled to a specific range for normalization. This ensures consistent distribution of input data and adapts to the network's batch processing mechanism. Of course, during neural network training, sequence data can also be normalized to eliminate the impact of signal dynamic range differences on model training, thereby improving the convergence speed and compensation accuracy of the neural network.

[0061] Furthermore, in order to ensure that the signal input to the multi-task neural network contains only nonlinear damage and removes interference factors such as linear distortion, the step S10 may include:

[0062] The optical signal transmitted in the link is converted into an electrical signal through a coherent receiver, and the electrical signal is sampled based on a digital sampling oscilloscope to obtain a target signal; the target signal is digitally processed to generate an equalized signal containing only nonlinear damage, wherein the digital signal processing includes dispersion compensation, frequency offset estimation and adaptive equalization correction.

[0063] It's important to note that at the receiving end, the optical signal is converted into an electrical signal by a coherent receiver. A digital sampling oscilloscope samples the electrical signal at a specific sampling frequency and converts the resulting analog signal into a digital signal for subsequent digital processing and analysis. The target signal is the digital signal obtained after the coherent receiver converts the optical signal into an electrical signal and samples it on the digital sampling oscilloscope. This signal contains the information carried by the optical signal during link transmission, as well as any impairments introduced. In optical communication systems, digital signal processing is used to compensate for any impairments incurred during optical signal transmission and restore the original information signal.

[0064] It can be understood that dispersion compensation is the process of processing signals in digital signal processing to offset pulse broadening caused by dispersion and restore the signal's original shape. Frequency offset estimation analyzes the received signal to estimate the magnitude of the frequency offset, which is then compensated in subsequent processing to ensure correct signal demodulation. Adaptive equalization is used to compensate for linear impairments such as polarization mode dispersion (PMD) incurred by optical signals during transmission. By continuously adjusting the equalizer coefficients, it adaptively tracks and compensates for signal changes, keeping the output signal as close to the original as possible. After the equalized signal undergoes digital signal processing steps such as dispersion compensation, frequency offset estimation, and adaptive equalization, most linear impairments have been eliminated, leaving only nonlinear impairments, ready for subsequent nonlinear compensation.

[0065] In step S20, the target input signal is input into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training according to the optimization target constructed based on computational complexity.

[0066] It should be noted that in the field of optical communications, multi-task neural networks can process nonlinear compensation tasks of multiple symbols in parallel, and share some network parameters between different tasks through the weight sharing mechanism, thereby improving processing efficiency and the generalization ability of the model. The step-by-step traversal method is used in this application to determine the number of neurons in the output layer of the neural network. By traversing and gradually changing the number of neurons in each layer, different combinations of the number of neurons are tried and evaluated to find the optimal number of neurons under specific performance indicators. Bayesian optimization predicts the impact of hyperparameters on the objective function by constructing a Gaussian process regression model, and dynamically balances between exploration and development to approach the global optimal solution faster.

[0067] It should be understood that computational complexity refers to the measurement of the computational resources (such as time, memory, etc.) required during the execution of an algorithm or model. In neural networks, computational complexity is usually related to factors such as the number of layers in the network, the number of neurons in each layer, and the connection method. The optimization objective is a function constructed based on specific application requirements and performance indicators. It is used to measure the performance of the neural network and ensure that the trained multi-task neural network significantly reduces computational complexity while ensuring compensation accuracy. The compensated signal is the signal obtained after nonlinear compensation processing of the improved multi-task neural network. Compared with the input target input signal, this signal reduces the influence of nonlinear damage and is closer to the original undamaged signal.

[0068] It is understandable that a convolutional neural network (CNN) can also be used instead of a feedforward neural network for nonlinear compensation. Its local perception and weight sharing characteristics can enhance feature extraction capabilities and improve the accuracy of modeling nonlinear effects. However, the introduction of CNN also increases computational complexity, making the equalization process more time-consuming. Therefore, in scenarios where computing resources are sufficient and higher compensation accuracy is required, a convolutional neural network can also be used for compensation.

[0069] It is understandable that multi-task neural networks also use grid search or random search methods to optimize the number of neurons in the input layer and the intermediate layer, realize automatic tuning of the model structure and parameters, and thus improve compensation performance and applicability.

[0070] In one example, reference Figure 2 and Figure 3 , Figure 2 This is a schematic diagram of the network structure based on the single-task feedforward neural network balancing solution. Figure 3 Schematic diagram of the network structure based on the multi-task feedforward neural network balancing scheme.

[0071] In traditional single-task neural network methods, the nonlinear compensation process is performed symbol by symbol, and each network calculation processes only one received symbol. This symbol-by-symbol processing method has high computational complexity and fails to fully utilize the correlation between adjacent symbols. Each received symbol is independently input into the neural network for processing, and the network calculates layer by layer through multiple layers of neurons to finally obtain the output result. Figure 2 As shown, a common two-layer feedforward neural network (FNN) has an input layer, an intermediate layer, and an output layer, and the corresponding number of neurons is n. i ′,n′ h and n′ o .in, Represents the received symbol, and X represents the input vector of the neural network: X i ,X q ,Y i ,Y q They represent the in-phase and quadrature signals of X polarization, and the in-phase and quadrature signals of Y polarization, respectively. represents the neurons in the middle layer, where n′ h Represents the number of neurons in the middle layer. s1 represents the output symbol, and Y represents the output vector of the neural network: Y = [s1].

[0072] The layers are connected by weights, W [1] Represents the connection weight between the input layer and the intermediate layer, W [2] Represents the connection weight between the intermediate layer and the output layer. The threshold of the intermediate layer is represented by b [1] , the threshold of the output layer is represented by b [2] , the calculation process of nonlinear compensation using neural network is as follows:

[0073] y=f [2] (W [2] f [1] (W [1] x+b [1] ) T +b [2] )

[0074] Where x represents the received symbol to be processed, f [1] (·) represents the activation function of the middle layer of the neural network, f [2] (·) represents the activation function of the output layer of the neural network, the superscript T represents the matrix transpose, and y represents the signal after the equalizer calculation. The number of multiplication operators required for the nonlinear compensation process is:

[0075] N mul =4n' h (n i'+n' o )

[0076] Since each symbol needs to be processed individually, the computational complexity increases linearly with the number of symbols. This makes the computational burden very heavy in systems that need to process a large number of symbols, and cannot effectively utilize the correlation between multiple symbols.

[0077] This application introduces a multi-task learning framework and proposes a new nonlinear compensation method that can process multiple symbols simultaneously. From the perspective of network structure, the multi-task neural network supports multi-task processing by increasing the number of neurons in the input and output layers, while appropriately increasing the number of neurons in the middle layer to meet more complex signal processing requirements. Figure 3 As shown, it is assumed that the number of neurons in the input layer, intermediate layer and output layer are n i , n h and n o , Represents the received symbol, where n i Represents the number of symbols in the input layer. X represents the input vector of the neural network: X i ,X q ,Y i ,Y q Represent the in-phase and quadrature signals of X polarization, and the in-phase and quadrature signals of Y polarization respectively. [1] Represents the connection weight between the input layer and the intermediate layer, W [2] Represents the connection weight between the intermediate layer and the output layer. represents the neurons in the middle layer, where n h Represents the number of neurons in the middle layer. Represents the symbol of the output, where n o Represents the number of symbols in the input layer. Y represents the output vector of the neural network:

[0078] The process of nonlinear compensation using a multi-task neural network is the same as above. The number of multiplication operators required to restore a single symbol is:

[0079] N mul =4n h (n i +n o ) / n o

[0080] Although the number of multi-task neurons increases compared to single-task neural networks, the multi-task learning structure allows the network to process multiple symbols simultaneously, and the network weights are shared between symbols. This allows the nonlinear compensation operators of adjacent symbols to be reused, reducing the computational complexity required for each symbol.

[0081] Through the weight-sharing mechanism of multi-task learning, the balancing tasks of multiple symbols are jointly processed, reducing the redundant computational operations required by traditional single-task neural networks to independently train and process each symbol. During symbol processing, multiple symbols share the feature extraction layer and some hidden layers of the network structure, significantly reducing computational complexity, the number of parameters, and memory usage, meeting the real-time processing requirements of high-data-rate transmission scenarios.

[0082] Furthermore, in order to ensure the ultimate reliability of data transmission, the entire signal processing method can be stably operated in an actual communication system, thereby improving the integrity and applicability of the system. After step S20, the method further includes:

[0083] The compensated signal is transmitted to a carrier phase recovery module to extract phase information; and the compensated signal is demodulated into digital data output through a decision module according to the phase information.

[0084] In one example, reference Figure 4 , Figure 4 This is a schematic diagram of the structure of a coherent optical communication system based on nonlinear compensation of a multi-task neural network in this application. The structure of the coherent optical communication system includes a transmitter, a transmission link and a receiver.

[0085] At the transmitter, an optical signal is generated by a laser. An arbitrary waveform generator (AWG) and a dual-polarization IQ modulator (DP IQ Mod) modulate the baseband electrical signal onto an optical carrier, generating a dual-polarization coherent optical signal. The modulated signal is transmitted over a long-distance link consisting of a standard single-mode fiber (SSMF) and an erbium-doped fiber amplifier (EDFA). N represents the number of spans in the link, each span consisting of an SSMF and an EDFA. At the receiver, the optical signal is converted to an electrical signal by a laser and an integrated coherent receiver (ICR). This signal is sampled by a digital sampling oscilloscope and fed into a digital signal processing (DSP) module. In the DSP module, the final signal is output by a digital storage oscilloscope (DSO). The signal first undergoes dispersion compensation to eliminate inter-symbol interference and frequency offset estimation to compensate for the frequency offset of the laser at the transmitting and receiving ends. Then, adaptive equalization is used to correct polarization mode distortion to generate an equalized signal containing only nonlinear damage. Then, carrier phase recovery and decision recovery are performed. i ,X q ,Y i ,Y qThey represent the in-phase and quadrature signals of X polarization, and the in-phase and quadrature signals of Y polarization, respectively.

[0086] The nonlinear compensation scheme proposed in this application focuses on processing the equalized signal, in stark contrast to traditional compensation methods that directly target unequalized signals. The specific compensation process includes: normalizing the equalized signal to ensure that the data format is suitable for neural network input; extracting the nonlinear characteristics of the signal using a multi-task neural network, where the input layer captures the signal characteristics, the middle layer models the nonlinear distortion pattern, and the output layer generates the compensated signal value; and improving processing efficiency through the weight sharing mechanism of multi-task learning.

[0087] After completing the nonlinear compensation, the signal is transmitted to the carrier phase recovery module to extract the phase information, and is finally demodulated into digital data output by the decision module.

[0088] In this embodiment, a post-equalization signal with nonlinear damage is obtained, and the post-equalization signal is normalized to generate a target input signal; the target input signal is input into an improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training based on the optimization target constructed based on computational complexity. Through the weight sharing mechanism of multi-task learning, the joint processing of multiple symbols is achieved, which reduces the redundant computational operations in the traditional single-task neural network. While ensuring the compensation accuracy, the computational complexity is significantly reduced to meet the real-time processing requirements in high data rate transmission scenarios. By introducing a multi-task neural network equalization scheme, it is different from the traditional neural network equalizer. Traditional neural networks can only process one symbol at a time, and the complexity increases linearly with the increase in the number of input symbols. On this basis, the present invention increases the number of neurons in the output layer so that it can process multiple symbols simultaneously.

[0089] Reference Figure 5 , Figure 5 This is a flow chart of the second embodiment of the signal processing method of the present application. Based on the above-mentioned first embodiment, the second embodiment of the signal processing method of the present application is proposed.

[0090] In the second embodiment, before step S20, the method further includes:

[0091] Step S201: obtaining an initial multi-task neural network, where the initial multi-task neural network includes an input layer, an intermediate layer, and an output layer.

[0092] It should be noted that the initial multi-task neural network is a basic neural network model in an untrained and unadjusted state. The input layer receives the target input signal. The intermediate layer, also known as the hidden layer, lies between the input and output layers and is a key component of the neural network for feature extraction and nonlinear transformation. Each layer contains multiple neurons, which are connected to neurons in the previous and next layers via connection weights. The output layer is where the neural network outputs the processed results. In nonlinear compensation tasks, the output layer outputs the compensated signal after the neural network has processed it.

[0093] Understandably, due to the large number of parameters that require optimization in multi-task neural networks, the number of neurons at different layers is determined differently. Because the output layer has a relatively small number of neurons, its search space is limited, and traversal can be used to find the optimal configuration. However, due to the larger parameter space of the input and intermediate layers, the search complexity is high, and stepwise traversal methods are computationally prohibitively expensive for such high-dimensional optimization problems. Therefore, Bayesian optimization is employed.

[0094] It should be understood that the multi-task neural network framework is naturally suitable for the parallel computing characteristics in the hardware architecture, such as GPU and FPGA acceleration, and processes multiple symbols simultaneously through a batch processing mechanism, effectively improving the processing speed and balancing efficiency. Compared with the traditional symbol-by-symbol processing method, the present invention uses a batch processing mechanism to reduce data loading and storage overhead, further shortens the calculation delay, and meets the needs of high-bandwidth, low-latency communication systems. The multi-task neural network framework can map multiple tasks to shared layers and task-specific layers to achieve the reuse of model parameters, reduce memory and computing resource consumption, and is suitable for the batch computing processing mechanism of hardware accelerators, thereby improving computing efficiency.

[0095] Step S202 , starting from the number of neurons that meets the minimum task requirements, gradually increasing the configuration by a step-by-step traversal method, to determine the first number of neurons in the optimal configuration of the output layer in the initial multi-task neural network.

[0096] It should be noted that in the neural network configuration determination scenario of this application, the step-by-step traversal method starts from a certain starting point in a certain order, tries different parameter configurations one by one, and evaluates each configuration to find the optimal configuration that meets specific conditions. For the number of neurons in the output layer, the minimum number of neurons required for the task is the minimum number of neurons that can enable the neural network to initially complete the task and produce meaningful output. The first number of neurons is the number of neurons that can enable the multi-task neural network to achieve the best performance when completing the nonlinear compensation task in the configuration of a large number of neurons in the output layer.

[0097] It's understandable that an appropriate number of output layer neurons enables multi-task neural networks to better perform nonlinear compensation tasks, improving both accuracy and efficiency. The stepwise traversal method systematically searches for the optimal configuration, avoiding the uncertainty associated with empirical or random selection. Some automated hyperparameter tuning methods may utilize complex algorithms, but these methods are often computationally expensive. The stepwise traversal method is relatively simple, intuitive, and easy to implement, offering advantages for small-scale problems or when resources are limited.

[0098] Specifically, based on the nature and requirements of the task, combined with theoretical analysis or experience, the number of neurons in the output layer that meets the minimum task requirements is determined as a starting point. For example, in some simple classification tasks, the number of neurons in the output layer may be related to the number of categories; in the nonlinear compensation task of this application, it may be related to the dimensionality of the signal or the number of parameters that need to be compensated. Then, starting with the minimum number of neurons required for the task, the number of neurons in the output layer is gradually increased according to a pre-set step size (for example, adding one neuron at a time). Each increase results in a new configuration of the number of neurons in the output layer. For each new configuration of the number of neurons in the output layer, the initial multi-task neural network is trained and tested using a set of pre-prepared validation data. Evaluation metrics can include the accuracy of nonlinear compensation (such as mean squared error, bit error rate, etc.) and computational complexity (such as training time, memory usage, etc.). The evaluation results under all configurations are compared, and the output layer neuron number that achieves the optimal evaluation metric is selected as the first neuron number for the optimal configuration. For example, performance metrics such as bit error rate and loss function value can be evaluated in combination with cross-validation. The traversal is stopped when the performance metrics converge or the benefits of increasing the number of neurons are insufficient, and the optimal balance between performance and computational complexity is determined.

[0099] Step S203 , determining the number of second neurons in the input layer and the number of third neurons in the middle layer of the initial multi-task neural network by constructing a Gaussian process regression model based on Bayesian optimization.

[0100] It should be noted that Bayesian optimization is used to find the optimal solution in a complex, high-dimensional parameter space. It models prior knowledge of the objective function and uses previously evaluated points to construct a probabilistic model. The model then predicts the next most promising point for evaluation, iterating and updating the model until the optimal solution is found. In neural networks, it can be used to determine the optimal values ​​of hyperparameters such as network structural parameters (such as the number of neurons). Gaussian process regression is a nonparametric probabilistic model used to model and predict functions. By estimating the mean and covariance function of a Gaussian process using known input-output data points, it is possible to predict the objective function value and its uncertainty for unknown input points. The second and third numbers of neurons are determined through Bayesian optimization to ensure that the initial multi-task neural network achieves optimal performance in the input layer and intermediate layer when performing nonlinear compensation.

[0101] Furthermore, in order to quickly converge to the optimal configuration of the number of neurons in the input layer and the intermediate layer under limited computing resources, the computational overhead of blind search is reduced. The step S203 may include:

[0102] Generating initial hyperparameter combinations by random sampling, and training and evaluating the performance of the multi-task neural network corresponding to each of the initial hyperparameter combinations to obtain initial function values, wherein the initial hyperparameter combinations include different combinations of the number of neurons in the input layer and the intermediate layer;

[0103] Constructing a Gaussian process regression model based on the initial hyperparameter combination and the corresponding initial function value;

[0104] Selecting currently unevaluated hyperparameter combinations in sequence according to a preset acquisition function, evaluating the objective function values ​​of the selected hyperparameter combinations, and updating the Gaussian process regression model until a preset iteration termination condition is met;

[0105] After the Gaussian process regression model is updated, the number of neurons in the input layer and the middle layer that minimizes the objective function value are selected from all evaluated hyperparameter combinations as the second number of neurons in the input layer and the third number of neurons in the middle layer.

[0106] It should be noted that random sampling involves randomly selecting some combinations from the parameter space consisting of all possible combinations of the number of neurons in the input and intermediate layers as initial hyperparameter combinations. The initial hyperparameter combination is a set of hyperparameters obtained through random sampling, including different combinations of the number of neurons in the input and intermediate layers. Performance evaluation can be performed using metrics such as mean squared error and bit error rate, measuring the performance of the network under different parameter combinations. The initial function value is a quantitative performance value obtained after training and performance evaluation of the multi-task neural network corresponding to each initial hyperparameter combination, which is used to measure the performance of the neural network.

[0107] In addition, it should be noted that the acquisition function is a function used in Bayesian optimization to determine the next hyperparameter combination to be evaluated. It comprehensively considers the predicted value and uncertainty of the objective function value to balance exploration (finding new potential areas) and utilization (utilizing known better areas). The objective function value is used to evaluate the performance of the hyperparameter combination to be evaluated on the multi-task neural network. The preset iteration termination condition refers to the pre-set condition for stopping the iterative process, which may include reaching the maximum number of iterations, the change in the objective function value being less than a certain threshold, etc.

[0108] Specifically, we first generate a set of initial hyperparameter combinations using random sampling within the feasible range of the number of neurons in the input and intermediate layers. For example, assuming the possible range of the number of neurons in the input layer is [10, 50] and the possible range of the number of neurons in the intermediate layers is [20, 100], we randomly select a number of combinations within these two ranges as the initial hyperparameter combinations. For each initial hyperparameter combination, we construct a corresponding multi-task neural network and train it using the training data. After training, we evaluate the network's performance using the validation data to obtain the initial function value corresponding to each combination.

[0109] Next, a Gaussian process regression model is constructed using the initial hyperparameter combinations and their corresponding initial function values. This model learns the relationship between inputs (hyperparameter combinations) and outputs (objective function values), providing a basis for subsequently predicting the objective function values ​​of unknown hyperparameter combinations. Based on a preset acquisition function, the most promising hyperparameter combination is selected from the currently unevaluated hyperparameter combinations. A multi-task neural network is constructed for the selected hyperparameter combination, trained, and evaluated to obtain the objective function value for that combination. The newly evaluated hyperparameter combination and its objective function value are added to the existing dataset, and the Gaussian process regression model is updated to more accurately predict the objective function value for unknown hyperparameter combinations. The pre-set termination criteria are checked, such as reaching the maximum number of iterations or the change in the objective function value being less than a certain threshold. If not, the next round of selection and evaluation is continued; if satisfied, the iterations are terminated. After the Gaussian process regression model is updated, the combination that minimizes the objective function value is found from all evaluated hyperparameter combinations. The number of neurons in the input and intermediate layers of this combination is the number of neurons in the second and third layers of the input and intermediate layers, respectively.

[0110] Step S204: Optimize the network structure of the initial multi-task neural network according to the first number of neurons, the second number of neurons, and the third number of neurons to obtain a target multi-task neural network.

[0111] It should be noted that the target multi-task neural network is an optimized but untrained version of the initial multi-task neural network. By adjusting the network structure based on the previously determined number of neurons in the first output layer, the second input layer, and the third intermediate layer, a more optimized target multi-task neural network is obtained, improving its performance in tasks such as nonlinear compensation.

[0112] Step S205 , formulating an optimization target according to the signal recovery loss and the computational complexity loss, and training the target multi-task neural network based on the optimization target to obtain an improved multi-task neural network.

[0113] Furthermore, in order to introduce a penalty term related to computational complexity into the loss function during the optimization process, so that the training process not only optimizes the signal recovery performance but also controls the consumption of computing resources, the step S205 may include:

[0114] The computational complexity is obtained by using the number of multiplication operations generated by the connection of neurons in each layer of the target multi-task neural network as a quantitative indicator of computational complexity loss;

[0115] The mean square error between the output signal of the target multi-task neural network and the original undamaged signal is used as the signal recovery loss;

[0116] constructing an optimization target according to the computational complexity, the signal recovery loss, and a hyperparameter between the signal recovery accuracy and the computational complexity;

[0117] The target multi-task neural network is trained using the optimization objective to obtain an improved multi-task neural network.

[0118] It should be noted that computational complexity loss refers to the loss incurred due to the consumption of computational resources during the operation of a neural network. This loss can be quantified by the number of multiplication operations generated by the connections between neurons in each layer of a multi-task neural network. The greater the number of multiplication operations, the higher the computational complexity, which means that more computational resources and time are required to complete the task. Signal recovery loss is used to measure the degree of difference between the output signal of the target multi-task neural network and the original undamaged signal. The mean square error can be used as a measurement indicator. Signal recovery accuracy refers to the accuracy with which the target multi-task neural network can restore the original undamaged signal. Hyperparameters are used to balance the relationship between signal recovery accuracy and computational complexity. Different hyperparameter settings will affect the focus of the optimization objective.

[0119] It should also be noted that the optimization objective is a function that comprehensively considers computational complexity, signal recovery loss, and the relationship between signal recovery accuracy and computational complexity. By adjusting the parameters of the target multi-task neural network to achieve the optimal value for the optimization objective, the network is optimized. The improved multi-task neural network, trained based on the optimization objective, achieves a better balance between signal recovery accuracy and computational complexity than the target multi-task neural network, resulting in superior performance.

[0120] Specifically, the computational complexity is first calculated by counting the number of multiplication operations generated by the neuronal connections in each layer of the target multitask neural network and using this as a quantitative indicator of computational complexity loss. Specifically, for each layer of neuronal connections, the number of multiplication operations between the input signal and the weight is calculated, and then the number of multiplication operations for each layer is summed. Next, the mean squared error between the output signal of the target multitask neural network and the original undamaged signal is calculated as the signal recovery loss. Simultaneously, an optimization objective is constructed based on the computational complexity, signal recovery loss, and the hyperparameter between signal recovery accuracy and computational complexity. Finally, the target multitask neural network is trained using the optimization objective. During the training process, an appropriate optimization algorithm (such as stochastic gradient descent, Adam, etc.) is used to continuously adjust the network parameters (weights and biases) to minimize the value of the optimization objective. The training data can use a large number of historical supplementary samples and their corresponding original undamaged signals.

[0121] In one example, optimizing the network structure and training strategy of the multi-task neural network is the key to improving the efficiency of nonlinear compensation. Traditional loss functions (such as mean square error (MSE) or cross entropy) usually only focus on signal recovery accuracy, while ignoring the impact of computational complexity on system real-time performance. Therefore, during the optimization process, it is possible to consider introducing a penalty term related to computational complexity into the loss function, so that the training process not only optimizes signal recovery performance, but also controls the consumption of computing resources. Specifically, the following optimization objectives can be designed:

[0122] L total =L rec +λL cc

[0123] Among them, L rec Indicates signal recovery loss (such as MSE), L cc Represents the computational complexity loss (the computational complexity approximation index can be used, such as the number of multiplication operations N mul), λ is a hyperparameter that balances signal recovery accuracy and computational complexity and can be prioritized based on the optimization objective. This approach dynamically balances compensation accuracy and computational complexity during the optimization process, resulting in a neural network that achieves both efficient nonlinear compensation and the real-time performance requirements of optical communication systems.

[0124] Without sacrificing signal transmission performance, the computational complexity of the receiving end is reduced; the optimization of nonlinear compensation calculation is achieved, so that the system can cope with complex transmission environments more efficiently; the utilization efficiency of hardware resources is improved, so that the system has real-time processing capabilities in high-speed transmission scenarios; and this application provides technical support for the actual deployment of high-performance optical communication systems, which will help further improve the transmission capacity of the network.

[0125] In this embodiment, a layered optimization strategy was implemented, employing a stepwise traversal method for the output layer of the initial multi-task neural network and Bayesian optimization for the input and intermediate layers, achieving systematic optimization of the network structure. Furthermore, by introducing a penalty term related to computational complexity into the loss function, a balance was ensured between compensation accuracy and computational complexity. The resulting neural network achieves both efficient nonlinear compensation and the real-time performance requirements of optical communication systems.

[0126] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the signal processing method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0127] This application also provides a signal processing device, please refer to Figure 6 , the signal processing device includes:

[0128] A signal preprocessing module, configured to obtain an equalized signal with nonlinear impairments and perform normalization processing on the equalized signal to generate a target input signal;

[0129] A signal compensation module is used to input the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and performs optimization training based on the optimization target constructed according to computational complexity.

[0130] The signal processing device provided in this application, using the signal processing method in the above-mentioned embodiment, can solve the technical problem that existing single-task neural network models require a large amount of sample data and high computing power for nonlinear compensation, and face high costs and energy consumption issues. Compared with the existing technology, the beneficial effects of the signal processing device provided in this application are the same as the beneficial effects of the signal processing method provided in the above-mentioned embodiment, and the other technical features of the signal processing device are the same as those disclosed in the above-mentioned embodiment method, and are not further described here.

[0131] The present application provides a signal processing device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the signal processing method in the above-mentioned embodiment 1.

[0132] Reference below Figure 7 , which shows a schematic diagram of the structure of a signal processing device suitable for implementing the embodiments of the present application. The signal processing device in the embodiments of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7 The signal processing device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0133] like Figure 7As shown, the signal processing device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the signal processing device. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the signal processing device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows a signal processing device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or provided instead.

[0134] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are performed.

[0135] The signal processing device provided in this application, using the signal processing method in the above-mentioned embodiment, can solve the technical problem that existing single-task neural network models require a large amount of sample data and high computing power for nonlinear compensation, and face high costs and energy consumption issues. Compared with the existing technology, the beneficial effects of the signal processing device provided in this application are the same as the beneficial effects of the signal processing method provided in the above-mentioned embodiment, and the other technical features of the signal processing device are the same as those disclosed in the method of the previous embodiment, and are not further described here.

[0136] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0137] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0138] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer program) stored thereon, wherein the computer-readable program instructions are used to execute the signal processing method in the above-mentioned embodiment.

[0139] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0140] The computer-readable storage medium may be included in the signal processing device, or may exist independently without being incorporated into the signal processing device.

[0141] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the signal processing device, the signal processing device executes the signal processing method described above.

[0142] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of code, and the module, program segment or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0144] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0145] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-mentioned signal processing method. It can solve the technical problem that the existing single-task neural network model requires a large amount of sample data and high computing power in nonlinear compensation, and faces high cost and energy consumption problems. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the signal processing method provided in the above-mentioned embodiment, and will not be repeated here.

[0146] The present application also provides a computer program product, comprising a computer program, which implements the steps of the above-mentioned signal processing method when executed by a processor.

[0147] The computer program product provided in this application can address the technical issues faced by existing single-task neural network models in nonlinear compensation, which require large amounts of sample data and high computing power, resulting in high costs and energy consumption. Compared to the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the signal processing methods provided in the above-mentioned embodiments, and are not further elaborated here.

[0148] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A signal processing method, characterized in that: The signal processing method comprises: Acquiring an equalized signal with nonlinear impairment, and performing normalization processing on the equalized signal to generate a target input signal; The target input signal is input into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal method and Bayesian optimization, and performs optimization training according to the optimization target constructed according to computational complexity.

2. The signal processing method according to claim 1, wherein: Before the step of inputting the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, the method further includes: Obtaining an initial multi-task neural network, wherein the initial multi-task neural network includes an input layer, an intermediate layer, and an output layer; Determine the first number of neurons in the optimal configuration of the output layer in the initial multi-task neural network by using a step-by-step traversal method starting from the number of neurons that meets the minimum task requirements and gradually increasing the configuration; Determining the number of second neurons in the input layer and the number of third neurons in the middle layer of the initial multi-task neural network by constructing a Gaussian process regression model based on Bayesian optimization; Optimizing the network structure of the initial multi-task neural network according to the first number of neurons, the second number of neurons, and the third number of neurons to obtain a target multi-task neural network; An optimization target is formulated according to the signal recovery loss and the computational complexity loss, and the target multi-task neural network is trained based on the optimization target to obtain an improved multi-task neural network.

3. The signal processing method according to claim 2, wherein: The step of determining the number of second neurons in the input layer and the number of third neurons in the intermediate layer in the initial multi-task neural network by constructing a Gaussian process regression model based on Bayesian optimization includes: Generating initial hyperparameter combinations by random sampling, and training and evaluating the performance of the multi-task neural network corresponding to each of the initial hyperparameter combinations to obtain initial function values, wherein the initial hyperparameter combinations include different combinations of the number of neurons in the input layer and the intermediate layer; Constructing a Gaussian process regression model based on the initial hyperparameter combination and the corresponding initial function value; Selecting currently unevaluated hyperparameter combinations in sequence according to a preset acquisition function, evaluating the objective function values ​​of the selected hyperparameter combinations, and updating the Gaussian process regression model until a preset iteration termination condition is met; After the Gaussian process regression model is updated, the number of neurons in the input layer and the middle layer that minimizes the objective function value are selected from all evaluated hyperparameter combinations as the second number of neurons in the input layer and the third number of neurons in the middle layer.

4. The signal processing method according to claim 2, wherein: The step of formulating an optimization target according to the signal recovery loss and the computational complexity loss, and training the target multi-task neural network based on the optimization target to obtain an improved multi-task neural network includes: The computational complexity is obtained by using the number of multiplication operations generated by the connection of neurons in each layer of the target multi-task neural network as a quantitative indicator of computational complexity loss; The mean square error between the output signal of the target multi-task neural network and the original undamaged signal is used as the signal recovery loss; constructing an optimization target according to the computational complexity, the signal recovery loss, and a hyperparameter between the signal recovery accuracy and the computational complexity; The target multi-task neural network is trained using the optimization objective to obtain an improved multi-task neural network.

5. The signal processing method according to any one of claims 1 to 4, characterized in that: The step of obtaining an equalized signal with nonlinear damage and performing normalization processing on the equalized signal to generate a target input signal includes: Converting the optical signal transmitted in the link into an electrical signal through a coherent receiver, and sampling the electrical signal based on a digital sampling oscilloscope to obtain a target signal; Performing digital signal processing on the target signal to generate an equalized signal containing only nonlinear impairments, wherein the digital signal processing includes dispersion compensation, frequency offset estimation, and adaptive equalization correction.

6. The signal processing method according to any one of claims 1 to 4, characterized in that: After the step of inputting the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, the method further includes: Passing the compensated signal to a carrier phase recovery module to extract phase information; The compensated signal is demodulated into digital data for output through a decision module according to the phase information.

7. A signal processing device, characterized in that: The device comprises: A signal preprocessing module, configured to obtain an equalized signal with nonlinear impairments and perform normalization processing on the equalized signal to generate a target input signal; A signal compensation module is used to input the target input signal into the improved multi-task neural network for nonlinear compensation to obtain a compensated signal, wherein the multi-task neural network determines the number of neurons in each layer through step-by-step traversal and Bayesian optimization, and optimizes according to the optimization target constructed according to computational complexity.

8. A signal processing device, characterized in that The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the signal processing method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the signal processing method according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the signal processing method according to any one of claims 1 to 6 are implemented.

Images