Multi-channel adaptive neural network equalization method, system and device, and storage medium

Abstract

A multi-channel adaptive neural network equalization method, system and device and a storage medium relate to the field of optical communication signal processing, and specifically comprise: for each sub-channel task, constructing a first sample pair based on a first received signal in the sub-channel task, the first sample pair comprising an input feature vector and a tag vector; performing optimization training on the initial equalizer model based on the first sample pair and a preset meta-learning algorithm to obtain a general equalizer model; for any new sub-channel task, constructing a second sample pair according to a second receiving signal corresponding to the new sub-channel task, and performing fine tuning on the universal equalizer model based on a target training set in the second sample pair to obtain a target equalizer model corresponding to the new sub-channel task; and performing equalization processing on a target received signal corresponding to the new sub-channel task according to the target equalizer model so as to realize multi-channel adaptive neural network equalization. According to the invention, model generalization among different channels can be realized.

Description

Technical Field

[0001] This application relates to the field of optical communication signal processing, specifically to a multi-channel adaptive neural network equalization method, system, device, and storage medium. Background Technology

[0002] Currently, wavelength division multiplexing (WDM) technology is widely used in high-capacity coherent optical communication systems to improve the transmission bandwidth of optical fibers. Each wavelength sub-channel may employ different modulation formats, have different optical signal-to-noise ratios (OSNR), device response characteristics, or user configurations, which leads to differences in the statistical distribution of the signal across the spectral dimension.

[0003] However, traditional digital signal processing (DSP) methods and fixed-parameter neural network equalizers are difficult to generalize and deploy across different channels in wavelength division multiplexing (WDM) systems, requiring retraining of the model for each channel. This not only increases deployment costs but also affects the system's real-time performance. Furthermore, although probabilistic shaping (PS) modulation techniques have improved constellation structure and channel matching, the resulting entropy configuration differences also pose a challenge to the model's generalization ability.

[0004] Therefore, how to provide a multi-channel adaptive neural network equalization method to achieve model generalization across different channels is a problem that urgently needs to be solved. Summary of the Invention

[0005] This application provides a multi-channel adaptive neural network equalization method, system, device, and storage medium, which can achieve model generalization across different channels.

[0006] In a first aspect, embodiments of this application provide a multi-channel adaptive neural network equalization method, the multi-channel adaptive neural network equalization method comprising: For each sub-channel task, a first sample pair is constructed based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector. The initial equalizer model is optimized and trained based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model. For any new sub-channel task, construct a second sample pair based on the second received signal corresponding to the new sub-channel task, and fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task. The target received signal corresponding to the new sub-channel task is equalized according to the target equalizer model to achieve multi-channel adaptive neural network equalization.

[0007] In conjunction with the first aspect, in one embodiment, the input feature vector includes the in-phase component sequence of the X-polarization channel, the orthogonal component sequence of the X-polarization channel, the in-phase component sequence of the Y-polarization channel, and the orthogonal component sequence of the Y-polarization channel.

[0008] In conjunction with the first aspect, in one implementation, the step of optimizing and training the initial equalizer model based on the first sample pair and a preset meta-learning algorithm to obtain a general equalizer model includes: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample pair to obtain task-specific parameters. Calculate the loss on the first test set in the first sample pair with task-specific parameters; The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated by the outer loop gradient based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

[0009] In conjunction with the first aspect, in one implementation, the expression for the task-specific parameter is:

[0010] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; Set the preset learning rate for the inner loop; Preset global parameters The gradient; To use preset global parameters The initial equalizer model configured; For the initial equalizer model In the training set The loss function on the given surface is expressed as:

[0011] In the formula, The number of the first sample pairs; Let be the input feature vector of the j-th first sample pair; Let be the label vector of the j-th first sample pair.

[0012] In conjunction with the first aspect, in one implementation, the expression for the final parameter is:

[0013] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; To use task-specific parameters A configured general-purpose equalizer model; For a universal equalizer model The first test set in the sub-channel task loss function on; The preset outer loop learning rate; This is the final parameter.

[0014] In conjunction with the first aspect, in one implementation, the expression for the target parameters of the target equalizer model is:

[0015] In the formula, For final parameters; For the target training set; For the final parameters on the target training set loss function on; Set the preset learning rate for the inner loop; For final parameters The gradient; These are the target parameters corresponding to the target equalizer model.

[0016] Secondly, embodiments of this application provide a multi-channel adaptive neural network equalization system, the multi-channel adaptive neural network equalization system comprising: The first processing module is used to construct a first sample pair for each sub-channel task based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector. The second processing module is used to optimize and train the initial equalizer model based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model. The third processing module is used to construct a second sample pair based on the second received signal corresponding to the new sub-channel task for any new sub-channel task, and to fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task. The fourth processing module is used to perform equalization processing on the target received signal corresponding to the new sub-channel task according to the target equalizer model, so as to realize multi-channel adaptive neural network equalization.

[0017] In conjunction with the second aspect, in one implementation, the second processing module is specifically used for: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample to obtain task-specific parameters. Calculate the loss on the first test set with the first sample and the task-specific parameters. The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated in an outer loop based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

[0018] Thirdly, embodiments of this application provide a multi-channel adaptive neural network equalization device, the multi-channel adaptive neural network equalization device including a processor, a memory, and a multi-channel adaptive neural network equalization program stored in the memory and executable by the processor, wherein when the multi-channel adaptive neural network equalization program is executed by the processor, it implements the steps of the multi-channel adaptive neural network equalization method as described in any of the preceding claims.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a multi-channel adaptive neural network equalization program, wherein when the multi-channel adaptive neural network equalization program is executed by a processor, it implements the steps of the multi-channel adaptive neural network equalization method as described in any of the preceding claims.

[0020] The beneficial effects of the technical solutions provided in this application include: For each sub-channel task, a first sample pair, including an input feature vector and a label vector, is constructed based on the first received signal in the sub-channel task. By constructing sample pairs separately for each sub-channel task, the model can learn the statistical differences between different sub-channels. Based on the first sample pair and a preset meta-learning algorithm, the initial equalizer model is optimized and trained to obtain a general equalizer model. The meta-learning algorithm ensures that the training process considers multiple sub-channel tasks simultaneously, rather than a single channel task, ultimately resulting in a general equalizer model with cross-channel generalization ability, which becomes the basic model applicable to multiple sub-channels. For any new sub-channel task, a second sample pair is constructed based on the second received signal corresponding to the new sub-channel task. The general equalizer model is fine-tuned based on the target training set in the second sample pair to obtain a target equalizer model corresponding to the new sub-channel task. Only a small amount of target training set for the new sub-channel is needed to obtain a target equalizer model adapted to a specific sub-channel, which directly reflects the cross-channel generalization ability of the general equalizer model. It can quickly adapt to new sub-channels without retraining the complete model for each channel. Finally, the target received signal corresponding to the new sub-channel task is equalized according to the target equalizer model to achieve multi-channel adaptive neural network equalization. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the multi-channel adaptive neural network equalization method of this application; Figure 2 This is a schematic diagram of the Transformer structure in an embodiment of the multi-channel adaptive neural network equalization method of this application; Figure 3 This is a schematic diagram of the meta-learning training process in an embodiment of the multi-channel adaptive neural network equalization method of this application; Figure 4 This is a functional module diagram of an embodiment of the multi-channel adaptive neural network equalization system of this application; Figure 5 This is a schematic diagram of the hardware structure of the multi-channel adaptive neural network equalization device involved in the embodiments of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0024] In a first aspect, embodiments of this application provide a multi-channel adaptive neural network equalization method.

[0025] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the multi-channel adaptive neural network equalization method of this application. Figure 1 As shown, the multi-channel adaptive neural network equalization method includes: Step S10: For each sub-channel task, construct a first sample pair based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector.

[0026] In an exemplary embodiment of this application, in a wavelength division multiplexing optical communication system, each sub-channel task refers to a signal equalization processing task for a specific wavelength channel in the system; specifically, each sub-channel task The i-th wavelength channel in the WDM system contains different modulation types (such as 16QAM, 32QAM, 64QAM, probabilistically shaped PS-64QAM, and probabilistically shaped PS-256QAM), different probabilistically shaped entropy levels (from 4 to 7.8), different optical signal-to-noise ratio (OSNR) parameters, transmitter / receiver device differences, and other characteristics; each task sample set (corresponding to a specific sub-channel task) ) contains 2 n Each send and receive sample pair can be denoted as ,in, This represents the j-th input feature of the model, i.e., the input feature vector. j ; This represents the j-th equilibrium objective of the model, which is also known as the label vector. j .

[0027] Specifically, for each sub-channel task, the first received signal corresponding to that sub-channel is acquired. This received signal includes the timing data of the I-component (in-phase component) and Q-component (quadrature component) of the damaged X-polarization channel and Y-polarization channel. Subsequently, for each discrete time j in the first received signal, a symmetrical sliding window mechanism is used to extract a signal segment centered at time j with a window length of 2m+1, where m is a preset window radius parameter, the specific value of which can be determined according to actual needs and is not limited here. The I-component sequence of the X-polarization channel, the Q-component sequence of the X-polarization channel, the I-component sequence of the Y-polarization channel, and the Q-component sequence of the Y-polarization channel are extracted from the received signal respectively, and these four-dimensional sequence data are sequentially concatenated to form a 4(2m+1)-dimensional input feature vector Input. j .

[0028] It should be noted that the X-polarization channel I component, X-polarization channel Q component, Y-polarization channel I component, and Y-polarization channel Q component at this discrete time j are extracted from the ideal transmitted signal corresponding to the first received signal (referring to the pure signal state that the transmitter originally intended to transmit in the optical communication system, unaffected by any channel impairment, representing the original signal form that the communication system expects to recover at the receiver) and combined to form a 4-dimensional label vector. j As the goal of supervised learning; specifically, [ , , ,in , This represents the ideal transmitted signal with X-polarization; Represents the ideal transmitted signal with Y polarization; the input feature vectors corresponding to the same discrete time j are... j With label vector j Pairing together forms the first sample pair (Input) j Label j This sample fully characterizes the mapping relationship from the damaged received signal to the ideal transmitted signal, providing a basic training unit for the subsequent learning of the neural network equalizer. The input feature vector captures the temporal correlation and polarization coupling characteristics of the signal through a sliding window, while the label vector represents the distortion-free signal state to be recovered. Together, they constitute the complete sample required for supervised learning, enabling the model to learn the nonlinear inverse mapping relationship from the channel-damaged signal to the original transmitted signal, thereby achieving effective compensation for optical communication channel distortion and noise.

[0029] Step S20: Optimize and train the initial equalizer model based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model.

[0030] In this embodiment, as an example, the predefined meta-learning algorithm refers to a predetermined meta-learning framework (such as a variant algorithm based on MAML) suitable for multi-sub-channel tasks. This algorithm simulates the process of "learning how to learn," enabling the equalizer model to extract common features from multiple sub-channel tasks and gain the ability to quickly adapt to new sub-channel tasks. Specifically, the first sample pair in each sub-channel task can be divided into a training set and a test set, wherein the training set (i.e., the support set) can be represented as... This is used to perform an inner loop update on the initial equalizer model to obtain task-specific parameters. The test set (i.e., the query set) can be represented as... It is used to evaluate the generalization performance of task-specific parameters on the sub-channel task; it should be noted that in meta-learning, the inner loop refers to the fast learning process on a single task, in which the model uses a small amount of support set data to quickly update the parameters to adapt to the current task.

[0031] Specifically, the initial equalizer model refers to the neural network equalizer infrastructure pre-built before meta-learning training. Its network structure (such as the multi-head attention mechanism based on the Transformer encoder) has been pre-designed according to the characteristics of the optical communication signal equalization task, but the model parameters (i.e., the preset global parameters) are in a random initialization state or are set to initial values ​​using standard initialization methods (such as Xavier initialization). As the starting point of the meta-learning algorithm, this model has not yet learned any specific features of the sub-channel task and does not have the ability to directly and effectively equalize optical communication signals.

[0032] It should be noted that, referring to Figure 2 The Transformer architecture diagram shown below, with input signals (such as...) / , / First, the input undergoes embedding processing, followed by an encoder consisting of multiple stacked identical layers. Within each layer, the input is split into multiple subspaces using a multi-head segmentation mechanism, generating query (Q), key (K), and value (V) vectors, such as Q1. Q4, K1 K4, V1 V4; Next, using the scaled dot product attention mechanism, the similarity between the query vector and the key vector is calculated and scaled, then subjected to a softmax operation, and finally weighted and summed with the value vector to obtain the attention output Z1 of each head. Z4; These attention outputs are then concatenated and transformed through a linear layer to obtain the final output signal (e.g., / , / This structure can effectively capture long-distance dependencies and complex patterns in input signals, and has important application value in tasks such as optical communication signal equalization.

[0033] It should be understood that the pre-defined meta-learning algorithm iteratively optimizes the pre-defined global parameters of the initial equalizer model by minimizing the cumulative loss of all sub-channel tasks on the test set in the outer loop. This results in a general equalizer model that can quickly adapt to the characteristics of each sub-channel task. This model retains the ability to learn the common features of multiple sub-channels, enabling efficient signal equalization to be achieved through fine-tuning with only a small number of support set samples when facing new sub-channel tasks, without having to retrain the complete model for each sub-channel. It should be noted that the outer loop refers to the learning process performed on multiple tasks. By optimizing meta-parameters (such as the parameters of the meta-learner), the model can better generalize to new tasks.

[0034] Step S30: For any new sub-channel task, construct a second sample pair based on the second received signal corresponding to the new sub-channel task, and fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task.

[0035] In an exemplary embodiment of this application, for any new sub-channel task, a second received signal corresponding to a wavelength channel that has not appeared in the training phase is obtained, and a second sample pair is constructed based on the second received signal. The principle and implementation process are the same as constructing a first sample pair based on the first received signal, and will not be repeated here for the sake of brevity. A portion of the sample pairs are selected from the second sample pairs as a target training set. The size of the target training set is smaller than the size of the training set for each sub-channel task in the meta-learning training phase, and contains only a small number of continuous signal segments. The target training set is then used to perform gradient-based inner loop updates on the trained general equalizer model. Specifically, the gradient of the loss function with respect to the model parameters is calculated and the parameters are adjusted along the negative gradient direction. The fine-tuning process is completed after a finite number of iterations (preferably 5-10 times). Finally, a target equalizer model matching the characteristics of the new sub-channel task is obtained. This model retains the basic structure of the general equalizer model and only performs local optimization for the specific channel impairment characteristics of the new sub-channel. Thus, it can achieve effective equalization of the signal of the new sub-channel without retraining the complete model, which significantly reduces the computational complexity and time delay of the system adapting to the new channel.

[0036] Step S40: Equalize the target received signal corresponding to the new sub-channel task according to the target equalizer model to achieve multi-channel adaptive neural network equalization.

[0037] In an exemplary embodiment of this application, the target received signal in the new sub-channel task is input to the finely tuned target equalizer model. By sequentially processing the signal segments at discrete moments in the target received signal, the target equalizer model can continuously output the equalized signal sequence to effectively compensate for the channel distortion and noise interference unique to the new sub-channel. Since the target equalizer model is obtained by fine-tuning a general equalizer model with a small number of samples, its parameters have been adaptively optimized for the transmission characteristics of the sub-channel. Therefore, high-quality signal equalization can be achieved without retraining the complete model. This processing mechanism enables the system to build dedicated target equalizer models for different sub-channel tasks, thereby supporting the signal equalization requirements of multiple wavelength channels simultaneously in a wavelength division multiplexing system, and ultimately achieving multi-channel adaptive neural network equalization.

[0038] This application constructs a first sample pair, including an input feature vector and a label vector, based on the first received signal in each sub-channel task. By constructing sample pairs separately for each sub-channel task, the model can learn the statistical differences between different sub-channels. Based on the first sample pair and a preset meta-learning algorithm, the initial equalizer model is optimized and trained to obtain a general equalizer model. The meta-learning algorithm ensures that the training process considers multiple sub-channel tasks simultaneously, rather than a single channel task, ultimately resulting in a general equalizer model with cross-channel generalization ability, which becomes the basic model applicable to multiple sub-channels. For any new sub-channel task, a second sample pair is constructed based on the second received signal corresponding to the new sub-channel task. The general equalizer model is fine-tuned based on the target training set in the second sample pair to obtain a target equalizer model corresponding to the new sub-channel task. Only a small amount of target training set for the new sub-channel is needed to obtain a target equalizer model adapted to a specific sub-channel, which directly reflects the cross-channel generalization ability of the general equalizer model. It can quickly adapt to new sub-channels without retraining the complete model for each channel. Finally, the target received signal corresponding to the new sub-channel task is equalized according to the target equalizer model to achieve multi-channel adaptive neural network equalization.

[0039] Furthermore, in one embodiment, the input feature vector includes the in-phase component sequence of the X-polarization channel, the orthogonal component sequence of the X-polarization channel, the in-phase component sequence of the Y-polarization channel, and the orthogonal component sequence of the Y-polarization channel.

[0040] As an example, in this embodiment, the input feature vector includes the in-phase component sequence of the X-polarization channel, the orthogonal component sequence of the X-polarization channel, the in-phase component sequence of the Y-polarization channel, and the orthogonal component sequence of the Y-polarization channel; specifically, the I component sequence of the X-polarization channel... Q component sequence of X-polarization channel I-component sequence of Y-polarization channel and the Q component sequence of the Y polarization channel .

[0041] Further, in one embodiment, the step of optimizing and training the initial equalizer model based on the first sample pair and a preset meta-learning algorithm to obtain a general equalizer model includes: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample pair to obtain task-specific parameters. Calculate the loss on the first test set in the first sample pair with task-specific parameters; The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated by the outer loop gradient based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

[0042] In an exemplary embodiment of this application, for each sub-channel task, a first training set is extracted from the first sample pair corresponding to that sub-channel task. This first training set includes sample pairs composed of multiple input feature vectors and label vectors constructed from the first received signal. Subsequently, using the preset global parameter θ of the initial equalizer model as an initial value, the predicted output of each sample in the first training set is calculated through forward propagation, and compared with the label vector to calculate the loss value. This loss value uses mean squared error to measure the difference between the predicted received signal and the ideal transmitted signal. Based on this, the gradient of the loss function with respect to the preset global parameter θ is calculated using the backpropagation algorithm, and the preset global parameter is updated along the negative gradient direction according to a preset inner loop learning rate α, ultimately obtaining the task-specific parameters. The model retains the general feature extraction capability of the initial equalizer model, while performing local optimizations for the channel characteristics of the current sub-channel task. This enables the model to quickly adapt to the signal equalization requirements of this specific wavelength channel, providing task-level adaptation results for subsequent outer loop global parameter optimization based on the test set.

[0043] It should be noted that, for each sub-channel task, the system first applies the task-specific parameters obtained through the inner loop gradient update to the first test set of the first sample pair corresponding to that sub-channel task. This first test set consists of sample pairs that did not participate in the inner loop update and is used to evaluate the generalization performance of the task-specific parameters on that sub-channel task. Specifically, each input feature in the first test set is sequentially input into the system using the task-specific parameters. The equalizer model calculates the predicted output through forward propagation and compares it with the corresponding label vector to calculate the loss value. This loss value uses mean square error to measure the difference between the predicted received signal and the ideal transmitted signal, and can be expressed as follows: Then, the weighted sum or arithmetic average of the individual losses across all sub-channel tasks is taken to determine the cumulative loss. Based on the cumulative loss, its gradient with respect to the preset global parameter θ is calculated, and the preset global parameter is updated along the negative gradient direction according to the preset outer loop learning rate β. This update process is repeatedly executed iteratively during the meta-learning training phase. Each iteration includes inner and outer loop steps for all sub-channel tasks until the change in the preset global parameter θ is less than the preset convergence threshold or the decrease in the cumulative loss is less than the preset stopping criterion (the specific values ​​of the preset convergence threshold and the preset stopping criterion can be determined according to actual needs and are not limited here). At this point, the preset global parameter is considered to have converged, and the optimized general equalizer model is obtained. This model learns a general feature representation applicable to multi-sub-channel tasks through iterative updates, so that it can quickly adapt through the inner loop with only a small number of samples when facing new sub-channel tasks.

[0044] It should be noted that after the preset global parameters have reached convergence, the converged preset global parameters θ are used as the final parameters of the initial equalizer model. These parameters retain the common features extracted from multiple sub-channel tasks through the meta-learning framework, forming a general representation capability for optical communication channel distortion modes. Specifically, using the converged preset global parameters as the final parameters of the initial equalizer model enables the generated general equalizer model to have a "rapid adaptation" characteristic. That is, when faced with a new sub-channel task, only a small number of training samples are needed to achieve good equalization performance through fine-tuning in the inner loop. This parameter determination process marks the completion of the meta-learning training phase. The generated general equalizer model can serve as the base model for subsequent rapid fine-tuning for each sub-channel task, without the need to retrain the complete model for each new sub-channel, thereby achieving the effect of multi-channel adaptive neural network equalization.

[0045] Furthermore, in one embodiment, the expression for the task-specific parameter is:

[0046] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; Set the preset learning rate for the inner loop; Preset global parameters The gradient; To use preset global parameters The initial equalizer model configured; For the initial equalizer model In the training set The loss function on the given surface is expressed as:

[0047] In the formula, The number of the first sample pairs; Let be the input feature vector of the j-th first sample pair; Let be the label vector of the j-th first sample pair.

[0048] As an example, in the embodiments of this application, the number of the first sample pairs is... The input feature vector of the j-th first sample pair The label vector of the j-th first sample pair Substituting into the following formula yields the initial equalizer model. In the training set loss function on The expression:

[0049] Specifically, the preset global parameters of the initial equalizer model are... Initial equalizer model In the training set loss function on Preset inner loop learning rate Preset global parameters gradient And with preset global parameters Initial equalizer model configuration Substituting into the following formula yields the task-specific parameters updated using the first training set for the i-th sub-channel task. The expression:

[0050] Furthermore, in one embodiment, the expression for the final parameter is:

[0051] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; To use task-specific parameters A configured general-purpose equalizer model; For a universal equalizer model The first test set in the sub-channel task loss function on; The preset outer loop learning rate; This is the final parameter.

[0052] As an example, in the embodiments of this application, the preset global parameters of the initial equalizer model are... The task-specific parameters updated using the first training set are applied to the i-th sub-channel task. , with task-specific parameters Configurable general equalizer model Universal equalizer model The first test set in the sub-channel task loss function on and preset outer loop learning rate Substituting into the following formula yields the final parameters. The expression:

[0053] Furthermore, in one embodiment, the expression for the target parameters of the target equalizer model is:

[0054] In the formula, For final parameters; For the target training set; For the final parameters on the target training set loss function on; Set the preset learning rate for the inner loop; For final parameters The gradient; These are the target parameters corresponding to the target equalizer model.

[0055] As an example, in the embodiments of this application, the final parameters are... Target training set The final parameters in the target training set loss function on Preset inner loop learning rate Final parameters gradient Substituting into the following formula yields the target parameters corresponding to the target equalizer model. The expression:

[0056] It should be noted that, referring to Figure 3 As shown, the signal data is first divided into S-band, C-band, and L-band according to band characteristics, and further subdivided into meta-training set and meta-test set. For each band, a corresponding channel i (e.g., channel 1 to channel N) is constructed, and each channel obtains a support set and a query set from the meta-training set. The system uses the meta-learner F as the core guide to generate multiple learners f corresponding to specific channels. i The learner employs a multi-head attention mechanism, a feedforward network, and a residual connection structure internally; in the inner loop, the learner f... i First use support set Si Training is performed by calculating the support set loss function. And update the parameters using gradient descent. To quickly adapt to the channel task, the updated parameters are then used in the query set Q. i Test and calculate the loss. The outer loop calculates the meta-loss function based on the query set loss of each channel in the inner loop. The meta-learner F parameters are updated using the meta-learning rate β via gradient backpropagation. , To optimize strategies; the system also introduces a meta-test set (new task) By calculating its loss function and update rules Optimize the parameters of the meta-equalizer to enhance its generalization ability to new tasks, and form an overall optimization closed loop from local to global and from known tasks to new tasks to improve the performance and adaptability of multi-channel signal processing.

[0057] Secondly, embodiments of this application also provide a multi-channel adaptive neural network equalization system.

[0058] In one embodiment, reference is made to Figure 4 , Figure 4 This is a functional module diagram of an embodiment of the multi-channel adaptive neural network equalization system of this application. Figure 4 As shown, the multi-channel adaptive neural network equalization system includes: The first processing module is used to construct a first sample pair for each sub-channel task based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector. The second processing module is used to optimize and train the initial equalizer model based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model. The third processing module is used to construct a second sample pair based on the second received signal corresponding to the new sub-channel task for any new sub-channel task, and to fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task. The fourth processing module is used to perform equalization processing on the target received signal corresponding to the new sub-channel task according to the target equalizer model, so as to realize multi-channel adaptive neural network equalization.

[0059] Furthermore, in one embodiment, the second processing module is specifically used for: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample to obtain task-specific parameters. Calculate the loss on the first test set with the first sample and the task-specific parameters. The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated in an outer loop based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

[0060] Furthermore, in one embodiment, the first processing module is specifically used for: The input feature vector includes the in-phase component sequence of the X-polarization channel, the orthogonal component sequence of the X-polarization channel, the in-phase component sequence of the Y-polarization channel, and the orthogonal component sequence of the Y-polarization channel.

[0061] Furthermore, in one embodiment, the second processing module is specifically used for: The expression for the task-specific parameter is:

[0062] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; Set the preset learning rate for the inner loop; Preset global parameters The gradient; To use preset global parameters The initial equalizer model configured; For the initial equalizer model In the training set The loss function on the given surface is expressed as:

[0063] In the formula, The number of the first sample pairs; Let be the input feature vector of the j-th first sample pair; Let be the label vector of the j-th first sample pair.

[0064] Furthermore, in one embodiment, the second processing module is specifically used for: The expression for the final parameter is:

[0065] In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; To use task-specific parameters A configured general-purpose equalizer model; For a universal equalizer model The first test set in the sub-channel task loss function on; The preset outer loop learning rate; This is the final parameter.

[0066] Furthermore, in one embodiment, the third processing module is specifically used for: The expression for the target parameters of the target equalizer model is:

[0067] In the formula, For final parameters; For the target training set; For the final parameters on the target training set loss function on; Set the preset learning rate for the inner loop; For final parameters The gradient; These are the target parameters corresponding to the target equalizer model.

[0068] This application constructs a first sample pair, including an input feature vector and a label vector, based on the first received signal in each sub-channel task. By constructing sample pairs separately for each sub-channel task, the model can learn the statistical differences between different sub-channels. Based on the first sample pair and a preset meta-learning algorithm, the initial equalizer model is optimized and trained to obtain a general equalizer model. The meta-learning algorithm ensures that the training process considers multiple sub-channel tasks simultaneously, rather than a single channel task, ultimately resulting in a general equalizer model with cross-channel generalization ability, which becomes the basic model applicable to multiple sub-channels. For any new sub-channel task, a second sample pair is constructed based on the second received signal corresponding to the new sub-channel task. The general equalizer model is fine-tuned based on the target training set in the second sample pair to obtain a target equalizer model corresponding to the new sub-channel task. Only a small amount of target training set for the new sub-channel is needed to obtain a target equalizer model adapted to a specific sub-channel, which directly reflects the cross-channel generalization ability of the general equalizer model. It can quickly adapt to new sub-channels without retraining the complete model for each channel. Finally, the target received signal corresponding to the new sub-channel task is equalized according to the target equalizer model to achieve multi-channel adaptive neural network equalization.

[0069] The functions of each module in the above-mentioned multi-channel adaptive neural network equalization system correspond to the steps in the above-mentioned multi-channel adaptive neural network equalization method embodiment, and their functions and implementation processes will not be described in detail here.

[0070] Thirdly, embodiments of this application provide a multi-channel adaptive neural network equalization device, which can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0071] Reference Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of the multi-channel adaptive neural network equalization device involved in the embodiments of this application. In the embodiments of this application, the multi-channel adaptive neural network equalization device may include a processor, a memory, a communication interface, and a communication bus.

[0072] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0073] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting devices within the multi-channel adaptive neural network equalization device, as well as interfaces used for interconnecting the multi-channel adaptive neural network equalization device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0074] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0075] The processor can be a general-purpose processor, which can call a multi-channel adaptive neural network equalization program stored in memory and execute the multi-channel adaptive neural network equalization method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the multi-channel adaptive neural network equalization program is called can be referred to in the various embodiments of the multi-channel adaptive neural network equalization method of this application, and will not be repeated here.

[0076] Those skilled in the art will understand that Figure 5 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0077] Fourthly, embodiments of this application also provide a readable storage medium.

[0078] This application stores a multi-channel adaptive neural network equalization program on a readable storage medium, wherein when the multi-channel adaptive neural network equalization program is executed by a processor, it implements the steps of the multi-channel adaptive neural network equalization method as described above.

[0079] The method implemented when the multi-channel adaptive neural network equalization program is executed can be referred to in various embodiments of the multi-channel adaptive neural network equalization method of this application, and will not be repeated here.

[0080] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0081] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0082] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0083] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0084] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0086] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-channel adaptive neural network equalization method, characterized in that, The multi-channel adaptive neural network equalization method includes: For each sub-channel task, a first sample pair is constructed based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector. The initial equalizer model is optimized and trained based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model. For any new sub-channel task, construct a second sample pair based on the second received signal corresponding to the new sub-channel task, and fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task. The target received signal corresponding to the new sub-channel task is equalized according to the target equalizer model to achieve multi-channel adaptive neural network equalization.

2. The multi-channel adaptive neural network equalization method as described in claim 1, characterized in that, The input feature vector includes the in-phase component sequence of the X-polarization channel, the orthogonal component sequence of the X-polarization channel, the in-phase component sequence of the Y-polarization channel, and the orthogonal component sequence of the Y-polarization channel.

3. The multi-channel adaptive neural network equalization method as described in claim 1, characterized in that, The process of optimizing and training the initial equalizer model based on the first sample pair and a preset meta-learning algorithm to obtain a general equalizer model includes: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample pair to obtain task-specific parameters. Calculate the loss on the first test set in the first sample pair with task-specific parameters; The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated by the outer loop gradient based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

4. The multi-channel adaptive neural network equalization method as described in claim 3, characterized in that, The expression for the task-specific parameter is: In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; Set the preset learning rate for the inner loop; Preset global parameters The gradient; To use preset global parameters The initial equalizer model configured; For the initial equalizer model In the training set The loss function on the given surface is expressed as: In the formula, The number of the first sample pairs; Let be the input feature vector of the j-th first sample pair; Let be the label vector of the j-th first sample pair.

5. The multi-channel adaptive neural network equalization method as described in claim 3, characterized in that, The expression for the final parameter is: In the formula, These are the preset global parameters for the initial equalizer model; To use the updated task-specific parameters from the first training set on the i-th sub-channel task; To use task-specific parameters A configured general-purpose equalizer model; For a universal equalizer model The first test set in the sub-channel task loss function on; The preset outer loop learning rate; This is the final parameter.

6. The multi-channel adaptive neural network equalization method as described in claim 3, characterized in that, The expression for the target parameters of the target equalizer model is: In the formula, For final parameters; For the target training set; For the final parameters on the target training set loss function on; Set the preset learning rate for the inner loop; For final parameters The gradient; These are the target parameters corresponding to the target equalizer model.

7. A multi-channel adaptive neural network equalization system, characterized in that, The multi-channel adaptive neural network equalization system includes: The first processing module is used to construct a first sample pair for each sub-channel task based on the first received signal in the sub-channel task. The first sample pair includes an input feature vector and a label vector. The second processing module is used to optimize and train the initial equalizer model based on the first sample pair and the preset meta-learning algorithm to obtain a general equalizer model. The third processing module is used to construct a second sample pair based on the second received signal corresponding to the new sub-channel task for any new sub-channel task, and to fine-tune the general equalizer model based on the target training set in the second sample pair to obtain the target equalizer model corresponding to the new sub-channel task. The fourth processing module is used to perform equalization processing on the target received signal corresponding to the new sub-channel task according to the target equalizer model, so as to realize multi-channel adaptive neural network equalization.

8. The multi-channel adaptive neural network equalization system as described in claim 7, characterized in that, The second processing module is specifically used for: For each sub-channel task, the preset global parameters of the initial equalizer model are updated by an inner loop gradient based on the first training set in the first sample to obtain task-specific parameters. Calculate the loss on the first test set with the first sample and the task-specific parameters. The cumulative loss is determined based on the loss of all sub-channel tasks, and the preset global parameters are updated in an outer loop based on the cumulative loss until the preset global parameters converge. The converged preset global parameters are used as the final parameters of the initial equalizer model to generate a general equalizer model.

9. A multi-channel adaptive neural network equalization device, characterized in that, The multi-channel adaptive neural network equalization device includes a processor, a memory, and a multi-channel adaptive neural network equalization program stored in the memory and executable by the processor, wherein when the multi-channel adaptive neural network equalization program is executed by the processor, it implements the steps of the multi-channel adaptive neural network equalization method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a multi-channel adaptive neural network equalization program, wherein when the multi-channel adaptive neural network equalization program is executed by a processor, it implements the steps of the multi-channel adaptive neural network equalization method as described in any one of claims 1 to 6.

Images