Sign Detection Method Based on Meta-Regularization

Abstract

The present invention belongs to the field of radio communication symbol detection, and specifically designs a symbol detection method based on meta-regularization. The method comprises collecting the first channel output symbol of the symbol to be tested; inputting the first channel output symbol of the symbol to be tested into a trained self-supervised detection neural network model, and outputting the first channel input symbol of the symbol to be tested; the present invention adds consistent meta-regularization to the meta-learning framework MAML to align the models trained on the support set and query set in each task to eliminate data distribution differences between them and improve the ability of meta-knowledge to maintain data consistency; the present invention improves the accuracy of symbol detection, achieves more accurate detection performance, and solves the problems of channel resource waste and repeated training in traditional machine learning methods.

Description

Technical Field

[0001] The present invention belongs to the field of symbol detection in radio communications, and in particular relates to a symbol detection method based on element regularization. Background Art

[0002] The fundamental task of a digital receiver is to reconstruct the observed transmitted signal, which has been affected by noise and channel impairments, to reliably recover the transmitted symbols. This task is commonly referred to as symbol detection. Traditional symbol detection algorithms are not applicable in some situations. For example, algorithms based on the Maximum A Posteriori Probability (MAP) rule require a complete understanding of the channel model and parameters. When the channel model is complex, these conventional methods cannot effectively capture the underlying physical information of the system. Furthermore, traditional model-based detection algorithms rely on Channel State Information (CSI) and require the detection of instantaneous parameters. This process reduces the data rate, and inaccurate parameter estimates can degrade detection performance. In time-varying channels, due to the large pilot overhead, adding pilots to obtain instantaneous CSI results in a significant loss of spectral efficiency. Therefore, without adding pilots, reliably recovering the transmitted symbols in instantaneous channels is a significant challenge.

[0003] The development of meta-learning (ML) overcomes this drawback, enabling rapid learning of a target with minimal examples. Unlike currently popular learning methods, meta-learning ultimately aims to acquire prior knowledge and derive more optimized algorithms through multiple learning tasks. Rather than learning how to solve specific tasks, it aims to master the skill of learning through previously completed tasks. Traditional AI approaches rely on machines learning, while meta-learning aims to empower machines to learn. Generally speaking, meta-learning leverages existing knowledge and experience to generate autonomous learning models and acquire the skill of "learning to learn." Only when machines possess this capability can true AI be achieved, and therefore, meta-learning has become a hot research area.

[0004] Meta-learning-based symbol detection algorithms employ a model-agnostic meta-learning (MAML) framework to train optimal initialization parameters. Based on these parameters, not only can the model be retrained, but the hyperparameters that determine the retraining process can also be optimized, allowing the model to quickly adapt to new channel conditions in a small number of update steps. This significantly improves detection efficiency. In short, meta-learning-based symbol detection technology holds great promise and deserves further exploration.

[0005] In recent years, many scholars have conducted extensive research on meta-learning-based symbol detection algorithms. MJPark et al. (MJPark, J.Ok, Y.-S.Jeon and D.Kim, "MetaSSD: Meta-Learned Self-Supervised Detection," 2022 IEEE International Symposium on Information Theory (ISIT), Espoo, Finland, 2022, pp.480-485, doi:10.1109 / ISIT50566.2022.9834613.) proposed a new symbol detection framework. The meta-learning-based symbol detection algorithm MetaSSD combines MAML with symbol detection to promote rapid adaptation to channel state information. Although this method can be used to recover information when CSI is unknown, it does not significantly improve the performance of reducing the symbol error rate.

[0006] Based on the current research on symbol detection based on meta-learning, we find that there are still challenges in the process of using meta-learning and detection networks for symbol detection:

[0007] While MAML-based symbol detection algorithms overcome the limitations of traditional methods, the data distributions of the support set (training set) and query set (test set) in their meta-learning tasks are inconsistent, making it difficult to fully describe the true data distribution. This results in poor generalization of the model learned on the support set to the query set, hindering the achievement of better symbol error rate performance. Summary of the Invention

[0008] To solve the above technical problems, the present invention proposes a symbol detection method based on meta-regularization, comprising the following steps:

[0009] Collect the first channel output symbol of the symbol to be tested;

[0010] Inputting the first channel output symbol of the symbol to be tested into the trained self-supervised detection neural network model, and outputting the first channel input symbol of the symbol to be tested. The training process of the self-supervised detection neural network model includes:

[0011] Acquire a second channel input symbol, a second channel output symbol, and second channel information of a training symbol;

[0012] The obtained training symbols are divided into a support set and a query set, and the corresponding second channel output symbols and second channel information are respectively input into a multilayer perceptron to estimate the predicted distribution and corresponding weight of the second channel input symbols;

[0013] Calculating a cross entropy loss of a training symbol based on a cross entropy of the obtained second channel input symbol and the estimated predicted distribution of the second channel input symbol;

[0014] Calculating a self-supervisory loss of the training symbol based on an estimated difference between the acquired second channel output symbol and a second channel output signal reconstructed based on a predicted distribution of the second channel input symbol;

[0015] Calculate the consistent meta-regularization loss of the training symbols based on the difference in parameters between training the model on the support set and training the model on the query set;

[0016] Based on the cross entropy loss, self-supervision loss and consistent element regularization loss, the estimated loss of the training symbols is calculated, and the multilayer perceptron is optimized by the estimated loss until the maximum number of iterations or the accuracy requirement of the multilayer perceptron is reached, and the training is terminated.

[0017] Beneficial effects of the present invention:

[0018] The present invention uses symbol detection for different channels as different training tasks, and adopts different training tasks to train the self-supervised detection neural network model, thereby improving the accuracy of symbol detection; the present invention also takes into account the limitations of the single MAML meta-learning framework for symbol detection algorithms, in order to better detect symbols more accurately. By adding consistent meta-regularization to the meta-learning framework MAML, the models trained on the support set and query set in each task are aligned to eliminate the data distribution differences between them and improve the ability of meta-knowledge to maintain data consistency. The accuracy of symbol detection is improved, more accurate detection performance is achieved, and the problems of channel resource waste and repeated training in traditional machine learning methods are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of a symbol detection method based on element regularization according to an embodiment of the present invention;

[0020] Figure 2 This is a model diagram of a self-supervised detection neural network system with L=2 according to an embodiment of the present invention;

[0021] Figure 3 This is a flowchart of a self-supervised detection neural network training according to an embodiment of the present invention;

[0022] Figure 4 This is a data processing flow chart for self-supervised detection neural network training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] It should be noted that, in the embodiment of the present invention, the first channel output symbol, the first channel input symbol and the first channel information refer to the output symbol, input symbol and channel information of the training symbol; the second channel output symbol, the second channel input symbol and the second channel information refer to the output symbol, input symbol and channel information of the symbol to be tested; the "first channel" and the "second channel" here are only used to distinguish whether the symbol is a training symbol or a symbol to be tested, and are not intended to limit the number of channels.

[0025] Figure 1 Flowchart of the symbol detection method based on element regularization according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0026] S1: collect the first channel output symbol of the symbol to be tested;

[0027] In an embodiment of the present invention, the symbol to be tested is the received symbol to be detected, and the detected received symbol needs to be processed to obtain the input symbol corresponding to the received symbol to be detected; that is, the received symbol to be detected, that is, the first channel output symbol, is processed, and the corresponding first channel input symbol is detected by estimating its first channel information.

[0028] S2: Inputting the first channel output symbol of the symbol to be tested and the first channel information into the trained self-supervised detection neural network model, and outputting the first channel input symbol of the symbol to be tested.

[0029] In an embodiment of the present invention, considering that existing symbol detection methods rely on channel state information, it is necessary to detect instantaneous parameters, which will reduce the data transmission rate, and inaccurate parameter estimation will reduce detection performance. In addition, the data distribution of the support set and query set of the symbol detection algorithm based on meta-learning is inconsistent, so the model learned on the support set cannot be generalized well to the query set, which is not conducive to obtaining better symbol error rate performance. Therefore, the present invention considers a consistent meta-regularization method to optimize the self-supervised detection neural network model, thereby helping the meta-learning model learn how to reduce the data distribution difference between the support set and query set data in the meta-learning task. In this way, the models of the support set and query set are aligned to overcome this problem, enhance the ability of meta-knowledge to maintain the consistency of support and query data, and improve the performance of the meta-learning model.

[0030] In order to better illustrate the self-supervised detection neural network model of the present invention, the present invention describes the self-supervised detection neural network model and its corresponding training process in detail.

[0031] In the embodiment of the present invention, Figure 2 As shown, a self-supervised detection neural network model is constructed. The self-supervised detection neural network model includes N layers of perceptual linear layers and one softmax layer. The first to N-1 layers of perceptual linear layers use ReLu activation function, and the Nth layer of perceptual linear layer uses SGD optimizer. In this embodiment, a multi-layer perceptron is used. It consists of 6 linear layers. The first to fifth layers use ReLu as the activation function. The last layer does not use an activation function but uses the SGD optimizer. The first layer of the perceptual linear layer inputs the channel output signal, and the sixth layer of the perceptual linear layer outputs the channel input symbol probability. After integration through the softmax layer, the channel output signal is reconstructed and estimated, thereby obtaining the corresponding channel input signal.

[0032] Figure 3 This is a flow chart of the self-supervised detection neural network training process according to an embodiment of the present invention. Figure 3 As shown, the training process of the self-supervised detection neural network model includes:

[0033] 101. Obtain a second channel input symbol, a second channel output symbol, and second channel information of a training symbol;

[0034] In the embodiment of the present invention, it is necessary to obtain relevant information of the training symbols, where the relevant information is the second channel input symbol, the second channel output symbol and the second channel information; the training symbols here are the training set data for training and optimizing the self-supervised detection neural network model; the self-supervised detection neural network model of the present invention adopts a meta-training method, so the training sample of meta-learning is a task, and there is a The meta-training set of training tasks is used to train the meta-learning model, that is,

[0035] 102. Divide the acquired training symbols into a support set and a query set, and input the corresponding second channel output symbols and second channel information into a multilayer perceptron, respectively, to estimate the predicted distribution and corresponding weight of the second channel input symbols;

[0036] In order to better train the meta-learning model, each task in the present invention contains a support set (or training set) and a query set (or test set), which can be expressed as The present invention considers that the detection of different channels is a different task, that is, when the relationship between the channel input symbol X and the channel output symbol Y changes, it is considered to be the detection of different tasks. That is, the ISI channel is modeled as frequency selective Rayleigh fading with exponential power delay (Exp-PDP), and random sampling is performed from the Exp-PDP channel model with γ = 2. L represents the number of taps, and the data used for training can be obtained.

[0037] Among them, the output y of the channel at time slot i is i It can be expressed as:

[0038]

[0039] z i The distribution at time slot i is The noise signal, ρ 2 Denotes the signal-to-noise ratio. Using BPSK modulator, that is, x = {-1, 1}.

[0040] The present invention transmits K symbol blocks on a limited memory channel. Suppose the transmitted symbol X i ∈X, channel output Y i ∈Y, where i is the time index and i∈[1,2,3,...,K]. There are l paths between the transmitter and the receiver, and symbols transmitted along the longer path take more time to arrive. Therefore, the channel output Y at time slot i is i Not only affected by the symbol X i The influence of the previously transmitted symbol X i-1 In a limited memory channel, i The conditional distribution of depends on the code element from i-L+1 to i, that is, X i-L+1:i Therefore, the relationship between Y and X can be expressed as:

[0041]

[0042] Specifically, a self-supervised detection neural network is used to predict input symbols. Given a dataset D1 = {X 1:P}, D2={Y 1:K} to recover the original codeword X, where X 1:P is the training symbol, Y 1:K Is the channel output. Predict a set of L consecutive input symbols X from a series of channel outputs i-L+1:i Assuming that the neural network can approximate any relationship between input and output, the channel output symbol Y i-L+1:i+L-1 As the input of the neural network to predict the channel input symbol X as the output of the neural network i-L+1:i .

[0043] Therefore, the present invention can use a multi-layer perceptron Use θ parameterization to estimate the channel input symbol x i-L+1 ,…,x i The distribution of each Input the softmax layer to estimate the distribution of the input symbol of a single channel. Predict the i-th symbol L times at different positions of the network output and integrate the estimated distribution. For the ensemble prediction distribution, predict X with different weights i , where the weights are proportional to the estimated strength of the corresponding tap.

[0044] 103. Calculate a cross entropy loss of the training symbol based on the obtained second channel input symbol and the cross entropy of the estimated predicted distribution of the second channel input symbol;

[0045] In the embodiment of the present invention, the channel input symbol X is used i and the ensemble prediction distribution The cross entropy loss between the training symbols is expressed as:

[0046]

[0047] Among them, X i represents the second channel input symbol of the i-th training symbol obtained, represents the predicted distribution of the second channel input symbol of the estimated i-th training symbol, CE represents the cross entropy function, and P represents the length of the training symbol.

[0048] 104. Calculate a self-supervision loss of the training symbol based on an estimated difference between the acquired second channel output symbol and the second channel output signal reconstructed based on the predicted distribution of the second channel input symbol;

[0049] In an embodiment of the present invention, the model is trained by self-supervisory loss, and the self-supervisory loss of the training symbol is expressed as:

[0050]

[0051] Among them, y i represents the obtained second channel output symbol, represents the second channel output signal reconstructed based on the predicted distribution of the second channel input symbol of the i-th training symbol, K represents the number of transmitted symbol blocks, that is, a total of K symbol blocks are transmitted.

[0052] The second channel output signal reconstructed based on the predicted distribution of the second channel input symbol is expressed as:

[0053]

[0054] in, is based on the channel model i The estimation function of =i represents an estimated symbol of the second channel input symbol of L consecutive taps from the i-L+1th tap to the i-th tap; represents the estimated second channel information of the l-th tap; Represents the estimated symbol of the second channel input symbol of the (i-l+1)th tap.

[0055] 105. Calculate the consistent meta-regularization loss of the training symbols based on the parameter difference between the training model using the support set and the training model using the query set;

[0056] In an embodiment of the present invention, a method for calculating a consistent meta-regularization loss of training symbols includes training a basic learner using a support set of the m-th channel task; training a new specific learner using a query set of the m-th channel task; performing meta-regularization using the Frobenius norm of the model parameter difference between the basic learner and the specific learner; and determining a consistent meta-regularization loss of the training symbols with the goal of minimizing the meta-loss and meta-regularization loss between the basic learner and the specific learner.

[0057] A consistent meta-regularization method is used to help the meta-learning model learn how to reduce the data distribution difference between the support set and query set data in the meta-learning task. In this way, the models of the support set and query set are aligned to overcome this problem, enhance the ability of meta-knowledge to maintain the consistency of support and query data, and improve the performance of the meta-learning model. The data distribution of the support set and query set is aligned by introducing a loss function. For the task First, we get the support set Base learner trained This is the same as the traditional meta-learning method. Then, with the task QuerySet in To train a new specific learning model

[0058] and The difference between can be regarded as an indicator to measure the difference in data distribution between the support set and the query set. and The Frobenius norm of the parameter difference is used as meta-regularization, and minimizing it helps eliminate data distribution differences to obtain better meta-knowledge. The meta-regularization in the two-layer gradient-based method can be defined as:

[0059]

[0060]

[0061]

[0062] in is the meta-loss, θ is the initialized cross-task weight parameter that the meta-learning algorithm wants to learn, and μ is the regularization parameter, ‖.‖ F is the Frobenius norm.

[0063] Each training task contains a support set and a queryset right The loss function of MAML optimization for internal gradient updates of tasks can be written as:

[0064] in is the meta-loss and θ is the initialized cross-task weight parameter that MAML wants to learn. The purpose is to obtain the task A learning model for a specific task. When encountering a new task When θ is θ, a task-specific predictor can be easily obtained from the initial θ via one (or a few) internal gradient steps.

[0065] Compared with MAML, the MAML-Reg formula of integrated element regularization is:

[0066]

[0067]

[0068]

[0069] in, represents the loss of yuan, and represents the weight parameter of the mth channel task, represents the support set of the m-th channel task, represents the query set of the mth channel task; represents the support set of the i-th channel task trained model parameters; Represents the query set using the mth channel task The trained model parameters, represents the number of channel tasks; θ represents the model parameters of the self-supervised detection neural network model; represents the Frobenius norm; α represents the step size, is the improved regularization coefficient, L represents the channel tap, tasknum represents the number of training tasks, represents the gradient loss of the support set loss; represents the gradient of the query set meta-loss.

[0070] The present invention improves the traditional regularization coefficient, utilizes the channel information contained in the channel tap, updates the meta-parameters according to this channel information, directly affects the meta-parameters in the meta-learning model, and makes the trained network model more consistent with the channel characteristics.

[0071] Based on the above calculation analysis, the consistent meta-regularization loss of the training symbol is expressed as:

[0072] in, Indicates the model parameters of the mth channel task trained using the support set; Represents the model parameters for training the i-th channel task using the query set; represents the number of channel tasks; θ represents the model parameters of the self-supervised detection neural network model; represents the Frobenius norm.

[0073] 106. Calculate an estimated loss of the training symbols based on the cross entropy loss, the self-supervision loss, and the consistent element regularization loss, and optimize the multilayer perceptron using the estimated loss until the maximum number of iterations or the accuracy requirement of the multilayer perceptron is reached, thereby terminating the training.

[0074] According to each loss in steps 103-105, the estimated loss of the training symbol is calculated, and the estimated loss of the training symbol is expressed as:

[0075]

[0076] Among them, loss represents the estimated loss of training symbols, loss x Represents the cross entropy loss of the training symbol, loss su Represents the self-supervised loss of the training symbol, loss F represents the consistent meta-regularization loss of the training symbol, β controls the importance parameter of the self-supervisory loss, is the improved regularization coefficient, L represents the channel tap, and tasknum represents the number of training tasks.

[0077] Figure 4 This is a data processing flow chart for self-supervised detection neural network training according to an embodiment of the present invention. Figure 4As shown, in an embodiment of the present invention, after obtaining the training symbols, some data preprocessing is performed on these training symbols, and the detection tasks of different channels are respectively regarded as a channel task. After the meta-learning tasks are divided, the training set data can be divided into a support set and a query set. The model is trained by the support set and the query set, and these data are meta-regularized. The meta-regularized data is sent to the self-supervised detection neural network for training, thereby updating the model parameters until all tasks are trained, that is, the original channel input symbols can be predicted. The embodiment of the present invention adopts the MAML framework to overcome the problem of model retraining and find meta-initialization parameters. Through these parameters, only a few update steps are needed for the model to adapt to the new channel environment. Specifically, in the adaptive step of the meta-learning algorithm, each task trains the model separately. When a new training task is input, only a small number of update steps are needed, and the model can be updated by meta-optimizing the learning meta-initialization parameters to adapt to the new task.

[0078] In a preferred embodiment of the present invention, the present embodiment may also use a validation set to evaluate the model, using var_snr as an evaluation indicator, and saving the model with the smallest var_snr value.

[0079] This paper introduces consistent meta-regularization into the meta-learning framework MAML to align the models trained on the support set and query set for each task, eliminating data distribution differences between them and improving the ability of meta-knowledge to maintain data consistency. The aligned tasks are then fed into the detection network training model. After the model is trained, new training tasks are fed into the network, and the model is updated through several meta-algorithm update steps to adapt to the new training environment. Symbols are estimated using the updated model, and the estimated values are judged to obtain an estimated bitstream. This paper improves the accuracy of symbol detection, achieves more accurate detection performance, and addresses the problems of traditional machine learning methods, such as wasted channel resources and repeated training.

[0080] It is understandable that the terms "first" and "second" in the patent are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one such feature.

[0081] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A symbol detection method based on meta-regularization, characterized in that: The method includes collecting a first channel output symbol of a symbol to be tested; inputting the first channel output symbol of the symbol to be tested into a trained self-supervised detection neural network model, and outputting the first channel input symbol of the symbol to be tested; wherein the training process of the self-supervised detection neural network model includes: Obtain a second channel input symbol, a second channel output symbol, and second channel information of a training symbol; The obtained training symbols are divided into a support set and a query set, and the corresponding second channel output symbols and second channel information are respectively input into a multilayer perceptron to estimate the predicted distribution and corresponding weight of the second channel input symbols; Calculating a cross entropy loss of a training symbol based on a cross entropy of the obtained second channel input symbol and the estimated predicted distribution of the second channel input symbol; Calculating a self-supervisory loss of the training symbol based on an estimated difference between the acquired second channel output symbol and a second channel output signal reconstructed based on a predicted distribution of the second channel input symbol; Calculate the consistent meta-regularization loss of the training symbols based on the difference in parameters between training the model on the support set and training the model on the query set; Based on the cross entropy loss, self-supervision loss and consistent element regularization loss, the estimated loss of the training symbols is calculated, and the multilayer perceptron is optimized by the estimated loss until the maximum number of iterations or the accuracy requirement of the multilayer perceptron is reached, and the training is terminated.

2. The symbol detection method based on element regularization according to claim 1, characterized in that: The self-supervised detection neural network model includes N perceptual linear layers and one softmax layer. The first to N-1 perceptual linear layers use ReLu activation functions, and the Nth perceptual linear layer uses an SGD optimizer.

3. The symbol detection method based on element regularization according to claim 1, characterized in that: The estimated loss of the training symbol is expressed as: Among them, loss represents the estimated loss of training symbols, loss x Represents the cross entropy loss of the training symbol, loss su Represents the self-supervised loss of the training symbol, loss F represents the consistent meta-regularization loss of the training symbol, β controls the importance parameter of the self-supervisory loss, is the improved regularization coefficient, L represents the channel tap, and tasknum represents the number of training tasks.

4. The symbol detection method based on meta-regularization according to claim 3, characterized in that: The cross entropy loss of the training symbols is expressed as: Among them, X i represents the second channel input symbol of the i-th training symbol obtained, represents the predicted distribution of the second channel input symbol of the estimated i-th training symbol, CE represents the cross entropy function, and P represents the length of the training symbol.

5. The symbol detection method based on element regularization according to claim 3, characterized in that: The self-supervised loss of the training symbol is expressed as: Among them, y i represents the second channel output symbol of the i-th training symbol obtained, represents the second channel output signal reconstructed based on the predicted distribution of the second channel input symbol of the i-th training symbol, and K represents the number of transmitted symbol blocks.

6. The symbol detection method based on meta-regularization according to claim 5, characterized in that: The second channel output signal reconstructed based on the predicted distribution of the second channel input symbol is expressed as: in, is based on the channel model i The estimation function of =i represents an estimated symbol of the second channel input symbol of L consecutive taps from the i-L+1th tap to the i-th tap; represents the estimated second channel information of the l-th tap; Represents the estimated symbol of the second channel input symbol of the (i-l+1)th tap.

7. The symbol detection method based on meta-regularization according to claim 1 or 3, characterized in that: The calculation method of the consistent meta-regularization loss of training symbols includes training the basic learner with the support set of the m-th channel task; training the new specific learner with the query set of the m-th channel task; using the Frobenius norm of the difference in model parameters between the basic learner and the specific learner for meta-regularization; and determining the consistent meta-regularization loss of training symbols with the goal of minimizing the meta-loss and meta-regularization loss between the basic learner and the specific learner.

8. The symbol detection method based on element regularization according to claim 7, characterized in that: The loss objective function to be minimized is expressed as: C1: C2: in, represents the loss of yuan, and represents the weight parameter of the i-th channel task, represents the support set of the m-th channel task, represents the query set of the mth channel task; represents the support set of the mth channel task trained model parameters; Represents the query set using the mth channel task The trained model parameters, represents the number of channel tasks; θ represents the model parameters of the self-supervised detection neural network model; represents the Frobenius norm; α represents the step size, is the improved regularization coefficient, L represents the channel tap, tasknum represents the number of training tasks, represents the gradient loss of the support set loss; represents the gradient of the query set meta-loss.

9. The symbol detection method based on element regularization according to claim 7, characterized in that: The consistent meta-regularization loss of the training symbols is expressed as: in, Indicates the model parameters of the mth channel task trained using the support set; Indicates the model parameters for training the mth channel task using the query set; represents the number of channel tasks; θ represents the model parameters of the self-supervised detection neural network model; represents the Frobenius norm.

Images