Modulation signal identification method

Abstract

The invention relates to the technical field of wireless communication, and relates to a modulation signal identification method, which comprises the following steps: acquiring training set data, verification set data and test set data, and acquiring a modulation signal identification model with initialized parameters; training a modulation signal identification model based on the training set data until the loss function of the modulation signal identification model is converged or the training frequency reaches a threshold value; calculating the prediction accuracy of the modulation signal identification model in the training process based on the verification set data, and selecting a model parameter with the highest prediction accuracy as a training result after training is completed; testing the prediction accuracy of the trained modulation signal identification model based on the test set data, and continuing the training process of the training model when the prediction accuracy exceeds a threshold value; and performing modulation signal identification by using the modulation signal identification model. The method provided by the embodiment of the invention can accurately extract effective features, and the accuracy of identifying the modulation category of the signal is relatively high.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of wireless communication technology, and in particular, to a modulation signal recognition method. Background Art

[0002] In the process of wireless communication technology, communication modes can be divided into cooperative communication and non-cooperative communication. In the non-cooperative communication mode, the receiver is unknown to the modulation mode, symbol rate and other parameters of the transmitter's signal, and it is necessary to use parameter estimation methods to achieve the acquisition and demodulation of the transmitter signal, specifically to extract features from the signal received by the receiver and further classify or regress the signal to obtain various information of the transmitted signal. Deep neural networks can automatically extract the potential features of signals from massive data, improve classification accuracy and recognition efficiency, and introduce deep neural network technology into the modulation signal recognition process. Compared with the traditional expert experience recognition method, the recognition accuracy and efficiency are higher and easier to apply in practice. However, the existing convolutional neural network lacks interpretability, requires a cumbersome manual design process, feature extraction and classification are not flexible enough, and the manual design process has a large workload, which cannot meet the modulation signal recognition needs well. Summary of the invention

[0003] The embodiments of the present specification provide a modulation signal identification method, aiming to solve one or more of the above problems and other potential problems.

[0004] In order to achieve the above objectives, the following technical solutions are provided:

[0005] The present invention provides a method for identifying a modulated signal, the method comprising:

[0006] Acquire training set data, validation set data and test set data, and acquire a modulation signal recognition model with parameter initialization; wherein the training set data and the test set data include IQ modulation signals modulated in different modulation modes and transmitted in a communication system, and the IQ modulation signals are respectively provided with modulation category labels;

[0007] Training the modulation signal recognition model based on the training set data until the loss function of the modulation signal recognition model converges or the number of training times reaches a threshold, the input of the modulation signal recognition model is the IQ modulation signal, and the output is a predicted modulation category label;

[0008] Among them, the modulation signal recognition model is constructed based on a deep adaptive wavelet network, and the modulation signal recognition model includes a feature optimization module and a classification module. The feature optimization module performs wavelet decomposition and pooling processing on the IQ modulation signal respectively to obtain time domain features and frequency domain features of different decomposition levels and different pooling methods, and uses a self-attention algorithm to learn the importance of each feature to obtain the attention weight of each feature; the classification module includes K neurons, K is the number of modulation mode categories, and the activation function in the classification module converts the attention weight into a predicted probability label, and the predicted probability label with the highest predicted probability value is the predicted modulation category label;

[0009] Calculating the loss function value of the modulation signal recognition model during the training process based on the validation set data, and selecting the model parameters with the lowest loss function value as the training result after the training is completed;

[0010] Testing the loss function value of the modulated signal recognition model after training based on the test set data, and continuing the training model training process when the loss function value exceeds a threshold;

[0011] The modulation signal recognition model is used to perform modulation signal recognition.

[0012] The modulation signal recognition method of the embodiments of the present specification, the modulation signal recognition model combines wavelet transform with deep neural network, which is beneficial to increase the accuracy of the feature extraction process. Compared with the general wavelet transform process, it is more flexible and accurate, and the classification performance of the entire model is better and more robust. Specifically, wavelet decomposition and pooling processing are used to optimize the extracted signal features to obtain multi-level feature vectors, and then the importance of each feature in the feature vector is learned through the self-attention algorithm. Classification is completed based on different importances, and efficient extraction and classification of signal features are achieved through multi-level decomposition and feature focusing, with higher classification accuracy.

[0013] In some embodiments, the Adam optimization algorithm is used to update model parameters when training the modulation signal recognition model based on the training set data.

[0014] In some embodiments, the modulation signal recognition model also includes a feature extraction module, which extracts features from the IQ modulation signal to obtain a feature matrix, inputs the feature matrix into the feature optimization module, and outputs the attention weight of each feature.

[0015] In some embodiments, the feature extraction module includes a two-dimensional convolution layer and a convolution layer, the two-dimensional convolution layer fuses the I-channel signal and the Q-channel signal of the IQ modulated signal and inputs the information into the convolution layer, and the convolution layer outputs the feature matrix.

[0016] In some embodiments, the two-dimensional convolution layer and the convolution layer are both provided with a batch normalization layer and a LeakyReLU layer.

[0017] In some embodiments, the feature optimization module includes several decomposition layers, two average pooling layers, a maximum pooling layer and a self-attention network; one of the average pooling layers and the maximum pooling layer processes the feature matrix respectively to obtain an average eigenvector and a maximum eigenvector; several of the decomposition layers perform several wavelet decompositions on the feature matrix respectively, and except for the first wavelet decomposition, the input of each subsequent wavelet decomposition is the low-frequency component of the output of the previous wavelet decomposition, and the output of the last wavelet decomposition is average pooled by another average pooling layer, and then connected in series after average pooling to obtain the decomposed eigenvector; the average eigenvector, the maximum eigenvector and the decomposed eigenvector are connected in series to obtain a eigenvector; the self-attention network is used to learn the importance of each feature in the eigenvector, and the attention weight of each feature is output.

[0018] In some embodiments, the self-attention network includes two fully connected layers, the feature vector is sequentially input into the two fully connected layers, and the attention weight is output, and the attention weight is as follows:

[0019] w A =Sigmod(W M2 ReLU(W M1 f))

[0020] Among them, w A One-dimensional attention weight, W M1 and W M2 They are the weight matrices of the first fully connected layer and the second fully connected layer respectively. The activation function of the first fully connected layer is ReLU(), and the activation function of the second fully connected layer is Sigmod().

[0021] In some embodiments, the classification module includes two fully connected layers, the activation function of the first fully connected layer is ReLU(), the activation function of the second fully connected layer is Sigmod(), the first fully connected layer combines various features, and the second fully connected layer outputs a predicted probability label through the activation function Sigmod().

[0022] In some embodiments, when the modulation signal recognition model is trained based on the training set data, the training process is divided into two training stages, and the loss functions of the two training stages are respectively a first loss function and a second loss function. When the first loss function converges, the first training stage ends and enters the second training stage. When the second loss function converges, the training of the modulation signal recognition model ends;

[0023] The first loss function formula is as follows:

[0024] Loss 1 =L cross +ρ 1 L wave

[0025] Among them, Loss 1 represents the first loss function, L cross is the label cross entropy loss for predicting modulation class labels and modulation class labels, L wave is the wavelet decomposition error, ρ 1 To adjust L cross Hyperparameters of weights;

[0026] The second loss function formula is as follows:

[0027] Loss 2 =Loss 1 +ρ 2 L center

[0028]

[0029] Among them, Loss 2 Represents the second loss function, Loss 1 represents the first loss function, L center represents the eigenvector center constraint, f n represents the nth eigenvector, c yn represents the eigenvector center, φ represents the constraint coefficient, ρ 2 Indicates adjustment L center Hyperparameters of weights.

[0030] In some embodiments, the eigenvector center constraint L is updated using an exponentially weighted moving average method. center The eigenvector center c in yn . BRIEF DESCRIPTION OF THE DRAWINGS

[0031] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the embodiments of the present specification will become readily understood. In the accompanying drawings, several embodiments of the present specification are shown by way of example and not limitation.

[0032] Figure 1 A flow chart of a modulation signal identification method in an embodiment of this specification is shown;

[0033] Figure 2 A comparison chart of the prediction accuracy of the modulation signal recognition model of the embodiment of this specification with other models under the 2016.10a data set is shown;

[0034] Figure 3 A comparison chart of the prediction accuracy of the modulation signal recognition model of the embodiment of this specification with other models under the 2016.10b data set is shown;

[0035] Figure 4 A comparison chart of the prediction accuracy of the modulation signal recognition model of the embodiment of this specification with other models under the 2018.01a data set is shown. DETAILED DESCRIPTION

[0036] The preferred embodiments of the present specification will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present specification are shown in the accompanying drawings, it should be understood that the present specification can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present specification more thorough and complete, and to fully convey the scope of the present specification to those skilled in the art.

[0037] The term "including" and its variations used in this document represent open inclusion, that is, "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example embodiment" and "an embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one other embodiment". Terms such as "upper", "lower", "front", and "rear" indicating placement or positional relationships are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the principles of this specification, and do not indicate or imply that the referred elements must have a specific orientation, be constructed or operate in a specific orientation, and therefore should not be understood as limitations on this specification.

[0038] A modulation signal recognition method of an embodiment of the present specification is described in detail below in conjunction with the accompanying drawings. The modulation signal recognition method of the present specification includes:

[0039] S1. Obtain training set data, validation set data and test set data, and obtain a modulation signal recognition model with parameter initialization; wherein the training set data and the test set data include IQ modulated signals modulated in different modulation modes and transmitted in a communication system, and the IQ modulated signals are respectively provided with modulation category labels;

[0040] S3. Train the modulation signal recognition model based on the training set data until the loss function of the modulation signal recognition model converges or the number of training times reaches a threshold, the input of the modulation signal recognition model is the IQ modulation signal, and the output is the predicted modulation category label; wherein the modulation signal recognition model is constructed based on a deep adaptive wavelet network, the modulation signal recognition model includes a feature optimization module and a classification module, the feature optimization module performs wavelet decomposition and pooling processing on the IQ modulation signal respectively, obtains time domain features and frequency domain features of different decomposition levels and different pooling methods, uses a self-attention algorithm to learn the importance of each feature, and obtains the attention weight of each feature; the classification module includes K neurons, K is the number of modulation mode categories, and the activation function in the classification module converts the attention weight into a predicted probability label, and the predicted probability label with the highest predicted probability value is the predicted modulation category label;

[0041] S5, calculating the loss function value of the modulation signal recognition model during the training process based on the validation set data, and after the training is completed, selecting the model parameter with the lowest loss function value as the training result;

[0042] S7, testing the loss function value of the modulated signal recognition model after training based on the test set data, and continuing the above training model training process when the loss function value exceeds a threshold;

[0043] S9. Use the modulation signal recognition model to perform modulation signal recognition.

[0044] In step S1, three public data sets generated by simulation using the open source software radio GNU Radio are used for model training, testing and verification. The data sets are 2016.10a, 2016.10b and 2018.01a. These data sets simulate various wireless channel conditions such as sampling rate offset, frequency offset, multipath fading, additive Gaussian white noise, etc. according to the actual environment. For example, the 2016.10a data set contains 11 commonly used modulation types of signals, including BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK, Gaussian frequency shift keying, 4PAM, 8 digital modulation methods, and 3 analog signal modulation methods, namely broadband frequency modulation, double-sideband amplitude modulation, and single-sideband amplitude modulation. The 2016.10a dataset contains a total of 220,000 signals, and the 2016.10b dataset has a relatively larger number of 1.2 million signals. The 2016.10b dataset contains 10 modulation types of signals except AM-SSB. Each signal consists of an in-phase part and an orthogonal part, stored as a 2×128 matrix. The signal-to-noise ratio ranges from -20dB to 18dB, with an interval of 2dB. The 2018.01a dataset contains 24 modulation types of signals, more than 200 million signal samples, each signal matrix size is 2×1024, the signal-to-noise ratio ranges from -20dB to 30dB, with an interval of 2dB. The signal data of different modulation types and different signal-to-noise ratios in the three public data sets are layered sampled. Specifically, the received signal of the communication system is sampled by randomly selecting time periods through slicing and sliding windows to obtain the training set data, the validation set data and the test set data. The ratio of the training set data, the validation set data and the test set data is 6:2:2.

[0045] In step S3, the blind reception process of the communication signal mainly covers two tasks: feature extraction and classification and recognition. Therefore, the modulation signal model mainly includes a feature extraction module, a feature optimization module and a classification module.

[0046] The feature extraction module is used to extract features from the IQ modulated signal to obtain a feature matrix, and specifically includes a two-dimensional convolution layer and a convolution layer. First, the IQ modulated signal containing two mutually orthogonal channels is converted into a tensor R, where the tensor R represents a signal data with a length of R and a channel number of 1; the two-dimensional convolution layer fuses the I-channel signal and the Q-channel signal in the tensor R, on the one hand converting the signal into features, and on the other hand avoiding the loss of relevant information between the I-channel signal and the Q-channel signal caused by directly processing the input signal in one dimension, and inputs the features after information fusion into the convolution layer, and further outputs a higher-level feature matrix while keeping the feature size unchanged. The two-dimensional convolution layer and the convolution layer are both provided with a batch normalization layer and a LeakyReLU layer to avoid gradient disappearance and improve the robustness of the model, and the slope of the LeakyReLU layer is set to 0.01.

[0047] The feature optimization module includes several decomposition layers, two average pooling layers, a maximum pooling layer and a self-attention network; one average pooling layer and the maximum pooling layer process the feature matrix respectively to obtain the average feature vector and the maximum feature vector; several decomposition layers perform several wavelet decompositions on the feature matrix respectively, and except for the first wavelet decomposition, the input of each subsequent wavelet decomposition is the low-frequency component of the output of the previous wavelet decomposition, and the output of the last wavelet decomposition is average pooled by another average pooling layer, and then the average pooling is combined in series to obtain the decomposed feature vector. That is, the decomposition layer performs one-dimensional wavelet decomposition on the vectors on each channel of the feature matrix, and then obtains the corresponding high-frequency components and low-frequency components, and average pools the output of the last wavelet decomposition along the channel dimension and then combines in series to obtain the decomposed feature vector. The average feature vector, the maximum feature vector and the decomposed feature vector are serially combined to obtain a feature vector, which contains various features obtained at different decomposition levels and different pooling methods. In order to focus on more important features, the self-attention network is used to learn the importance of each feature in the feature vector and output the attention weight of each feature. Specifically, the self-attention network includes two fully connected layers, and the feature vector is sequentially input into the two fully connected layers to output the attention weight. The attention weight is as follows:

[0048] w A =Sigmod(W M2 ReLU(W M1 f))

[0049] Among them, w A One-dimensional attention weight, W M1 and W M2They are the weight matrices of the first fully connected layer and the second fully connected layer respectively, the activation function of the first fully connected layer is ReLU(), the activation function of the second fully connected layer is Sigmod(), the output length of the first fully connected layer is 1 / 4 of the input, and the output length of the second fully connected layer is the same as the length of the feature vector.

[0050] The attention weight is multiplied by the feature vector to obtain a weighted feature vector for the subsequent classification process: W =w A f, where f W represents the weighted eigenvector, w A represents the one-dimensional attention weight, and f represents the feature vector.

[0051] The classification module includes two fully connected layers, the activation function of the first fully connected layer is ReLU(), and the activation function of the second fully connected layer is Sigmod(). The first fully connected layer is used to combine f W The second fully connected layer contains K neurons, where K is the number of signal modulation categories contained in the data set, and the activation function Sigmod() outputs the predicted probability label. The predicted probability label with the highest predicted probability value is the predicted modulation category label, and the modulation category corresponding to the subscript of the predicted modulation category label is the predicted modulation category of the signal.

[0052] In step S3, when training the modulation signal recognition model based on the training set data, the Adam optimization algorithm is used to update the model parameters, the initial learning rate of the training is set to 0.001, and the training set data is randomly divided into multiple small batch training sets for training. When the Adam optimization algorithm updates the model parameters, it can adaptively adjust the learning rate according to the gradient characteristics of different parameters, combining the advantages of momentum and adaptive learning rate, accelerating the convergence process and improving training efficiency.

[0053] In the early stage of model training, the parameters in the network change rapidly, and the derived feature vectors are not stable and effective. Therefore, the training of the modulation signal recognition model is divided into two training stages. The first E data sets of the training are traversed as the first training stage, and the purpose is to conduct preliminary training on the model network parameters so that they converge roughly to the final value. Specifically, the first loss function of the first training stage includes label cross entropy loss and wavelet decomposition error. In order to gradually adjust the neural network parameters of the wavelet decomposition function during the training process to achieve the ideal effect, it is necessary to add the error of the wavelet decomposition result for consideration. The formula of the first loss function is as follows:

[0054] Loss 1 =L cross +ρ1 L wave

[0055] Among them, Loss 1 represents the first loss function, L cross is the label cross entropy loss for predicting modulation class labels and modulation class labels, L wave is the wavelet decomposition error, ρ 1 To adjust L cross Hyperparameters of weights;

[0056]

[0057] Among them, C represents the number of channels, t represents the tth wavelet decomposition, T represents the number of wavelet decompositions, and H represents the channel dimension along the high-frequency component matrix. Represents the channel dimension of the matrix along the low-frequency components.

[0058] When the first loss function converges, the first training phase ends and enters the second training phase. In order to improve the classification accuracy, the feature vector should not only have the separability between categories, but also need to improve the compactness within each category. Therefore, in the second training phase, the feature vector center constraint is added to the first loss function to obtain the second loss function, so as to achieve the aggregation of feature vectors of the same category and the separation of feature vectors of different categories, so that the feature vectors can move closer to the corresponding center. When the second loss function converges, the training of the modulation signal recognition model ends.

[0059] The second loss function formula is as follows:

[0060] Loss 2 =Loss 1 +ρ 2 L center

[0061]

[0062] Among them, Loss 2 Represents the second loss function, Loss 1 represents the first loss function, L center represents the eigenvector center constraint, f n represents the nth eigenvector, c yn represents the center of the eigenvector, φ represents the constraint coefficient, which is taken as 0.05 in the embodiment of this specification, and ρ 2 Indicates adjustment L center Specifically, the exponentially weighted moving average method is used to update the feature vector center constraint L center The eigenvector center c in yn .

[0063] In step S5, the prediction accuracy is represented by the second loss function value. The smaller the second loss function value is, the higher the prediction accuracy is. Starting from the second training stage, when the second loss function value calculated by the verification set data does not decrease after 3 traversals, the learning rate is reduced by half. If the second loss function value does not decrease after 10 consecutive traversals, the training is stopped. The model parameters with the lowest second loss function value during training, that is, the highest prediction accuracy, are used as the training results. In step S7, the second loss function value calculated by the test set data also represents the model prediction accuracy. The above test process is calculated in the Pytorch1.12 environment. As Figure 2 , Figure 3 , Figure 4 As shown, there are comparison diagrams of the prediction accuracy of the model that does not distinguish the signal modulation category and the prediction accuracy of the modulation signal recognition model in the embodiment of this specification under different data sets. It can be seen from the right figure that the signal prediction accuracy of the modulation signal recognition model of this specification is higher.

[0064] Although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of this specification. Certain features described in the context of separate embodiments may also be implemented in a single implementation in combination. Conversely, various features described in the context of a single implementation may also be implemented in multiple implementations individually or in any suitable sub-combination.

[0065] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.

[0066] The embodiments of the present specification have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A modulation signal recognition method, characterized in that: Methods include: Acquire training set data, validation set data and test set data, and acquire a modulation signal recognition model with parameter initialization; wherein the training set data and the test set data include IQ modulation signals modulated in different modulation modes and transmitted in a communication system, and the IQ modulation signals are respectively provided with modulation category labels; Training the modulation signal recognition model based on the training set data until the loss function of the modulation signal recognition model converges or the number of training times reaches a threshold, the input of the modulation signal recognition model is the IQ modulation signal, and the output is a predicted modulation category label; The modulation signal recognition model is constructed based on a deep adaptive wavelet network. The modulation signal recognition model includes a feature optimization module and a classification module. The feature optimization module performs wavelet decomposition and pooling processing on the IQ modulation signal to obtain time domain features and frequency domain features of different decomposition levels and different pooling methods, and uses a self-attention algorithm to learn the importance of each feature to obtain the attention weight of each feature. The classification module includes K neurons. K is the number of modulation categories, and the activation function in the classification module converts the attention weight into a predicted probability label, and the predicted probability label with the highest predicted probability value is the predicted modulation category label; Calculate the loss function value of the modulation signal recognition model during the training process based on the validation set data, and select the model parameters with the lowest loss function value as the training result after the training is completed; test the loss function value of the modulation signal recognition model after training based on the test set data, and continue the above-mentioned training model training process when the loss function value exceeds a threshold; The modulation signal recognition model is used to perform modulation signal recognition.

2. The method according to claim 1, characterized in that When training the modulation signal recognition model based on the training set data, the Adam optimization algorithm is used to update the model parameters.

3. The method according to claim 2, characterized in that The modulation signal recognition model also includes a feature extraction module, which extracts features from the IQ modulation signal to obtain a feature matrix, inputs the feature matrix into the feature optimization module, and outputs the attention weight of each feature.

4. The method according to claim 3, characterized in that The feature extraction module includes a two-dimensional convolution layer and a convolution layer. The two-dimensional convolution layer fuses the I-channel signal and the Q-channel signal of the IQ modulated signal and inputs the information into the convolution layer. The convolution layer outputs the feature matrix.

5. The method according to claim 4, characterized in that The two-dimensional convolution layer and the convolution layer are both provided with a batch normalization layer and a LeakyReLU layer.

6. The method according to claim 3, characterized in that: The feature optimization module includes several decomposition layers, two average pooling layers, a maximum pooling layer and a self-attention network; one of the average pooling layers and the maximum pooling layer processes the feature matrix respectively to obtain an average feature vector and a maximum feature vector; several of the decomposition layers perform several wavelet decompositions on the feature matrix respectively, and except for the first wavelet decomposition, the input of each subsequent wavelet decomposition is the low-frequency component of the output of the previous wavelet decomposition, and the output of the last wavelet decomposition is average pooled by another average pooling layer, and then connected in series after average pooling to obtain the decomposed feature vector; the average feature vector, the maximum feature vector and the decomposed feature vector are connected in series to obtain a feature vector; the self-attention network is used to learn the importance of each feature in the feature vector, and the attention weight of each feature is output.

7. The method according to claim 6, characterized in that The self-attention network includes two fully connected layers, the feature vector is sequentially input into the two fully connected layers, and the attention weight is output, and the attention weight is as follows: In A =Sigmod(W M2 ReLU(W M1 f)) Among them, w A One-dimensional attention weight, W M1 and W M2 They are the weight matrices of the first fully connected layer and the second fully connected layer respectively. The activation function of the first fully connected layer is ReLU(), and the activation function of the second fully connected layer is Sigmod().

8. The method according to claim 1, characterized in that The classification module includes two fully connected layers, the activation function of the first fully connected layer is ReLU(), the activation function of the second fully connected layer is Sigmod(), the first fully connected layer combines various features, and the second fully connected layer outputs a predicted probability label through the activation function Sigmod().

9. The method according to claim 3, characterized in that: When the modulation signal recognition model is trained based on the training set data, the training process is divided into two training stages, and the loss functions of the two training stages are respectively a first loss function and a second loss function. When the first loss function converges, the first training stage ends and enters the second training stage. When the second loss function converges, the training of the modulation signal recognition model ends; The first loss function formula is as follows: Loss1=L cross +ρ1L wave Among them, Loss1 represents the first loss function, L cross is the label cross entropy loss for predicting modulation class labels and modulation class labels, L wave is the wavelet decomposition error, ρ1 is the adjustment L cross Hyperparameters of weights; The second loss function formula is as follows: Loss2=Loss1+ρ2L center Among them, Loss2 represents the second loss function, Loss1 represents the first loss function, and L center represents the eigenvector center constraint, f n represents the nth eigenvector, c yn represents the center of the eigenvector, φ represents the constraint coefficient, and ρ2 represents the adjustment L center Hyperparameters of weights.

10. The method according to claim 9, characterized in that Update the eigenvector center constraint L using the exponentially weighted moving average method center The eigenvector center c in yn .

Images