An incremental 5G antenna identification method based on synthetic feature replay

Abstract

The application discloses a kind of based on synthetic feature replay incremental 5G antenna identification method, comprising:1 using the 5G antenna data set of basic category to train basic classification network and variational autoencoder network;2 the feature vector extracted using feature module in basic classification network and the reconstructed feature vector generated by variational autoencoder network estimate the Gaussian distribution of synthetic feature, and store in memory bank;3 extract arbitrary category Gaussian distribution from memory bank, generate feature vector, and splice with the feature vector of new data set, to train basic classification network, using the feature vector of new data set to train variational autoencoder network;4 adopt the optimization method of combining multiscale gradient blur and momentum to update basic classification model, use gradient descent method to update variational autoencoder model.The application can effectively reduce memory cost, and relieve catastrophic forgetting and overfitting problem, so that 5G antenna classification model can be continuously learned and updated in dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of incremental learning, specifically an incremental 5G antenna identification method based on synthetic feature replay. Background Technology

[0002] With the rapid development of 5G communication technology, testing the performance of 5G antennas has become particularly crucial. In a robotic arm-based 5G antenna near-field detection system, it is necessary to identify the model of the 5G antenna to determine its operating frequency and size, thereby setting the size and height of the sampling surface. However, when a new 5G antenna dataset is added, the target recognition model needs to be retrained. Therefore, incremental learning technology is applied, so that when a new dataset is added, only the new dataset needs to be trained, which greatly saves time.

[0003] A key challenge in incremental learning is addressing the issues of learning new knowledge and catastrophic forgetting of old knowledge as data dynamically arrives at each learning stage. Typical incremental learning methods employ a hybrid training approach, combining samples from the old dataset with new data samples, to mitigate the catastrophic forgetting problem. However, this method involves storing both old sample data and old model parameters, requiring significant memory investment and potentially leading to data privacy breaches, thus presenting certain limitations. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing an incremental 5G antenna identification method based on synthetic feature replay. This method aims to effectively reduce memory costs and mitigate catastrophic forgetting and overfitting issues. It enables continuous learning and updating of the 5G antenna classification model in dynamic environments, thereby improving identification efficiency and accuracy while avoiding data privacy leaks.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The incremental 5G antenna identification method based on synthetic feature replay of the present invention is characterized by the following steps:

[0007] S1: Obtain the 5G antenna dataset for the basic categories, denoted as... ,in, express The Middle The first basic category One sample, express Category tags, express The number of basic categories included. Indicates the first The number of samples in each basic category;

[0008] S2: Construct a basic classification network and a variational autoencoder network;

[0009] S3: Training the basic classification network and variational autoencoder network:

[0010] S31: The feature extraction module in the basic classification network... Processing is performed to obtain eigenvectors ; After being input into the classification module of the basic classification network for recognition, it is then processed by the softmax layer to obtain... Category probability distribution ;

[0011] The training loss function of the basic classification network is established using equation (1). :

[0012] (1)

[0013] S32: Variational autoencoder network pair Processing is performed to obtain reconstructed features. Therefore, the loss function of the variational autoencoder network can be established using equation (2). :

[0014] (2)

[0015] In equation (2), Indicates divergence loss. Indicates the reconstruction loss. Indicates hyperparameters;

[0016] S33: Using equations (3) and (4), we obtain the first... Mean of the synthetic feature vectors of each category and the The synthesized feature vectors of the n categories in the nth category variance in each dimension Thus, the first Variance of the synthetic feature vectors of each category s c 2 = [ s c , 1 2 , s c , 2 2 , . . . , s c , d 2 , . . . , s c , D 2 ] And thus obtain the first Gaussian distribution of synthesized feature vectors for each category And store it in the memory bank:

[0017] (3)

[0018] (4)

[0019] S34: Use the gradient descent algorithm to optimize all parameters in the basic classification network and variational autoencoder network to obtain the trained basic classification model and variational autoencoder model;

[0020] S4: Train the base classification network for the kth iteration:

[0021] S41: Define the current iteration number as k, and initialize k=1;

[0022] The parameters of the trained base classification model are used as the base classification model for the (k-1)th iteration, and the parameters of the base classification model for the (k-1)th iteration are denoted as... ;

[0023] The parameters of the trained variational autoencoder model are used as the variational autoencoder model for the (k-1)th iteration, and the parameters of the variational autoencoder model for the (k-1)th iteration are denoted as... ;

[0024] S42: Collect the 5G antenna dataset of the incremental category in the k-th iteration, denoted as... ,in, express The Middle The first basic category One sample, express Category tags;

[0025] S43: Will The input is processed in the feature extraction module of the trained base classification model to obtain feature vectors. ;

[0026] S44: Randomly draw a Gaussian distribution of feature vectors of any class from the memory bank to generate the first feature vector of any class s. 1 synthesized feature vector and with By splicing, we obtain the first... splicing features Then, it is input into the basic classification model in the k-th iteration for processing, to obtain... probability distribution ;

[0027] S45: Construct the loss function of the base classification model for the kth iteration according to equation (5). :

[0028] (5)

[0029] In equation (5), Representative dataset The total number of categories;

[0030] S46: Update the parameters of the base classification model in the (k-1)th iteration using an optimization method combining multi-scale gradient fuzziness and momentum. Thus, the update strategy for the basic classification model in the k-th iteration is constructed;

[0031] S5: Training the variational autoencoder network for the k-th iteration:

[0032] S51: Will The input is processed in the encoder of the trained variational autoencoder model to obtain the reconstructed features of the k-th iteration. Therefore, the loss function of the variational autoencoder model in the kth iteration is established using equation (6). :

[0033] (6)

[0034] In equation (6), It is the divergence loss of the k-th iteration. It is the hyperparameter of the k-th iteration. It is the reconstruction loss of the k-th iteration;

[0035] S52: Update the parameters of the variational autoencoder model in the (k-1)th iteration using gradient descent. Thus, the update strategy of the variational autoencoder model in the k-th iteration is constructed;

[0036] S6: After assigning k+1 to k, return to S42 and execute sequentially until the maximum number of iterations is reached, thereby obtaining the optimal incremental classification model and its optimal parameters. And the optimal incremental variational autoencoder model and its optimal parameters ;

[0037] S7: Input any 5G antenna sample into the optimal incremental classification model for processing to obtain the predicted category label.

[0038] The incremental 5G antenna identification method based on synthetic feature replay described in this invention is also characterized in that the basic classification network in S2 includes: a feature extraction module and a classification module;

[0039] The variational autoencoder network includes: an encoder, a mean layer, a variance layer, a variational layer, and a decoder;

[0040] The feature extraction module and the encoder have the same network structure, and each consists of an input convolutional layer, Q downsampling convolutional layers, and W residual convolutional layers. Each convolutional layer is followed by an instance regularization layer and a ReLU activation function.

[0041] The classification module includes: a fully connected layer and a softmax layer.

[0042] S32 includes the following steps:

[0043] S321: Input into the encoder to obtain Another feature vector ;

[0044] S322: After processing through the mean layer and variance layer respectively, the corresponding feature vectors are obtained. mean and variance ;

[0045] S323: and After undergoing variable-level reparameterization, latent space samples are generated.

[0046] S324: The latent space samples are processed by the decoder to obtain reconstructed features. ;

[0047] S325: Establish using equations (7) and (8) respectively and :

[0048] (7)

[0049] (8)

[0050] In equation (7), express The Middle Variance in each dimension This represents the total number of dimensions. express The Middle The mean of each dimension.

[0051] S46 includes the following steps:

[0052] S461: Define and initialize the first-order momentum in the (k-1)th iteration. and second momentum ;

[0053] S462: Calculate the classification model loss gradient for the k-th iteration using equation (9). :

[0054] (9)

[0055] In equation (9), express exist gradient at;

[0056] S463: Using equation (10) Gradient fuzzing is performed to obtain the fuzzy gradient with scale s in the k-th iteration. :

[0057] (10)

[0058] In equation (10), This represents a Gaussian blur kernel with a scale of s;

[0059] S464: Use equation (11) to obtain the weighted sum of the k-th iteration. :

[0060] (11)

[0061] In equation (11), The weights representing a scale of s, Represents a nonlinear activation function; S denotes the maximum scale;

[0062] S465: The first-order momentum of the k-th iteration is obtained using equations (12) and (13). and second momentum :

[0063] (12)

[0064] (13)

[0065] In equations (12) and (13), and This indicates two attenuation factors;

[0066] S446: Use equation (14) to obtain the parameters of the base classification model in the k-th iteration. :

[0067] (14)

[0068] In equation (14), It's the learning rate. It is a parameter to prevent the denominator from being 0.

[0069] S51 includes the following steps:

[0070] S511: Will The input is processed in the encoder of the trained variational autoencoder model to obtain the feature vector. ; After further processing through the mean layer and variance layer, the feature vector is obtained. mean and variance ;

[0071] S512: and After reparameterization with varying hierarchies, the latent space sample for the k-th iteration is generated. This latent space sample is then input into the decoder for processing to obtain the reconstructed features for the k-th iteration. ;

[0072] S513: Using equations (15) and (16), establish the loss function of the variational autoencoder model for the k-th iteration. :

[0073] (15)

[0074] (16)

[0075] In equations (15) and (16), It is the kth iteration. The mean of the synthetic feature vectors of each category; It is the kth iteration. The synthesized feature vectors of the n categories in the nth category The variance in each dimension.

[0076] S52 includes the following steps:

[0077] S521: Calculate the loss gradient of the variational autoencoder model in the k-th iteration using equation (17). :

[0078] (17)

[0079] In equation (17), express exist gradient at;

[0080] S522: Calculate the parameters of the variational autoencoder model in the kth iteration using equation (18). :

[0081] (18)

[0082] In equation (18), This represents the learning rate.

[0083] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the incremental 5G antenna identification method, and the processor is configured to execute the program stored in the memory.

[0084] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the incremental 5G antenna identification method.

[0085] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0086] 1. This invention uses the feature vectors extracted by the feature module in the basic classification network and the reconstructed feature vectors generated by the variational autoencoder network to estimate the Gaussian distribution of the synthetic features, which does not require storing old data original samples, thus greatly reducing memory costs.

[0087] 2. This invention employs an optimization method combining multi-scale gradient fuzziness and momentum to update the parameters of the basic classification model, effectively mitigating catastrophic forgetting and improving the model's training stability and generalization ability.

[0088] 3. When identifying an ever-expanding dataset, this invention does not require retraining the entire dataset each time, nor does it require storing a large amount of historical data, which greatly improves the efficiency of model training and recognition. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the framework of the incremental target recognition method based on synthetic feature replay of the present invention;

[0090] Figure 2 This is a flowchart illustrating the incremental target recognition method based on synthetic feature replay of the present invention. Detailed Implementation

[0091] In this embodiment, as Figure 1 As shown, an incremental 5G antenna identification method framework based on synthetic feature replay includes: an encoder 1, a feature extraction module 2, a classifier 3, a decoder 4, a module for estimating the Gaussian distribution of synthetic features 5, and a memory bank 6. The encoder 1 and feature extraction module 2 have the same network structure and are used to extract sample features. The features extracted by the encoder 1 are input into the decoder 4 to generate reconstructed features. The features extracted by the feature extraction module 2 and the reconstructed features generated by the decoder 4 are input into the module for estimating the Gaussian distribution of synthetic features 5 to obtain the estimated Gaussian distribution of the synthetic features, which is then stored in the memory bank 6.

[0092] In this embodiment, an incremental 5G antenna identification method based on synthetic feature replay is described, such as... Figure 2As shown, it includes the following steps:

[0093] S1: Obtain the 5G antenna dataset for the basic categories, denoted as... ,in, express The Middle The first basic category One sample, express Category tags, express The number of basic categories included. Indicates the first The number of samples in each basic category;

[0094] S2: Construct a basic classification network and a variational autoencoder network;

[0095] The basic classification network includes: a feature extraction module and a classification module;

[0096] A variational autoencoder network consists of: an encoder, a mean layer, a variance layer, a variational layer, and a decoder;

[0097] The feature extraction module and the encoder have the same network structure, and each consists of an input convolutional layer, Q downsampling convolutional layers, and W residual convolutional layers. Each convolutional layer is followed by an instance regularization layer and a ReLU activation function.

[0098] The classification module includes: a fully connected layer and a softmax layer.

[0099] S3: Training the basic classification network and variational autoencoder network:

[0100] S31: The feature extraction module in the basic classification network... Processing is performed to obtain eigenvectors ; After being input into the classification module for recognition, it is then processed by the softmax layer to obtain... Category probability distribution ;

[0101] The training loss function of the basic classification network is established using equation (1). :

[0102] (1)

[0103] S32: Variational autoencoder network pair Processing:

[0104] S321: Input into the encoder, and get Another feature vector ;

[0105] S322: After processing through the mean layer and variance layer respectively, the corresponding feature vectors are obtained. mean and variance ;

[0106] S323: and After undergoing a variable-level reparameterization process, the latent space sample z is generated using equation (2):

[0107] (2)

[0108] In formula (2) Indicates from the normal distribution Medium sampling noise, .

[0109] S324: After the latent space samples are processed by the decoder, the reconstructed features are obtained using equation (3). :

[0110] (3)

[0111] In equation (3), This indicates the decoder.

[0112] S325: Establish using equations (4) and (5) respectively and :

[0113] (4)

[0114] (5)

[0115] In equation (4), express The Middle Variance in each dimension This represents the total number of dimensions. express The Middle The mean of each dimension.

[0116] S326: Thus, the loss function of the variational autoencoder network is established using equation (6). :

[0117] (6)

[0118] In equation (6), Indicates divergence loss. Indicates the reconstruction loss. This represents hyperparameters.

[0119] S33: Using equations (7) and (8), we obtain the first... Mean of the synthetic feature vectors of each category and the The feature vectors of the nth category are in the nth... variance in each dimension Thus, the first Variance of the synthesized feature vectors of each category s c 2 = [ s c , 1 2 , s c , 2 2 , . . . , s c , d 2 , . . . , s c , D 2 ] And thus obtain the first Gaussian distribution of synthesized feature vectors for each category And store it in the memory bank:

[0120] (7)

[0121] (8)

[0122] S34: Use the gradient descent algorithm to optimize all parameters in the base classification network and variational autoencoder network to obtain the trained base classification model and variational autoencoder model.

[0123] S4: Train the base classification network for the kth iteration:

[0124] S41: Define the current iteration number as k, and initialize k=1; use the parameters of the trained base classification model as the base classification model for the (k-1)th iteration, and denot the parameters of the base classification model for the (k-1)th iteration as... The parameters of the trained variational autoencoder model are used as the variational autoencoder model for the (k-1)th iteration, and the parameters of the variational autoencoder model for the (k-1)th iteration are denoted as... ;

[0125] S42: Collect the 5G antenna dataset of the incremental category in the k-th iteration, denoted as... ,in, express The Middle The first basic category One sample, express Category tags;

[0126] S43: Will The input is processed in the feature extraction module of the pre-trained basic classification model to obtain feature vectors. ;

[0127] S44: Randomly draw a Gaussian distribution of feature vectors of any class from the memory bank to generate the first feature vector of any class s. 1 synthesized feature vector and with By splicing, we obtain the first... splicing features Then, it is input into the basic classification model in the k-th iteration for processing, to obtain... probability distribution .

[0128] S45: Construct the loss function of the base classification model for the kth iteration according to equation (9). :

[0129] (9)

[0130] In equation (5), Representative dataset The total number of categories.

[0131] S46: Update strategy for building the base classification model:

[0132] S461: Define and initialize the first-order momentum in the (k-1)th iteration. and second momentum ;

[0133] S462: Calculate the classification model loss gradient for the kth iteration using equation (10). :

[0134] (10)

[0135] In equation (8), express exist gradient at;

[0136] S463: Using equation (11) Gradient fuzzing is performed to obtain the fuzzy gradient with scale s in the k-th iteration. :

[0137] (11)

[0138] In equation (11), The Gaussian blur kernel with scale s is specifically expressed as:

[0139] (12)

[0140] In equation (12), Represents the gradient matrix coordinates in Represents standard deviation, Is The corresponding Gaussian distribution weights at each location.

[0141] S464: Use equation (13) to obtain the weighted sum of the k-th iteration. :

[0142] (13)

[0143] In equation (13), The weights representing a scale of s, represents a nonlinear activation function; S represents the maximum scale.

[0144] S465: The first-order momentum of the k-th iteration is obtained using equations (14) and (15). and second momentum :

[0145] (14)

[0146] (15)

[0147] In equations (14) and (15), and This represents two attenuation factors.

[0148] S466: Use equation (16) to obtain the parameters of the base classification model in the k-th iteration. :

[0149] (16)

[0150] In equation (13), It's the learning rate. It is a parameter to prevent the denominator from being 0.

[0151] S5: Training the variational autoencoder network for the k-th iteration:

[0152] S51: Will The input is processed in the encoder of the trained variational autoencoder model to obtain the feature vector. ; After further processing through the mean layer and variance layer, the feature vector is obtained. mean and variance ;

[0153] S52: and After reparameterization with variable hierarchies, the latent space samples for the kth iteration are generated using equation (17). The latent space samples from the k-th iteration are then input into the decoder for processing, and the reconstructed features from the k-th iteration are obtained using equation (18). :

[0154] (17)

[0155] (18)

[0156] In equation (18), This represents the decoder for the (k-1)th iteration.

[0157] S53: Using equations (19)-(21), establish the loss function of the variational autoencoder model for the k-th iteration. :

[0158] (19)

[0159] (20)

[0160] (twenty one)

[0161] In equations (19)-(21), It is the kth iteration. Mean of the synthetic feature vectors of each category ; It is the kth iteration. The synthesized feature vectors of the n categories in the nth category are... variance in each dimension ; It is the divergence loss of the k-th iteration; The reconstruction loss in the kth iteration.

[0162] S54: Update strategy for constructing variational autoencoder models:

[0163] S541: Calculate the loss gradient of the variational autoencoder model in the k-th iteration using equation (22). :

[0164] (twenty two)

[0165] In equation (22), express exist gradient at;

[0166] S542: Calculate the parameters of the variational autoencoder model in the kth iteration using equation (23). :

[0167] (twenty three)

[0168] In equation (18), This represents the learning rate.

[0169] S6: After assigning k+1 to k, return to S413 and execute sequentially until the maximum number of iterations is reached, thus obtaining the incremental classification model and its parameters. Incremental variational autoencoder model and its parameters ;

[0170] S7: Input any 5G antenna sample into the incremental classification model for processing to obtain the predicted category label.

[0171] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0172] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. An incremental 5G antenna identification method based on synthetic feature replay, characterized in that, Includes the following steps: S1: Obtain the 5G antenna dataset for the basic categories, denoted as... ,in, express The Middle The first basic category One sample, express Category tags, express The number of basic categories included. Indicates the first The number of samples in each basic category; S2: Construct a basic classification network and a variational autoencoder network; S3: Training the basic classification network and variational autoencoder network: S31: The feature extraction module in the basic classification network... Processing is performed to obtain eigenvectors ; After being input into the classification module of the basic classification network for recognition, it is then processed by the softmax layer to obtain... Category probability distribution ; The training loss function of the basic classification network is established using equation (1). : (1) S32: Variational autoencoder network pair Processing is performed to obtain reconstructed features. Therefore, the loss function of the variational autoencoder network can be established using equation (2). : (2) In equation (2), Indicates divergence loss. Indicates the reconstruction loss. Indicates hyperparameters; S33: Using equations (3) and (4), we obtain the first... Mean of the synthetic feature vectors of each category and the The synthesized feature vectors of the n categories in the nth category variance in each dimension Thus, the first Variance of the synthetic feature vectors of each category And thus obtain the first Gaussian distribution of synthesized feature vectors for each category And store it in the memory bank: (3) (4) S34: Use the gradient descent algorithm to optimize all parameters in the basic classification network and variational autoencoder network to obtain the trained basic classification model and variational autoencoder model; S4: Train the base classification network for the kth iteration: S41: Define the current iteration number as k, and initialize k=1; The parameters of the trained base classification model are used as the base classification model for the (k-1)th iteration, and the parameters of the base classification model for the (k-1)th iteration are denoted as... ; The parameters of the trained variational autoencoder model are used as the variational autoencoder model for the (k-1)th iteration, and the parameters of the variational autoencoder model for the (k-1)th iteration are denoted as... ; S42: Collect the 5G antenna dataset of the incremental category in the k-th iteration, denoted as... ,in, express The Middle The first basic category One sample, express Category labels; S43: Will The input is processed in the feature extraction module of the trained base classification model to obtain feature vectors. ; S44: Randomly draw a Gaussian distribution of feature vectors of any class from the memory bank to generate the first feature vector of any class s. 1 synthesized feature vector and with By splicing, we obtain the first... splicing features Then, it is input into the basic classification model in the k-th iteration for processing, to obtain... probability distribution ; S45: Construct the loss function of the base classification model for the kth iteration according to equation (5). : (5) In equation (5), Representative dataset The total number of categories; S46: Update the parameters of the base classification model in the (k-1)th iteration using an optimization method combining multi-scale gradient fuzziness and momentum. Thus, the update strategy for the basic classification model in the k-th iteration is constructed; S461: Define and initialize the first-order momentum in the (k-1)th iteration. and second momentum ; S462: Calculate the classification model loss gradient for the k-th iteration using equation (9). : (9) In equation (9), express exist gradient at; S463: Using equation (10) Gradient fuzzing is performed to obtain the fuzzy gradient with scale s in the k-th iteration. : (10) In equation (10), This represents a Gaussian blur kernel with a scale of s; S464: Use equation (11) to obtain the weighted sum of the k-th iteration. : (11) In equation (11), The weights representing the scale s, Represents a nonlinear activation function; S denotes the maximum scale; S465: The first-order momentum of the k-th iteration is obtained using equations (12) and (13). and second momentum : (12) (13) In equations (12) and (13), and This indicates two attenuation factors; S446: Use equation (14) to obtain the parameters of the base classification model in the k-th iteration. : (14) In equation (14), It's the learning rate. It is a parameter to prevent the denominator from being zero; S5: Training the variational autoencoder network for the k-th iteration: S51: Will The input is processed in the encoder of the trained variational autoencoder model to obtain the reconstructed features of the k-th iteration. Therefore, the loss function of the variational autoencoder model in the kth iteration is established using equation (6). : (6) In equation (6), It is the divergence loss of the k-th iteration. It is the hyperparameter of the k-th iteration. It is the reconstruction loss of the k-th iteration; S52: Update the parameters of the variational autoencoder model in the (k-1)th iteration using gradient descent. Thus, the update strategy of the variational autoencoder model in the k-th iteration is constructed; S6: After assigning k+1 to k, return to S42 and execute sequentially until the maximum number of iterations is reached, thereby obtaining the optimal incremental classification model and its optimal parameters. And the optimal incremental variational autoencoder model and its optimal parameters ; S7: Input any 5G antenna sample into the optimal incremental classification model for processing to obtain the predicted category label.

2. The incremental 5G antenna identification method based on synthetic feature replay according to claim 1, characterized in that, The basic classification network in S2 includes: a feature extraction module and a classification module; The variational autoencoder network includes: an encoder, a mean layer, a variance layer, a variational layer, and a decoder; The feature extraction module and the encoder have the same network structure, and each consists of an input convolutional layer, Q downsampling convolutional layers, and W residual convolutional layers. Each convolutional layer is followed by an instance regularization layer and a ReLU activation function. The classification module includes: a fully connected layer and a softmax layer.

3. The incremental 5G antenna identification method based on synthetic feature replay according to claim 2, characterized in that, S32 includes the following steps: S321: Input into the encoder to obtain Another feature vector ; S322: After processing through the mean layer and variance layer respectively, the corresponding feature vectors are obtained. mean and variance ; S323: and After undergoing variable-level reparameterization, latent space samples are generated. S324: The latent space samples are processed by the decoder to obtain reconstructed features. ; S325: Establish using equations (7) and (8) respectively and : (7) (8) In equation (7), express The Middle Variance in each dimension This represents the total number of dimensions. express The Middle The mean of each dimension.

4. The incremental 5G antenna identification method based on synthetic feature replay according to claim 2, characterized in that, S51 includes the following steps: S511: Will The input is processed in the encoder of the trained variational autoencoder model to obtain the feature vector. ; After further processing through the mean layer and variance layer, the feature vector is obtained. mean and variance ; S512: and After reparameterization with varying hierarchies, the latent space sample for the k-th iteration is generated. This latent space sample is then input into the decoder for processing to obtain the reconstructed features for the k-th iteration. ; S513: Using equations (15) and (16), establish the loss function of the variational autoencoder model for the k-th iteration. : (15) (16) In equations (15) and (16), It is the kth iteration. The mean of the synthetic feature vectors of each category; It is the kth iteration. The synthesized feature vectors of the n categories in the nth category The variance in each dimension.

5. The incremental 5G antenna identification method based on synthetic feature replay according to claim 1, characterized in that, S52 includes the following steps: S521: Calculate the loss gradient of the variational autoencoder model in the k-th iteration using equation (17). : (17) In equation (17), express exist gradient at; S522: Calculate the parameters of the variational autoencoder model in the kth iteration using equation (18). : (18) In equation (18), This represents the learning rate.

6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing any of the incremental 5G antenna identification methods of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the incremental 5G antenna identification method according to any one of claims 1-5.

Images