A kind of K center feature distribution reconstruction-based incremental radar emitter individual intelligent identification method

Abstract

The application relates to a kind of incremental radar radiation source individual intelligent identification method based on K center feature distribution reconstruction, belong to radar data processing technical field. Including: initial category data set is generated by simulation, initial category data set is used as the input of feature extractor, feature extractor is trained based on improved supervised contrast loss function, feature extraction is carried out using feature extractor to obtain training set features;The training set features of each class are extracted based on K-means multi-center, obtain each cluster center of each class as the K center set of this class feature and the point number contained in the cluster where the K center is located, generate point number noise around K center as pseudo-feature;Pseudo-feature and new class feature are merged, and network classifier is trained. The method can process the task of radar radiation source individual incremental identification under small sample condition, and significantly improve the generalization of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar data processing, and particularly relates to a kind of incremental radar radiation source individual intelligent identification method based on K center feature distribution reconstruction. BACKGROUND

[0002] Radiation source individual identification refers to the identity identification of specific individuals or devices of electromagnetic signal sources, mainly through analyzing the characteristics of electromagnetic wave signals emitted by them to distinguish different radiation sources. This technology is widely used in radar, communication systems and radio monitoring fields, aiming to identify and track the unique characteristics of different transmitting devices to determine their identity or source. This technology has important application value in the fields of communication and monitoring.

[0003] With the development of electronic technology and information technology, it is necessary to identify the type and individual of radar radiation source signals in electronic warfare and electronic reconnaissance tasks. However, the types and numbers of radiation sources in electromagnetic space are more complex and changeable, which further increases the difficulty of electronic reconnaissance radiation source individual identification. The existing radar radiation source individual identification method based on closed set assumption can only identify known radiation source individual signals. However, in actual application scenarios, there will be a large number of new class radiation source individual signals. The radiation source individual identification system deployed in an open environment for a long time should have the ability to continuously learn new categories, so as to continuously meet the needs of new identification tasks.

[0004] In order to realize that the identification model can continuously learn new knowledge and avoid catastrophic forgetting, the class incremental learning algorithm is applied to the radar radiation source individual identification task. The radiation source individual identification method based on incremental learning algorithm refers to classifying radar signals or other types of electromagnetic radiation source signals, gradually learning the features of new individuals or new categories, and not forgetting the identification ability of old class individuals. This learning method can cope with gradually increasing radiation source individuals over time without using all training set data to retrain the entire model.

[0005] Therefore, it is of important academic and military application value to study the radar radiation source individual identification technology in an open environment, which ensures the identification accuracy of known radiation source individuals and enables the neural network model to have continuous learning ability.

[0006] In the existing radar emitter individual identification method based on incremental learning, the incremental learning algorithm based on data playback usually needs to save a part of the data samples of the old class for playback to alleviate the catastrophic forgetting when introducing a new class. However, this method does not meet the privacy protection requirements, especially for the radar emitter individual data set, directly saving and using the data samples may lead to data privacy leakage. In contrast, the class-incremental learning algorithm based on no example avoids data playback, but the identification accuracy of the old class after incremental training will decrease significantly, which fails to effectively cope with catastrophic forgetting. SUMMARY

[0007] The technical problem to be solved by the present application is:

[0008] In order to avoid the deficiencies of the prior art in radar emitter signal identification in the face of class-incremental situation, the present application provides a class-incremental radar emitter individual intelligent identification method based on K-center feature distribution reconstruction, which is used to solve the catastrophic forgetting phenomenon of the radar emitter individual identification method under the condition of class-incremental, improve the radar emitter individual identification accuracy under the condition of small sample, and ensure the privacy of the radar emitter individual data set.

[0009] In order to solve the above technical problems, the technical scheme adopted by the present application is:

[0010] A class-incremental radar emitter individual intelligent identification method based on K-center feature distribution reconstruction, characterized in that it comprises:

[0011] Simulate to generate a plurality of classes of radar emitter individual data sets with noise, and divide each class of samples into a training set and a test set;

[0012] Construct a pre-trained feature extractor, which comprises three multi-scale convolution modules fused with multi-head attention mechanism, a global adaptive average pooling layer, and a fully connected neural network layer;

[0013] Extract N sampling points of the training set as the first channel input of the feature extractor, and perform FFT transformation on the N sampling points, and take the real part and the imaginary part of the FFT transformed part as the second channel and the third channel input of the feature extractor respectively;

[0014] Train the feature extractor based on the improved supervised contrast loss function, and extract features from the training set using the trained feature extractor;

[0015] Perform multi-center extraction based on K-means on the training set features of each class to obtain each cluster center of each class as the K-center set of the features of this class and the number of points contained in the cluster where the K-center is located, and generate a number of noise points around the K-center as pseudo-features;

[0016] The pseudo features and the new class features are combined, a network classifier is trained, the network classifier is a two-layer fully connected layer neural network classifier, the number of input layer neurons is equal to the feature dimension number 512, the number of intermediate layer neurons is set to 128, and the number of output layer neurons is equal to the current total number of categories;

[0017] The trained feature extractor is used for feature extraction of a test set, and the features of the test set are input into the trained network classifier for classification test of the recognition accuracy.

[0018] Further technical solutions of the present application: further comprising normalizing the input samples of the first channel, the second channel and the third channel to satisfy the distribution with a mean of 0 and a standard deviation of 1.

[0019] Further technical solutions of the present application: after training the feature extractor, further comprising constructing and training an initial category classifier, the initial category classifier is used for classifying the initial category training set.

[0020] Further technical solutions of the present application: the multi-scale convolution module of the fusion multi-head attention mechanism comprises:

[0021] The input data is respectively subjected to one-dimensional convolution branches of three different scales:

[0022] Y convk =Conv1D(X,W convk )+b convk

[0023] Wherein, k is the scale of convolution, Y convk is the output of the convolution branch of the scale k, and X is the input;

[0024] The input is also subjected to a pooling convolution branch:

[0025] Y pool =Conv1D(MaxPool1D(X))

[0026] Wherein, Y pool is the output of the pooling convolution branch, and X is the input;

[0027] The outputs of the branches are spliced in the channel direction to obtain the output of the multi-scale convolution:

[0028] Y inception =Concat(Y conv1 ,Y conv3 ,Y conv5 ,Y pool )

[0029] Wherein, Y inception is the output of the multi-scale convolution;

[0030] The output of the multi-scale convolution is used to obtain a module output Y using a multi-head self-attention layer:

[0031]

[0032] The further technical solution of the application is to train the feature extractor based on an improved supervised contrast loss function, comprising:

[0033] In the training using the supervised contrast loss function, the same-class sample features of each sample feature in a batch are taken as positive sample pairs, and different-class sample features are taken as negative sample pairs.

[0034] The similarity Logits between the anchor point and the contrast feature is calculated by the dot product of the feature vectors:

[0035]

[0036] wherein z i and z j are the feature vectors of sample i and sample j respectively, and τ is a temperature parameter for controlling the similarity gradient.

[0037] The probability distribution of the contrast logits is calculated using a softmax function:

[0038]

[0039] The positive sample loss and the negative sample loss are calculated for the positive and negative sample pairs respectively:

[0040]

[0041]

[0042] wherein P(i) is the positive sample set of sample i, and N(i) is the negative sample set of sample i.

[0043] The difference between the logarithmic probabilities of the positive and negative sample pairs is taken as the final loss, and the average value in the entire batch is taken as the loss:

[0044] L=-(L pos -negative_weight·L neg )

[0045] wherein L is the final loss, and negative_weight is the negative sample weight used to control the importance of the negative sample in the loss calculation.

[0046] The further technical solution of the application is that the new class feature is obtained, comprising:

[0047] The trained feature extractor is used for feature extraction on the sample of the new class radar radiation source individual training set, and the training set features of the new class are obtained.

[0048] A further technical solution of the present application is that the training set features of each class are subjected to multi-center extraction based on K-means, and the multi-center extraction based on K-means comprises:

[0049] The number of clusters k is set to 10, and the optimization target is the sum of squares of distances from all data points to the cluster centers to which they belong:

[0050]

[0051] wherein x i is the i th data point, mu j is the j th cluster center, C j is the set of all data points in the j th cluster, ||x i -mu j || 2 is the square of the Euclidean distance;

[0052] The cluster assignment and center update are performed through iteration, and the K-means algorithm attempts to minimize the objective function J so as to divide the data set into 10 optimal clusters.

[0053] A further technical solution of the present application is that the pseudo features conform to a normal distribution with K centers as the mean and a standard deviation sigma:

[0054] X i ~ N (c i , sigma 2 )

[0055] wherein c i is the i th K center position.

[0056] A computer system comprises one or more processors, a computer readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0057] A computer readable storage medium stores computer executable instructions, and the instructions are used to implement the above method when executed.

[0058] The present application has the following beneficial effects:

[0059] The class incremental radar radiation source individual intelligent identification method based on K center feature distribution reconstruction provided by the present application can process the task of incremental identification of radar radiation source individuals under the condition of small samples, and significantly improves the generalization of the model.

[0060] 1、The present application only needs to train a classifier in the incremental training stage, which greatly reduces the calculation resources consumed by training when new class data arrives; and the present application only saves the centers of the multiple K-means after feature extraction, and does not save the original data, which ensures the privacy of the data.

[0061] 2、The present application designs a K-center feature reconstruction mechanism based on the feature space to reconstruct the feature distribution of the old class samples. When new class sample data arrives, the feature distribution of the new class sample data extracted by the pre-trained feature extractor is trained with the reconstructed old sample feature distribution to train a new classifier. In this way, the catastrophic forgetting phenomenon caused by class increment can be significantly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0062] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the principles of the application. In the drawings:

[0063] Figure 1 The present application is based on the K-center feature reconstruction mechanism of the class incremental radar radiation source individual identification algorithm flow chart.

[0064] Figure 2 The present application is based on the K-center feature reconstruction mechanism of the class incremental radar radiation source individual identification algorithm flow chart.

[0065] Figure 3 The present application is based on the K-center feature reconstruction mechanism of the class incremental radar radiation source individual identification algorithm flow chart.

[0066] Figure 4 The present application is based on the K-center feature reconstruction mechanism of the class incremental radar radiation source individual identification algorithm flow chart. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0068] It should be noted that the terms "first", "second", and the like in the description and in the claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the terms thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not necessarily have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0069] The application provides a kind of class incremental radar radiation source individual identification method based on K center feature reconstruction mechanism, utilize the feature extractor of initial class sample training using supervised contrast learning loss function, the pre-training feature extractor of the feature extractor can make the feature of same class gather, the feature of different class separates.For each class sample feature is extracted respectively to do multi-center K clustering, save K center as the prototype of feature reconstruction, to reconstruct the distribution of old class in feature space, realize the purpose of relieving the catastrophic forgetting of old class identification ability of incremental radar radiation source signal identification.

[0070] As shown in Figure 1 The following steps are included:

[0071] Step 1, generating data set

[0072] Simulate 36 classes of noisy radar radiation source individual data set, and 80% of the number of samples in each class is used as the training set, and 20% is used as the test set.

[0073] Step 2, data preprocessing

[0074] The first 1000 points of each sample in the radar radiation source individual data set are extracted as the first channel input of the feature extractor network. Then the FFT transform is performed on the one-dimensional sample data, and the real part and the imaginary part of the FFT transformed are taken as the second and third channel inputs of the feature extractor network respectively.

[0075] Specifically: the radar radiation source individual data set is composed of simulated radar pulse intermediate frequency sequences, and the first 1000 sample point amplitude values of each radar pulse intermediate frequency sequence in the radar radiation source individual data set are extracted as the first channel input of the feature extractor network. Then the FFT transform is performed on the 1000 sample point amplitude values obtained, and the real part and the imaginary part of the FFT transformed are taken as the second and third channel inputs of the feature extractor network respectively.

[0076] The three input channels are respectively subjected to Z-score normalization, and the mean of each sample is subtracted from the three channels respectively and then divided by the standard deviation. The normalized data will have a distribution with a mean of 0 and a standard deviation of 1.

[0077] The Z-score normalization formula is as follows:

[0078]

[0079] Wherein, X is the original data, mu is the mean of the sample, sigma is the standard deviation of the sample, X norm is the normalized data.

[0080] In this step, the radar radiation source individual data time domain information is taken as the first channel input of the feature extractor, and then the one-dimensional sample data is subjected to FFT transformation, and the real part and the imaginary part after the FFT transformation are taken as the second and third channel inputs of the feature extractor network respectively, so that the time and frequency domain information fusion of the radar radiation source individual data is realized, which is helpful to improve the recognition accuracy of the model.

[0081] Step 3, build a feature extractor integrating multi-head attention mechanism and multi-scale convolution module

[0082] The feature extractor used in the application is based on a multi-scale convolution architecture, and a multi-head attention mechanism is introduced to enhance the feature extraction capability of the model on sequence data. The structure of the feature extractor integrating the multi-head attention mechanism and the multi-scale convolution module is as shown in Figure 2 , which comprises: a multi-scale convolution module integrating a multi-head attention mechanism, a global adaptive average pooling layer and a fully connected neural network layer.

[0083] The mathematical formula of the multi-head attention layer calculation is explained as:

[0084]

[0085] Wherein, Q is a query matrix, K is a key matrix, V is a value matrix, d k is the dimension of the key.

[0086] The application uses a multi-head self-attention mechanism, Q, K and V are all input matrices X, so the mathematical formula of the multi-head self-attention layer calculation in the feature extractor is explained as:

[0087]

[0088] Wherein, X is the input matrix of the multi-head self-attention layer, and Y is the output matrix of the multi-head self-attention layer.

[0089] Specifically, the structure of the multi-scale convolution module integrating the multi-head attention mechanism is as shown in Figure 3 .

[0090] The multi-scale convolution module fused with the multi-head attention mechanism firstly respectively passes the input data through three one-dimensional convolution branches of different scales:

[0091] Y convk =Conv1D(X,W convk )+b convk

[0092] Wherein, k is the scale of convolution, Y convk is the output of the convolution branch of scale k, and X is the input.

[0093] The input also passes through a pooling convolution branch:

[0094] Y pool =Conv1D(MaxPool1D(X))

[0095] Wherein, Y pool is the output of the pooling convolution branch, and X is the input.

[0096] The outputs of each branch are spliced in the channel direction to obtain the output of the multi-scale convolution:

[0097] Y inception =Concat(Y conv1 ,Y conv3 ,Y conv5 ,Y pool )

[0098] Wherein, Y inception is the output of the multi-scale convolution.

[0099] Finally, the multi-head self-attention layer is used on the output of the multi-scale convolution to obtain the module output Y:

[0100]

[0101] In this step, the feature extractor used is based on a multi-scale convolution architecture, and a multi-head attention mechanism is introduced to enhance the feature extraction capability of the model on sequence data, so that the feature extractor can better extract the features of the radar emitter individual signal.

[0102] Step 4, training the feature extractor

[0103] The feature extractor is trained using the initial class training set and using an improved supervised contrastive loss function, so as to obtain a feature extractor that can make the intra-class feature distribution gather and the inter-class feature distribution separate.

[0104] In the training using the supervised contrastive loss function, the same-class sample features of each sample feature in a batch are respectively taken as positive sample pairs, and different-class sample features are taken as negative sample pairs.

[0105] First, the similarity between the anchor point and the contrast feature is calculated by the dot product of the feature vector:

[0106]

[0107] where z i and z j are the feature vectors of sample i and sample j, respectively, and τ is the temperature parameter that controls the gradient of similarity.

[0108] The probability distribution of the contrast logits is calculated using the softmax function:

[0109]

[0110] The positive sample loss and the negative sample loss are calculated for the positive and negative sample pairs, respectively:

[0111]

[0112] where P(i) is the positive sample set of sample i, and N(i) is the negative sample set of sample i.

[0113] The final loss is the difference between the logarithmic probabilities of the positive and negative sample pairs, and then the average value in the entire batch is taken as the loss.

[0114] L = -(L pos -negative_weight·L neg )

[0115] where L is the final loss, and negative_weight is the negative sample weight used to control the importance of negative samples in loss calculation.

[0116] In this step, the feature extractor is trained using an improved supervised contrast loss function. The supervised contrast loss function is improved by adding negative sample loss, which makes the features between classes more separated and helps to improve the recognition accuracy.

[0117] Step 5: Feature extraction on the initial class training set samples

[0118] The training set features are obtained by using the feature extractor on the preprocessed radar emitter individual signal training set samples in step 2. The feature dimension of each sample after feature extraction is 512.

[0119] Step 6: Multi-center extraction based on K-means for each class of training set features

[0120] K-means clustering algorithm is used for each class of training set features, the core idea of K-means algorithm is to divide the data set into k clusters, so that the distance between data points in each cluster and the cluster center is minimized. The number of clusters k is set to 10 in the application, and the optimization goal of K-means is to minimize the intra-cluster squared error (Inertia), that is, the sum of the squared distances of all data points to their cluster centers:

[0121]

[0122] where x i is the i-th data point, μ j is the j-th cluster center, C j is the set of all data points in the j-th cluster, ||x i -μ j || 2 is the squared Euclidean distance. Through iteration of cluster assignment and center update, K-means algorithm tries to minimize the objective function J to divide the data set into 10 optimal clusters.

[0123] Each cluster center of each class is extracted and saved as the K-center set of this class feature and the number of points n contained in the cluster of the K-center.

[0124] In this step, K-means algorithm is used to cluster the features of each class with cluster number k = 10, and the cluster centers are extracted as the K-center set of this class feature, and the number of points n contained in the cluster of each K-center is recorded i . Finally, n i points of noise are generated for each K-center as pseudo-features, and the pseudo-features generated based on K-center can achieve the purpose of reconstructing the distribution of old class features.

[0125] Step 7, construct and train the initial class classifier

[0126] A two-layer fully connected neural network classifier is constructed, the number of input layer neurons is equal to the feature dimension number 512, the number of intermediate layer neurons is set to 128, and the number of output layer neurons is equal to the number of initial classes 18. The classifier is trained using the initial class training set features, and the cross-entropy loss function is used to update the classifier parameters.

[0127] Step 8, test the recognition accuracy of the initial stage model

[0128] The feature extractor trained in step 4 is used to extract features for the initial class test set samples, and then the classifier trained in step 7 is used for classification test to test its recognition accuracy.

[0129] Step 9, feature extraction for new class radar radiation source individual training set samples

[0130] When new class data arrives, the feature extractor trained in step 4 is used to extract features from the new class radar emitter individual training set samples to obtain the training set features of the new class.

[0131] Step 10, generating pseudo features for old classes

[0132] The pseudo features are generated according to the K-center set of the old class features saved in step 6, and the pseudo features follow a normal distribution with K-center as the mean and standard deviation σ:

[0133] X i ~ N(c i ,σ 2 )

[0134] Where i is the i-th K-center, and the standard deviation σ is set to 0.1 in the present application. The number of pseudo features generated by each K-center is equal to the number of points n i contained in the cluster where the K-center is located, and the generated pseudo features are equivalent to reconstructing the old class feature distribution.

[0135] Step 11, constructing and training a new classifier

[0136] A two-layer fully connected neural network classifier is constructed, with the number of input layer neurons equal to the feature dimension number 512, the number of intermediate layer neurons set to 128, and the number of output layer neurons equal to the current total number of classes. The old class pseudo features generated in step 10 and the new class features extracted in step 9 are combined, and the combined features are used to train the new classifier.

[0137] In this step, when new class data arrives, a new classifier is trained using pseudo features and new class features, and this classifier has the ability to identify new class and old class radar emitter individuals.

[0138] Step 12, test the recognition accuracy of the incremental stage model

[0139] The feature extractor trained in step 4 is used to extract features from the test set samples, and then the classifier trained in step 11 is used for classification to test its recognition accuracy.

[0140] The above entire method does not save any old class sample data, realizing example-free incremental learning. The privacy of the radar emitter individual data set is guaranteed.

[0141] The present application has the following advantages:

[0142] 1. Considering that the radar emitter individual signal contains periodic and non-periodic information, steps 2, 3 and 4 apply new data preprocessing methods, new feature extractors and improved supervised contrast loss functions, so that the feature extraction effect is better and the model recognition accuracy is higher.

[0143] 2. Radar emitter individual data set requires its confidentiality, and old class samples cannot be used during incremental training. This is the significance of using an incremental learning algorithm in radar emitter individual identification. The incremental learning algorithm based on K-center feature distribution reconstruction designed by the present application achieves the purpose of example-free incremental learning. Compared with other example-free incremental learning algorithms, the forgetting rate of the algorithm of the present application is lower.

[0144] 3. The incremental learning algorithm designed by the present application only needs to train the classifier when a new class arrives for incremental training, without training the feature extractor, which significantly reduces the training cost and reduces the time of incremental training, making the algorithm more flexible in practical application.

[0145] The effect of the present application will be further described below in combination with a simulation experiment:

[0146] 1. Simulation conditions:

[0147] The hardware platform of the simulation experiment of the present application is: the processor is Intel(R) Core i9-10920X CPU, the main frequency is 3.5GHZ, the memory is 64GB; the graphics card is NVIDIA GeForce RTX 4090, and the video memory capacity is 24GB.

[0148] The software platform of the simulation experiment of the present application is: WINDOWS10 operating system, MATLAB R2021b, Python3.8, Pytorch 1.31.

[0149] 2. Radar emitter individual data set simulation:

[0150] The radar emitter individual data set used by the present application is generated by MATLAB simulation. By changing the pulse width T, bandwidth B, and carrier frequency f of the radar emitter c to simulate three different types of radar emitters, and then changing the individual feature parameters of the radar emitters, including carrier frequency offset and phase noise parameters to simulate different radar emitter individual signals. The present application simulates 12 types of emitters, each type having 3 different individual signals, generating a radar emitter individual data set of 36 classes with 2000 samples per class. Table 1 shows the radar emitter type parameter settings, and Table 2 shows the radar emitter individual feature parameter settings.

[0151] Table 1 Radar emitter type parameter table

[0152]

[0153]

[0154] Table 2 individual characteristic parameter table

[0155]

[0156] 3. Radar emitter individual incremental identification experiment process and result

[0157] The present application carries out the radar emitter individual incremental identification experiment, 80% of each class sample of radar emitter individual data set is used as the training set, and 20% is used as the test set; And the experiment is under the condition of small sample to radar emitter individual incremental identification, 10% of each class sample of radar emitter individual data set is used as the training set, and 90% is used as the test set.

[0158] The 18 radar emitter individuals of type one, type two, type three, type four, type five and type six in the data set are used as the initial category; 3 individuals of each type are added each time, and a total of 6 times are added. Radar emitter individual incremental identification experiment is carried out under the condition that the signal-to-noise ratio is 5dB and 10dB, and the experimental results are shown in Table 3 and Figure 4

[0159] Table 3 experimental results

[0160]

[0161] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered in the protection scope of the present application.​

Claims

1. A K-center feature distribution reconstruction-based incremental radar emitter individual intelligent identification method, characterized in that, The method comprises the steps of: imitating to generate several types of noisy radar emitter individual data sets, and dividing each type of sample into a training set and a test set; a pre-trained feature extractor is constructed, which comprises three multi-scale convolution modules with a multi-head attention mechanism, a global adaptive average pooling layer, and a fully connected neural network layer; The multi-scale convolution module with a multi-head attention mechanism comprises: The input data is respectively subjected to one-dimensional convolution branches of three different scales: wherein, is a scale of a convolution, is a scale of a convolution branch output, is an input; The input is also subjected to a pooling convolution branch: wherein is the output of the pooling convolution branch, is the input; The outputs of each branch are spliced in the channel direction to obtain the output of the multi-scale convolution: wherein, is the output of the multiscale convolution; using a multi-head self-attention layer on the output of the multi-scale convolution to obtain a module output : N sampling points of the training set are extracted as the first channel input of the feature extractor, and the N sampling points are subjected to FFT transformation, and the real part and the imaginary part after the FFT transformation are respectively taken as the second channel and the third channel input of the feature extractor; the input samples of the first channel, the second channel and the third channel are normalized to satisfy the distribution with a mean value of 0 and a standard deviation of 1; The feature extractor is trained based on an improved supervised contrast loss function, and the training set features are extracted by using the trained feature extractor; the feature extractor is trained based on the improved supervised contrast loss function, which comprises: In the training using the supervised contrast loss function, the same sample features of each sample feature in a batch are taken as positive sample pairs, and different sample features are taken as negative sample pairs; The similarity Logits between the anchor point and the contrast feature is calculated by the dot product of the feature vectors: wherein, and are eigenvectors of the samples and the samples respectively, is a temperature parameter controlling the similarity gradient. The probability distribution of the contrast logits is calculated using the softmax function: The positive sample loss and the negative sample loss are calculated for the positive and negative sample pairs respectively: wherein, is a positive sample set of samples ; is a negative sample set of samples ; The difference between the logarithmic probabilities of the positive sample and the negative sample is taken as the final loss, and the average value in the whole batch is taken as the loss: wherein, is the final loss, is the negative sample weight to control the importance of negative samples in loss calculation K-means-based multi-center extraction is performed on the training set features of each class to obtain each cluster center of each class as the K-center set of the features of this class and the number of points contained in the cluster where the K-center is located, and a number of noise points are generated around the K-center as pseudo-features; the K-means-based multi-center extraction of the training set features of each class comprises: Setting the number of clusters The optimization goal is the sum of squared distances of all data points to their belonging cluster centers: wherein, is the th data point, is the th cluster center, is the th cluster of all data points, is the squared Euclidean distance; The K-means algorithm attempts to minimize the objective function by iterating cluster assignment and center update Thus the dataset is partitioned into 10 optimal clusters; The pseudo-features and the new class features are merged, and a network classifier is trained, wherein the network classifier is a two-layer fully connected neural network classifier, the number of input layer neurons is equal to the feature dimension number 512, the number of intermediate layer neurons is set to 128, and the number of output layer neurons is equal to the current total number of classes; the new class feature comprises: The training set features of the new class are extracted by using the trained feature extractor, and the training set features of the new class are obtained. The pseudo features obey a normal distribution with K centers as mean values and standard deviations as ​ wherein is the first i K center position; The trained feature extractor is used to extract the features of the test set, and the features of the test set are input into the trained network classifier for classification to test the recognition accuracy.

2. The K-center feature distribution based reconstruction method for incremental radar emitter individual intelligent recognition according to claim 1, characterized in that, After training the feature extractor, an initial class classifier is constructed and trained, which is used for classifying the initial class training set.

3. A computer system, characterized by The method comprises the steps of: one or more processors, a computer readable storage medium storing one or more programs for execution by the one or more processors, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to carry out the method of claim 1.

4. A computer-readable storage medium, characterized in that a computer readable storage medium storing computer executable instructions which, when executed, implement the method of claim 1.

Images