Class incremental radar radiation source individual intelligent identification method based on K center feature distribution reconstruction

Abstract

The invention relates to a class incremental radar radiation source individual intelligent identification method based on K center feature distribution reconstruction, and belongs to the technical field of radar data processing. Comprising the steps that an initial category data set is generated through simulation, the initial category data set serves as input of a feature extractor, the feature extractor is trained based on an improved supervised contrast loss function, and the feature extractor is adopted for feature extraction to obtain training set features; performing K-means-based multi-center extraction on the training set features of each class to obtain each cluster center of each class as a K center set of the class of features and points contained in a cluster where the K center is located, and generating points of noise around the K center as pseudo features; and combining the pseudo features and the new features, and training a network classifier. According to the method, the task of radar radiation source individual incremental recognition under the small sample condition can be processed, and the generalization of the model is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar data processing, and in particular relates to an individual intelligent identification method of quasi-incremental radar radiation sources based on K-center feature distribution reconstruction. Background Art

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

[0003] With the development of electronic technology and information technology, it is necessary to identify the type and individual of radar emitter signals in electronic warfare and electronic reconnaissance missions. However, the types and numbers of emitters in electromagnetic space are more complex and changeable, which further increases the difficulty of identifying individual emitters in electronic reconnaissance. The existing radar emitter individual identification method based on closed set assumption can only identify individual signals of known emitters. However, in actual application scenarios, a large number of new types of emitter individual signals will appear. The emitter 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 achieve the goal of enabling the recognition model to continuously learn new knowledge and avoid catastrophic forgetting, the class-based incremental learning algorithm is applied to the radar emitter individual recognition task. The emitter individual recognition method based on the incremental learning algorithm refers to gradually learning the characteristics of new individuals or new categories by classifying radar signals or other types of electromagnetic radiation source signals, and will not forget the recognition ability of old categories. This learning method can cope with the gradual increase of emitter individuals over time without using all the training set data to retrain the entire model.

[0005] Therefore, studying the individual identification technology of radar emitters in an open environment can enable the neural network model to have the ability of continuous learning while ensuring the accuracy of individual identification of known emitters, which has important academic and military application value.

[0006] Among the existing radar emitter individual recognition methods based on incremental learning, the incremental learning algorithm based on data replay usually needs to save a part of the data samples of the old categories for replay when introducing new categories to alleviate catastrophic forgetting. However, this method does not meet the privacy protection requirements. In particular, for radar emitter individual data sets, directly saving and using data samples may lead to data privacy leakage. In contrast, although the example-free incremental learning algorithm avoids data replay, the recognition accuracy of the old categories will drop significantly after incremental training, and it fails to effectively deal with catastrophic forgetting. Summary of the invention

[0007] The technical problems to be solved by the present invention are:

[0008] In order to avoid the shortcomings of the prior art in radar emitter signal recognition when facing a quasi-incremental situation, the present invention provides a quasi-incremental radar emitter individual intelligent recognition method based on K-center feature distribution reconstruction, which is used to solve the catastrophic forgetting phenomenon of the radar emitter individual recognition method under the quasi-incremental situation, improve the accuracy of radar emitter individual recognition under small sample conditions, and ensure the privacy of radar emitter individual data sets.

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

[0010] A method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction, characterized by comprising:

[0011] The simulation generates several types of individual data sets of noisy radar emitters, and each type of sample is divided into a training set and a test set;

[0012] Construct a pre-trained feature extractor, which includes three multi-scale convolution modules fused with a 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, perform FFT transformation on the N sampling points, and use the real part and imaginary part after FFT transformation as the second channel and third channel input of the feature extractor respectively;

[0014] The feature extractor is trained based on the improved supervised contrast loss function, and the trained feature extractor is used to extract features to obtain the training set features;

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

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

[0017] Use the trained feature extractor to extract features from the test set, and input the features of the test set into the trained network classifier for classification to test its recognition accuracy.

[0018] A further technical solution of the present invention also includes normalizing the input samples of the first channel, the second channel, and the third channel so that they satisfy a distribution with a mean of 0 and a standard deviation of 1.

[0019] A further technical solution of the present invention: after training the feature extractor, it also includes constructing and training an initial category classifier, wherein the initial category classifier is used to classify the initial category training set.

[0020] A further technical solution of the present invention: the multi-scale convolution module integrating the multi-head attention mechanism comprises:

[0021] The input data passes through three one-dimensional convolution branches of different scales:

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

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

[0024] The input also passes through a pooling convolution branch:

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

[0026] Among them, Y pool is the output of the pooled convolution branch, and X is the input;

[0027] The outputs of each branch are concatenated in the channel direction to obtain the output of multi-scale convolution:

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

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

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

[0031]

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

[0033] In the supervised contrast loss function training, the sample features of the same type of each sample feature in a batch are used as positive sample pairs, and the sample features of different types are used 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 vector:

[0035]

[0036] Among them, 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 similarity gradient;

[0037] Use the softmax function to calculate the probability distribution of contrast logits:

[0038]

[0039] For positive and negative sample pairs, calculate the positive sample loss and negative sample loss respectively:

[0040]

[0041]

[0042] Where P(i) is the positive sample set of sample i; N(i) is the negative sample set of sample i;

[0043] The difference in the log probability of the positive and negative sample pairs is taken as the final loss, and the average is taken over the entire batch as the loss:

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

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

[0046] A further technical solution of the present invention: the acquisition of the new class features includes:

[0047] The trained feature extractor is used to extract features from the individual training set samples of the new type of radar emitter to obtain the training set features of the new type.

[0048] A further technical solution of the present invention is: the multi-center extraction of the training set features of each category based on K-means comprises:

[0049] Set the number of clusters k = 10, and the optimization target is the sum of the squares of the distances from all data points to the center of their clusters:

[0050]

[0051] Among them, x i is the ith data point, μ j is the jth cluster center, C j is the set of all data points in the jth cluster, ||x i -μ j || 2 is the square of the Euclidean distance;

[0052] By iteratively performing cluster assignment and center update, the K-means algorithm attempts to minimize the objective function J to divide the data set into 10 optimal clusters.

[0053] A further technical solution of the present invention is that the pseudo-features obey a normal distribution with K centers as the mean and a standard deviation as σ:

[0054] X i ~N(c i ,σ 2 )

[0055] Among them, c i is the i-th K center position.

[0056] A computer system, characterized in that it comprises: one or more processors, and 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-mentioned method.

[0057] A computer-readable storage medium, characterized in that it stores computer-executable instructions, which are used to implement the above method when executed.

[0058] The beneficial effects of the present invention are:

[0059] The invention provides a quasi-incremental radar emitter individual intelligent identification method based on K-center feature distribution reconstruction, which can handle the task of incremental identification of radar emitter individuals under small sample conditions and significantly improve the generalization of the model.

[0060] 1. The present invention only needs to train one classifier in the incremental training stage, which greatly reduces the computing resources consumed by training when new types of data arrive; and the present invention only saves the centers of multiple K-means after feature extraction, and does not save the original data, thus ensuring the privacy of the data.

[0061] 2. The present invention designs a K-center feature reconstruction mechanism to reconstruct the feature distribution of old class samples in the feature space. When new class sample data arrives, the feature distribution of the new class sample data extracted by the pre-trained feature extractor is combined with the reconstructed feature distribution of the old sample to train a new classifier. This can significantly reduce the catastrophic forgetting phenomenon caused by class increment. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0063] Figure 1 Flow chart of the quasi-incremental radar emitter individual identification algorithm based on the K-center feature reconstruction mechanism of the present invention.

[0064] Figure 2 The structure diagram of the feature extractor of the present invention integrating the multi-head attention mechanism and the multi-scale convolution module

[0065] Figure 3 Structural diagram of the multi-head attention multi-scale convolution module of the present invention.

[0066] Figure 4 Simulation result diagram of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention 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", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] The present invention provides a quasi-incremental radar emitter individual identification method based on a K-center feature reconstruction mechanism, which uses a supervised contrast learning loss function to train a feature extractor for samples of the initial category, and the feature extractor is a pre-trained feature extractor that can aggregate features of the same category and separate features of different categories. The features of each category of samples are extracted and multi-center K clustering is performed respectively, and the K centers are saved as prototypes for feature reconstruction to reconstruct the distribution of the old category in the feature space, so as to achieve the purpose of alleviating the catastrophic forgetting of the old category recognition ability of the incremental radar emitter signal recognition.

[0070] like Figure 1 As shown, the following steps are included:

[0071] Step 1: Generate a dataset

[0072] The simulation generates 36 types of individual data sets of noisy radar emitters, and 80% of the samples in each type are used as training sets and 20% as test sets.

[0073] Step 2: Data preprocessing

[0074] For each radar emitter individual data set, the first 1000 points are intercepted as the first channel input of the feature extractor network. Then, the one-dimensional sample data is transformed by FFT, and the real and imaginary parts after FFT transformation are used as the second and third channel inputs of the feature extractor network respectively.

[0075] Specifically, the radar emitter individual data set consists of simulated radar pulse intermediate frequency sequences. The amplitude values ​​of the first 1000 sampling points in the radar pulse intermediate frequency sequence in each radar emitter individual data set are intercepted as the first channel input of the feature extractor network. Then, the intercepted 1000 sampling point amplitude values ​​are subjected to FFT transformation, and the real and imaginary parts after FFT transformation are used as the second and third channel inputs of the feature extractor network respectively.

[0076] The three input channels are Z-score normalized, and the three channels of each sample are subtracted from the mean and 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] Among them, X is the original data, μ is the mean of the sample, σ is the standard deviation of the sample, and X norm is the normalized data.

[0080] In this step, the time domain information of the individual data of the radar emitter is used as the first channel input of the feature extractor, and then the one-dimensional sample data is transformed by FFT. The real part and imaginary part after FFT transformation are used as the second and third channel inputs of the feature extractor network respectively, realizing the fusion of the time and frequency domain information of the individual data of the radar emitter, which helps to improve the recognition accuracy of the model.

[0081] Step 3: Build a feature extractor that integrates multi-head attention mechanism and multi-scale convolution module

[0082] The feature extractor used in the present invention is based on a multi-scale convolution architecture and introduces a multi-head attention mechanism to enhance the model's ability to extract features from sequence data. The structure of the feature extractor that integrates the multi-head attention mechanism and the multi-scale convolution module is as follows: Figure 2 As shown in the figure, it includes: a multi-scale convolution module integrating a multi-head attention mechanism → a global adaptive average pooling layer → a fully connected neural network layer.

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

[0084]

[0085] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key.

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

[0087]

[0088] Among them, 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 multi-scale convolution module structure of this embodiment integrates the multi-head attention mechanism as follows: Figure 3 shown.

[0090] The multi-scale convolution module that integrates the multi-head attention mechanism first passes the input data through three one-dimensional convolution branches of different scales:

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

[0092] Among them, k is the scale of the 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] Among them, Y pool is the output of the pooled convolution branch, and X is the input.

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

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

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

[0099] Finally, a 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 convolutional architecture, and a multi-head attention mechanism is introduced to enhance the model's feature extraction capability for sequence data, so that the feature extractor can better extract the features of individual signals of radar emitters.

[0102] Step 4: Train the feature extractor

[0103] The present invention uses an initial category training set and an improved supervised contrast loss function to train a feature extractor, thereby obtaining a feature extractor that can aggregate feature distribution within a class and separate feature distribution between classes.

[0104] In the training using the supervised contrast loss function, the sample features of the same type for each sample feature in a batch are taken as positive sample pairs, and the sample features of different types are taken as negative sample pairs.

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

[0106]

[0107] Among them, 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 similarity gradient.

[0108] Use the softmax function to calculate the probability distribution of contrast logits:

[0109]

[0110] For positive and negative sample pairs, calculate the positive sample loss and negative sample loss respectively:

[0111]

[0112] Among them, P(i) is the positive sample set of sample i; N(i) is the negative sample set of sample i.

[0113] The final loss is the difference in log probability of positive and negative pairs, which is then averaged over the entire batch as the loss.

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

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

[0116] In this step, the present invention uses an improved supervised contrast loss function to train a feature extractor, improves the supervised contrast loss function, increases negative sample loss, and makes inter-class features more separated, which is beneficial to improving recognition accuracy.

[0117] Step 5: Extract features from the initial category training set samples

[0118] The radar radiation source individual signal training set samples preprocessed in step 2 are subjected to feature extraction using a feature extractor to obtain training set features. The feature dimension of each sample after feature extraction is 512 dimensions.

[0119] Step 6: Perform K-means-based multi-center extraction on the training set features of each category

[0120] The K-means clustering algorithm is used for each class of training set features. The core idea of ​​the K-means algorithm is to divide the data set into k clusters so that the distance between the data points in each cluster and the cluster center is minimized. The present invention sets the number of clusters k=10, and the optimization goal of K-means is to minimize the intra-cluster square error (Inertia), that is, the sum of the squares of the distances from all data points to the cluster center to which they belong:

[0121]

[0122] Among them, x i is the ith data point, μ j is the jth cluster center, C j is the set of all data points in the jth cluster, ||x i -μ j || 2 is the square of the Euclidean distance. By iteratively performing cluster assignment and center update, the K-means algorithm attempts to minimize the objective function J to divide the data set into 10 optimal clusters.

[0123] Each cluster center of each category is extracted and saved as the K center set of such features and the number of points n contained in the cluster where the K center is located.

[0124] In this step, the K-means algorithm is used to cluster the features of each category into 10 clusters, extract the cluster centers as the K center sets of such features, and record the number of points n contained in the cluster where each K center is located. i Finally, generate n for each K center i The noise of each point is used as a pseudo feature, and the pseudo feature generated based on the K centers can achieve the purpose of reconstructing the feature distribution of the old class.

[0125] Step 7: Build and train the initial category classifier

[0126] A two-layer fully connected neural network classifier is constructed, the number of neurons in the input layer is equal to the number of feature dimensions 512, the number of neurons in the middle layer is set to 128, and the number of neurons in the output layer is equal to the number of initial categories 18. The classifier is trained using the features of the initial category training set, and the classifier parameters are updated using the cross entropy loss function.

[0127] Step 8: Test the recognition accuracy of the model in the initial stage

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

[0129] Step 9: Extract features from the individual training set samples of the new radar emitter type

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

[0131] Step 10: Generate pseudo features for old classes

[0132] Generate pseudo features based on the K center set of old class features saved in step 6. The pseudo features obey a normal distribution with K centers 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 invention. The number of pseudo features generated by each K center is equal to the number of points n contained in the cluster where the K center is located. i , the generated pseudo features are equivalent to reconstructing the distribution of old class features.

[0135] Step 11: Build and train a new classifier

[0136] Construct a two-layer fully connected neural network classifier, with the number of neurons in the input layer equal to the feature dimension of 512, the number of neurons in the middle layer set to 128, and the number of neurons in the output layer equal to the current total number of categories. Merge the old class pseudo features generated in step 10 and the new class features extracted in step 9, and use the merged features to train a new classifier.

[0137] In this step, when new class data arrives, pseudo features and new class features are used to train a new classifier, which has the ability to identify new and old class radar emitters individually.

[0138] Step 12: Test the recognition accuracy of the incremental 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 to classify and test its recognition accuracy.

[0140] In the above method, no sample data of old classes is saved, and no-example incremental learning is achieved. The privacy of individual datasets of radar emitters is guaranteed.

[0141] The present invention has the following advantages:

[0142] 1. Considering that the individual signals of radar emitters contain periodic and non-periodic information, steps 2, 3, and 4 apply a new data preprocessing method, a new feature extractor, and an improved supervised contrast loss function, which makes the feature extraction effect better and the model recognition accuracy higher.

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

[0144] 3. The incremental learning algorithm designed by the present invention only needs to train the classifier when a new category is added for incremental training, without the need to train the feature extractor. This significantly reduces the training cost and the time for incremental training, making the algorithm more flexible in practical applications.

[0145] The effect of the present invention is further described below in conjunction with simulation experiments:

[0146] 1. Simulation conditions:

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

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

[0149] 2. Simulation of individual data sets of radar emitters:

[0150] The radar radiation source individual data set used in the present invention is generated by MATLAB simulation, and the pulse width T, bandwidth B, and carrier frequency f of the radar radiation source are changed. c To simulate three different types of radar emitters, and then change the individual characteristic parameters of the radar emitters, including carrier frequency offset and phase noise parameters, to simulate different individual signals of radar emitters. The present invention simulates and generates 12 types of emitters, each of which has 3 different individual signals, generating 36 types of radar emitter individual data sets with 2000 samples per type. Table 1 gives the type parameter settings of the radar emitters, and Table 2 gives the individual characteristic parameter settings of the radar emitters.

[0151] Table 1 Radiation source type parameter table

[0152]

[0153]

[0154] Table 2 Individual characteristic parameter table

[0155]

[0156] 3. Experimental process and results of individual incremental identification of radar emitters

[0157] The present invention conducts an experiment on incremental identification of individual radar emitters, using 80% of samples of each category of the radar emitter individual data set as a training set and 20% as a test set; and the experiment conducts incremental identification of individual radar emitters under small sample conditions, using 10% of samples of each category of the radar emitter individual data set as a training set and 90% as a test set.

[0158] In the data set, all individuals 1, 2, and 3 of type 1, type 2, type 3, type 4, type 5, and type 6, a total of 18 types of radar emitter individuals, are used as initial categories; each increment is 6 times for 3 types of individuals of one type. The radar emitter individual incremental recognition experiment is carried out under the conditions of signal-to-noise ratio of 5dB and 10dB respectively. The experimental results are shown in Table 3 and Figure 4 shown.

[0159] Table 3 Experimental results

[0160]

[0161] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A method for intelligent identification of quasi-incremental radar emitter individuals based on K-center feature distribution reconstruction, characterized in that: include: The simulation generates several types of individual data sets of noisy radar emitters, and each type of sample is divided into a training set and a test set; Construct a pre-trained feature extractor, which includes three multi-scale convolution modules fused with a multi-head attention mechanism, a global adaptive average pooling layer, and a fully connected neural network layer; Extract N sampling points of the training set as the first channel input of the feature extractor, perform FFT transformation on the N sampling points, and use the real part and imaginary part after FFT transformation as the second channel and third channel input of the feature extractor respectively; The feature extractor is trained based on the improved supervised contrast loss function, and the trained feature extractor is used to extract features to obtain the training set features; Perform multi-center extraction based on K-means for the training set features of each category, obtain each cluster center of each category as the K center set of such features and the number of points contained in the cluster where the K center is located, and generate point noise around the K center as pseudo features; The pseudo features and the new class features are combined to train a network classifier, wherein the network classifier is a two-layer fully connected neural network classifier, the number of neurons in the input layer is equal to the number of feature dimensions 512, the number of neurons in the middle layer is set to 128, and the number of neurons in the output layer is equal to the current total number of categories; Use the trained feature extractor to extract features from the test set, and input the features of the test set into the trained network classifier for classification to test its recognition accuracy.

2. According to claim 1, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that The method also includes normalizing the input samples of the first channel, the second channel, and the third channel so that the samples satisfy a distribution with a mean of 0 and a standard deviation of 1.

3. According to claim 1, a method for intelligent identification of quasi-incremental radar emitter individuals based on K-center feature distribution reconstruction is characterized in that: After training the feature extractor, the method further includes constructing and training an initial category classifier, wherein the initial category classifier is used to classify the initial category training set.

4. According to claim 1, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that The multi-scale convolution module integrating the multi-head attention mechanism includes: The input data passes through three one-dimensional convolution branches of different scales: Y convk =Conv1D(X,W convk )+b convk Among them, k is the scale of the convolution, Y convk is the output of the convolution branch of scale k, and X is the input; The input also passes through a pooling convolution branch: Y pool =Conv1D(MaxPool1D(X)) Among them, Y pool is the output of the pooled convolution branch, and X is the input; The outputs of each branch are concatenated in the channel direction to obtain the output of multi-scale convolution: AND inception =Concat(And conv1 ,AND conv3 ,AND conv5 ,AND pool ) Among them, Y inception is the output of multi-scale convolution; Apply a multi-head self-attention layer to the output of the multi-scale convolution to obtain the module output Y:

5. According to claim 4, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that: The feature extractor is trained based on an improved supervised contrastive loss function, including: In the supervised contrast loss function training, the sample features of the same type of each sample feature in a batch are used as positive sample pairs, and the sample features of different types are used as negative sample pairs; The similarity Logits between the anchor point and the contrast feature is calculated by the dot product of the feature vector: Among them, 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 similarity gradient; Use the softmax function to calculate the probability distribution of contrast logits: For positive and negative sample pairs, calculate the positive sample loss and negative sample loss respectively: Where P(i) is the positive sample set of sample i; N(i) is the negative sample set of sample i; The difference in the log probability of the positive and negative sample pairs is taken as the final loss, and the average is taken over the entire batch as the loss: L=-(L pos -negative_weight·L neg ) Among them, L is the final loss, and negative_weight is the negative sample weight used to control the importance of negative samples in the loss calculation.

6. According to claim 1, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that The acquisition of the new class features includes: The trained feature extractor is used to extract features from the individual training set samples of the new type of radar emitter to obtain the training set features of the new type.

7. According to claim 1, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that The multi-center extraction of the training set features of each category based on K-means includes: Set the number of clusters k = 10, and the optimization target is the sum of the squares of the distances from all data points to the center of their clusters: Among them, x i is the ith data point, μ j is the jth cluster center, C j is the set of all data points in the jth cluster, ||x i -μ j || 2 is the square of the Euclidean distance; By iteratively performing cluster assignment and center update, the K-means algorithm attempts to minimize the objective function J to divide the data set into 10 optimal clusters.

8. According to claim 1, a method for intelligently identifying individual quasi-incremental radar emitters based on K-center feature distribution reconstruction is characterized in that The pseudo-features follow a normal distribution with K centers as mean and standard deviation σ: X i ~N(c i ,s 2 ) Among them, c i is the i-th K center position.

9. A computer system, characterized in that include: One or more processors, and 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 are enabled to implement the method of claim 1.

10. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images