Radar interference open set identification method based on sustainable learning algorithm

Abstract

The invention discloses a radar interference open-set identification method based on a sustainable learning algorithm, and the method comprises the steps: 1, building an interference time-frequency image open-set identification network based on an OpenMax layer, and 2, carrying out the intra-class classification of a plurality of types of interference time-frequency images in an unknown class through clustering analysis, obtaining the detailed classification information of unknown interference, and carrying out the recognition of the open-set identification network. And 3, creating an interference time-frequency image of unknown category interference through the generative adversarial network, and completing a sustainable image recognition task. According to the method, the classification capability of a radar system on known interference and the identification capability on unknown interference can be improved, so that continuous learning and updating of the system on the interference can be realized, and the method has certain reference significance on the radar interference open set identification direction.

Description

Technical Field

[0001] The invention belongs to the field of radar interference recognition, in particular to a radar interference open set recognition method based on a sustainable learning algorithm. Background Art

[0002] Radar interference identification is an important research direction in the field of radar countermeasures. How to effectively identify interference is an important prerequisite for achieving interference suppression. Only by determining the specific type of interference received can corresponding reasonable anti-interference measures be taken. The general process of interference identification is to preprocess the radar received signal first, then extract the feature parameters of the processed signal in the time-frequency domain and transform domain, and finally input the feature parameters into a suitable classifier for processing to identify the type of interference signal. At present, algorithms for radar interference identification are mainly divided into two categories: traditional algorithms based on feature extraction and classification algorithms based on deep learning models.

[0003] Traditional algorithms based on feature extraction require manual extraction of interference signal features in the time, frequency, time-frequency and other transform domains, and design of classifiers to achieve interference identification. Such algorithms require a lot of prior knowledge, have high computational complexity and limited application scenarios. Classification algorithms based on deep learning models can learn directly from inputs without manual feature extraction. However, they rely heavily on the quality and diversity of training data, and also have high demands on computing resources.

[0004] With the rapid development of electronic countermeasure technology and the emergence of signal processing technology, interference signals in modern electromagnetic environments have become more and more complex and diverse. In practical applications, interference signals of unknown types are often encountered. The current radar interference recognition method mainly solves the closed set classification problem, that is, the interference training set contains all categories of the test set. The existing open set recognition model has limited recognition effect, and will uniformly identify multiple types of unknown interference as unknown classes, causing the model to lose the ability to re-learn unknown interference in the future. In addition, the existing methods require a large amount of labeled sample data and prior knowledge during the model training process, which is difficult to meet in the recognition scenario of unknown interference types. In a complex electromagnetic environment, small sample conditions and unknown interference problems coexist. Considering how to achieve correct identification of unknown interference in small sample conditions has become a challenge that needs to be solved. Summary of the invention

[0005] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a radar interference open set identification method based on a sustainable learning algorithm, so as to effectively identify unknown category interference, thereby enabling the system to continuously learn and update the unknown interference and enhance the interference identification capability and adaptability of the radar system.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The characteristic of the radar interference open set identification method based on the sustainable learning algorithm of the present invention is that it is carried out according to the following steps:

[0008] Step 1: Perform short-time Fourier transform on the interference training signal received by the radar to obtain a training time-frequency image dataset containing N known categories. ,in, Indicates The training time-frequency image set of categories, and , express Middle training time-frequency images; express The total number of training time-frequency images in ;

[0009] Perform short-time Fourier transform on the interference test signal received by the radar to obtain Test time-frequency image dataset of categories ,in, Indicates The test time-frequency image set of categories, and , express Middle Test time-frequency images; express The total number of test time-frequency images in ; Represents the total number of categories in the unknown category unknown; let The true category label is recorded as ;

[0010] Step 2, construct an interference time-frequency image open set recognition network, including: the first layer is composed of a convolution layer, a RELU activation function, a batch normalization layer, a maximum pooling layer and an SA attention mechanism layer; the second to fifth layers are composed of a convolution layer, a RELU activation function and a batch normalization layer, wherein the input of the second layer is residually connected to the batch normalization layer of the third layer; the input of the fourth layer is residually connected to the batch normalization layer of the fifth layer; the sixth layer is composed of a convolution layer, a RELU activation function, a convolution layer, an SA attention mechanism layer, a flattening layer, two fully connected layers and a RELU activation function; the seventh layer is a classification layer;

[0011] Step 2.1: Use the interference time-frequency image open set to identify the first layer of the network during training Perform feature extraction and obtain the Category Differentiated time-frequency features ;

[0012] Step 2.2 It is processed in the subsequent second to sixth layers of the network structure and outputs Class The activation vector of the training time-frequency image ;

[0013] Step 2.3 In the input classification layer, the Softmax classifier is used to output the Category Predicted class labels for interference time-frequency images ;

[0014] Step 2.4, based on and Constructing a cross entropy loss function and using it to perform preliminary training on the interference time-frequency image open set recognition network, and completing the training when the cross entropy loss function tends to be stable, to obtain the interference time-frequency image open set recognition model after preliminary training;

[0015] Step 2.5: During the test, replace the Softmax classifier in the classification layer of the interference time-frequency image open set recognition model after preliminary training with the OpenMax layer, so as to Process and obtain Activation values ​​belonging to unknown class unknown ;

[0016] Step 2.6: According to and calculate Belong to Probability of class and the probability of belonging to the unknown category unknown ;

[0017] if and The maximum value of the two is less than the set threshold Δ or the maximum value of the two is , then It is regarded as an interference test time-frequency image in the unknown category unknown, thereby obtaining an interference test time-frequency image dataset of the unknown category unknown ;

[0018] Step 3: After clustering, we get A test time-frequency image dataset of unknown categories ,in, Indicates A set of test interference time-frequency images of unknown categories;

[0019] Step 4: Establish a generative adversarial network, including a generator and a discriminator, to process test time-frequency images of unknown categories or interfering virtual samples of unknown categories to train the generative adversarial network and obtain a trained generative adversarial model for generating A set of interfering time-frequency images of unknown categories ,in, Indicates A set of interfering time-frequency images of unknown categories;

[0020] Step 5: , and The training is completed when the cross entropy loss function tends to be stable, and the optimal interference time-frequency image open set recognition model after the final training is obtained for recognition. The interference time-frequency image of the class.

[0021] The radar interference open set identification method based on the sustainable learning algorithm described in the present invention is also characterized in that step 2.1 is performed according to the following steps:

[0022] Step 2.1.1. After being processed by the convolution layer, RELU activation function, batch normalization layer, and maximum pooling layer in the first layer, the first Category Training time-frequency features ;

[0023] Step 2.1.2 In the first layer of the SA attention mechanism, the Divide into time-frequency sub-features, among which Time-frequency sub-feature Divided into the first First frequency sub-feature and The second time-frequency sub-feature ;

[0024] Step 2.1.3 The input channel attention branch performs global average pooling to obtain the Channel 1 statistics , and then use formula (1) to get the First channel attention features :

[0025] (1)

[0026] In formula (1), , Separate a 1st scaling parameter and a 1st shift parameter; is the sigmoid activation function;

[0027] Step 2.1.4 The spatial attention branch in the SA attention mechanism layer of the first layer is input and group normalized to obtain the Second space statistics , and then use formula (2) to get the Second spatial attention feature :

[0028] (2)

[0029] In formula (2), , Separate a second scaling parameter and a second shift parameter;

[0030] Step 2.1.5: The aggregation module in the SA attention mechanism layer will and After feature fusion, the fused features are shuffled by the Channel Shuffle operator to obtain the first Category The image Time-frequency fusion features , thus After performing aggregation operations along the channel dimension, we get Category Differentiated time-frequency features .

[0031] Furthermore, step 2.5 is performed as follows:

[0032] Step 2.5.1: Input the first to sixth layers of the interference time-frequency image open set recognition model after preliminary training for processing, and obtain the Category The activation vector of the test time-frequency image ;

[0033] Step 2.5.2: According to Calculate the The average activation vector of the class training time-frequency images , and calculate and The distance between The distance between the activation vector of each training time-frequency image is used to construct the Weibull-like distribution model ;

[0034] Step 2.5.3, according to , calculate the Class Test time-frequency images Belong to Probability of class ;

[0035] Step 2.5.4: Use formula (3) to get Belong to Test activation vector for the class :

[0036] (3)

[0037] In formula (3), represents the Hadamard product;

[0038] Step 2.5.5: Use formula (4) to get Activation values ​​belonging to unknown class unknown :

[0039] (4).

[0040] Further, step 4 is performed as follows:

[0041] Step 4.1. Enter A random noise vector To the generator for processing, the output is Unknown category Interference virtual samples ;

[0042] Step 4.2: Unknown category Test interference time-frequency image and Input the discriminator module, the corresponding output belong Likelihood Rating and belong Likelihood score ;

[0043] Step 4.3: and A set of interfering virtual samples of unknown categories Merge into A mixed dataset of interfering time-frequency images of unknown categories, and extract the A set of interfering time-frequency interpolated images of unknown categories ;

[0044] Step 4.4: Use formula (5) to construct the loss function of the discriminator :

[0045] (5)

[0046] In formula (5), Express expectations, , , Respectively represent the A set of unknown test interference time-frequency images , interference virtual sample set and interference time-frequency interpolation image set Samples of express belong Likelihood Rating right The Euclidean norm of the gradient of , is a hyperparameter; express The Interference time-frequency interpolation image;

[0047] Step 4.5: Use formula (6) to construct the loss function of the generator :

[0048] (7)

[0049] Step 4.6, based on and , the weights of the generator and the discriminator are updated alternately using the RMSprop optimization method, thereby training the generative adversarial network until and When the training is stable, the training is completed and a trained generative adversarial model is obtained;

[0050] Step 4.7: Set the random noise vector Input the generator in the trained generative adversarial model and output A set of interfering time-frequency images of unknown categories ,and ,in, For the Unknown category The interference time-frequency image is generated. For the The total number of interfering time-frequency generated images of unknown categories.

[0051] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the radar interference open set identification method, and the processor is configured to execute the program stored in the memory.

[0052] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program, when executed by a processor, executes the steps of the radar interference open set identification method. Compared with the prior art, the present invention has the following advantages:

[0053] 1. The present invention can effectively identify unknown interference in real time by integrating the three modules of open set recognition, type internal classification and interference image generation, continuously optimize and update the current radar interference recognition model, solve the problem that the existing interference recognition algorithm cannot effectively identify multiple types of unknown interference, and realize the sustainable learning of adaptive radar interference recognition.

[0054] 2. The present invention introduces the SA attention mechanism into the open set recognition network of interference time-frequency images, which can adaptively focus on the differentiated feature information of interference, and uses the OpenMax layer to change the output dimension of the classifier from N to N+1, thereby realizing the recognition of unknown interference and improving the recognition accuracy of the model for unknown interference;

[0055] 3. The type internal classification module of the present invention obtains the classification information of the interference time-frequency image of the unknown category through the clustering algorithm, and uses the generator and discriminator of the generative adversarial network to optimize each other to generate a realistic interference time-frequency generated image set of the unknown category, thus completing the sustainable learning of unknown interference under the concurrent conditions of small sample and open set recognition;

[0056] 4. The present invention can effectively identify 8 common types of suppression interference and deception interference, including most of the interference types currently popular in this field, which shows the effectiveness and universality of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the time-frequency diagram of the target signal and 8 interference signals in the present invention;

[0058] Figure 2 It is the interference time-frequency image open set recognition network diagram of the present invention;

[0059] Figure 3 is a confusion matrix diagram of the interference time-frequency image open set recognition model when the interference-to-noise ratio is 0 dB in the present invention;

[0060] Figure 4It is a confusion matrix diagram of the interference time-frequency image open set recognition model after type internal division when the interference-to-noise ratio is 0 dB in the present invention;

[0061] Figure 5 It is the recognition accuracy of the interference time-frequency image open set recognition model of the present invention for samples containing new unknown interference. DETAILED DESCRIPTION

[0062] In this embodiment, a radar interference open set identification method based on a sustainable learning algorithm is performed in the following steps:

[0063] Step 1: Perform short-time Fourier transform on the interference training signal received by the radar to obtain a training time-frequency image dataset containing N known categories. ,in, Indicates The training time-frequency image set of categories, and , express Middle training time-frequency images; express The total number of training time-frequency images in ;

[0064] Perform short-time Fourier transform on the interference test signal received by the radar to obtain Test time-frequency image dataset of categories ,in, Indicates The test time-frequency image set of categories, and , express Middle Test time-frequency images; express The total number of test time-frequency images in ; Represents the total number of categories in the unknown category unknown; let The true category label is recorded as ; In this example, 8 types of interference are simulated according to the generation mechanisms of linear frequency modulation radars, including noise amplitude modulation interference NAM, noise product interference NP, noise convolution interference NC, comb spectrum interference CS, spectrum dispersion interference SS, slice reconstruction interference CI, intermittent sampling forwarding interference ISR and multiple false target interference MFT. The linear frequency modulation signal is selected as the transmitting signal with a pulse width of 5 μs and a sampling frequency of 200 MHz. According to the general amplitude ratio of the interference signal to the echo signal (1.3~1.5), the interference signal power ratio is set to 3 dB. The specific range of the interference parameters is shown in Table 1. Among them, f0 is the center frequency, B is the bandwidth, and B n is the noise bandwidth, N cis the number of teeth, R s is the sampling rate, N p is the number of pulses, N timeslots is the number of time slots, N s is the number of sampling times, N targets is the target quantity.

[0065] Table 1 Specific range of interference parameters of the present invention

[0066]

[0067] The short-time Fourier transform images of the target signal and 8 types of interference signals generated by the simulation of the parameters in Table 1 are as follows: Figure 1 shown.

[0068] N is 5 (known interferences are NAM, NP, NC, ISR, CS, corresponding to Known Jamming1-5), K is 2 (unknown interferences are SMSP, CI, corresponding to Unknown Jamming1-2), is 800, is 200.

[0069] Step 2: Construct an open set recognition network for interference time-frequency images, such as Figure 2 As shown. It includes: the first layer is composed of a convolution layer, a RELU activation function, a batch normalization layer, a maximum pooling layer and a SA attention mechanism layer; the second to fifth layers are composed of a convolution layer, a RELU activation function and a batch normalization layer, wherein the input of the second layer is residually connected to the batch normalization layer of the third layer; the input of the fourth layer is residually connected to the batch normalization layer of the fifth layer; the sixth layer is composed of a convolution layer, a RELU activation function, a convolution layer, an SA attention mechanism layer, a flattening layer, two fully connected layers and a RELU activation function; the seventh layer is a classification layer;

[0070] Step 2.1: Use the interference time-frequency image open set to identify the first layer of the network during training Perform feature extraction and obtain the Category Differentiated time-frequency features ;

[0071] Step 2.1.1. After being processed by the convolution layer, RELU activation function, batch normalization layer, and maximum pooling layer in the first layer, the first Category Training time-frequency features ;

[0072] Step 2.1.2 In the first layer of the SA attention mechanism, the Divide into time-frequency sub-features, among which Time-frequency sub-feature Divided into the first First frequency sub-feature and The second time-frequency sub-feature In this example, is 8.

[0073] Step 2.1.3 The input channel attention branch performs global average pooling to obtain the Channel 1 statistics , and then use formula (1) to get the First channel attention features :

[0074] (1)

[0075] In formula (1), , Separate a 1st scaling parameter and a 1st shift parameter; is the sigmoid activation function;

[0076] Step 2.1.4 The spatial attention branch in the SA attention mechanism layer of the first layer is input and group normalized to obtain the Second space statistics , and then use formula (2) to get the Second spatial attention feature :

[0077] (2)

[0078] In formula (2), , Separate a second scaling parameter and a second shift parameter;

[0079] Step 2.1.5: The aggregation module in the SA attention mechanism layer will and After feature fusion, the fused features are shuffled by the Channel Shuffle operator to obtain the first Category The image Time-frequency fusion features , thus After performing aggregation operations along the channel dimension, we get Category Differentiated time-frequency features .

[0080] Step 2.2 It is processed in the subsequent second to sixth layers of the network structure and outputs Class The activation vector of the training time-frequency image ;

[0081] Step 2.3 In the input classification layer, the Softmax classifier is used to output the Category Predicted class labels for interference time-frequency images ;

[0082] Step 2.4, based on and A cross entropy loss function is constructed and used to perform preliminary training on the interference time-frequency image open set recognition network, and the training is completed when the cross entropy loss function tends to be stable, so as to obtain the interference time-frequency image open set recognition model after preliminary training; in this example, the learning rate of the interference time-frequency image open set recognition network is 0.0001, and the back propagation algorithm adopts AdamOptimizer.

[0083] Step 2.5: During the test, replace the Softmax classifier in the classification layer of the interference time-frequency image open set recognition model after preliminary training with the OpenMax layer, so as to Process and obtain Activation values ​​belonging to unknown class unknown ;

[0084] Step 2.5.1: Input the first to sixth layers of the interference time-frequency image open set recognition model after preliminary training for processing, and obtain the Category The activation vector of the test time-frequency image ;

[0085] Step 2.5.2: According to Calculate the The average activation vector of the class training time-frequency images , and calculate and The distance between The distance between the activation vector of each training time-frequency image is used to construct the Weibull-like distribution model ;

[0086] Step 2.5.3, according to , calculate the Class Test time-frequency images Belong to Probability of class ;

[0087] Step 2.5.4: Use formula (3) to get Belong to Test activation vector for the class :

[0088] (3)

[0089] In formula (3), Denotes the Hadamard product.

[0090] Step 2.5.5: Use formula (4) to get Activation values ​​belonging to unknown class unknown :

[0091] (4)

[0092] Step 2.6: According to and calculate Belong to Probability of class and the probability of belonging to the unknown category unknown ;

[0093] if and The maximum value of the two is less than the set threshold Δ or the maximum value of the two is , then It is regarded as an interference test time-frequency image in the unknown category unknown, thereby obtaining an interference test time-frequency image dataset of the unknown category unknown In this example, the threshold Δ is 0.9.

[0094] Step 3: After clustering, we get A test time-frequency image dataset of unknown categories ,in, Indicates A set of test interference time-frequency images of unknown categories; in this example, ResNet-18 network is used before clustering After extracting the features of the samples, k-means++ clustering analysis is performed after obtaining the feature point data.

[0095] Step 4: Establish a generative adversarial network, including a generator and a discriminator, to process test time-frequency images of unknown categories or interfering virtual samples of unknown categories to train the generative adversarial network and obtain a trained generative adversarial model for generating A set of interfering time-frequency images of unknown categories ,in, Indicates A set of interfering time-frequency images of unknown categories;

[0096] Step 4.1. Enter A random noise vector To the generator module for processing, the output is Unknown category Interference virtual image ,The goal of the generator is to generate realistic interference virtual images, making it as difficult as possible for the discriminator to distinguish between the interference virtual images and the real test interference time-frequency images;

[0097] Step 4.2: Unknown category Test interference time-frequency image and Input the discriminator module, the corresponding output belong Likelihood Rating and belong Likelihood score , the goal of the discriminator is to determine whether the input image is real or generated by the generator. The discriminator hopes to output a higher score for the real test interference time-frequency image and a lower score for the interference virtual image;

[0098] Step 4.3: and A set of interfering virtual samples of unknown categories Merge into The mixed dataset of interference time-frequency images of unknown categories is used, and the first A set of interfering time-frequency interpolated images of unknown categories :

[0099] (5)

[0100] In formula (5), represents the interpolation coefficient, and ;

[0101] Step 4.4: Use formula (6) to construct the loss function of the discriminator :

[0102] (6)

[0103] In formula (6), express belong Likelihood Rating expectations, among which Indicates that it is taken from A set of unknown test interference time-frequency images Samples of express belong Likelihood Rating expectations, among which Indicates that it is taken from A set of interfering virtual samples of unknown categories Samples of is a hyperparameter, express expectations, among which Indicates that it is taken from A set of interfering time-frequency interpolated images of unknown categories Samples of express belong Likelihood Rating right The Euclidean norm of the gradient of ; in this example, The value of is 10.

[0104] Step 4.5: Use formula (7) to construct the loss function of the generator :

[0105] (7)

[0106] Step 4.6, based on and , the weights of the generator and the discriminator are updated alternately using the RMSprop optimization method, thereby training the generative adversarial network until and When it tends to be stable, the training is completed and a trained generative adversarial model is obtained; In this example, the learning rates of the generator and the discriminator are both 0.0001.

[0107] Step 4.7: Set the random noise vector Input the generator in the trained generative adversarial model and output A set of interfering time-frequency images of unknown categories ,and ,in, For the Unknown category The interference time-frequency image is generated. For the The total number of interfering time-frequency images generated by unknown categories; in this example, Q is 800.

[0108] Step 5: , and The training is completed when the cross entropy loss function tends to be stable, and the optimal interference time-frequency image open set recognition model after the final training is obtained for recognition. In this example, the optimal interference time-frequency image open set recognition model after final training is not only used to identify The time-frequency images of the interference of this type are also used to identify a new type of unknown interference (MFT, corresponding to Unknown Jamming3);

[0109] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0110] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0111] from Figure 3 From the confusion matrix diagram of the open set recognition model of interference time-frequency images, it can be seen that the open set recognition model of interference time-frequency images using OpenMax combined with threshold can accurately classify interference time-frequency images of known categories and effectively identify unknown interference time-frequency images. Figure 4 When it is 0 dB, the confusion matrix diagram of the interference time-frequency image open set recognition model after type internal division. After cluster analysis, information about identifying various unknown types of interference is obtained. Here, the unknown class unknown time-frequency data set is successfully divided into two categories according to the category information of Unknown Jamming1-2. Figure 5It is the recognition accuracy of the interference time-frequency image open set recognition model for samples containing new unknown interference. The interference time-frequency generated images of unknown categories are obtained by using the generative adversarial network and then combined with the training time-frequency image set and the test time-frequency image set and put back into the interference time-frequency image open set recognition network. After the network performs feature learning on Unknown Jamming 1 and 2, it can effectively recognize these two previously unknown interference time-frequency images. The trained interference time-frequency image open set recognition model has strong classification ability and generalization, and can also accurately recognize a new unknown type of interference Unknown Jamming 3.

[0112] In order to verify the effectiveness of the method of the present invention, this embodiment selects three methods based on method type, interference recognition ability, open set recognition ability and continuous learning ability to compare the effects with the method of the present invention. The selected methods are Open Set Recognition model (OSR model), Weighted Ensemble CNN with Transfer Learning (WECNN-TL) and Jamming Recognition Networks (JRNet). According to the experimental results, the results are shown in Table 2:

[0113] Table 2 Experimental results of the method of the present invention and the three selected comparative methods in terms of method type, interference recognition ability, open set recognition ability and continuous learning ability

[0114]

[0115] The experimental results show that compared with the other three methods, the method proposed in this paper can not only more efficiently identify known types of interference images, but also effectively identify unknown interference images, and has the ability to continuously learn unknown types of interference.

Claims

1. A radar interference open set identification method based on a sustainable learning algorithm, characterized in that: The steps are as follows: Step 1: Perform short-time Fourier transform on the interference training signal received by the radar to obtain a training time-frequency image dataset containing N known categories. ,in, Indicates The training time-frequency image set of categories, and , express Middle training time-frequency images; express The total number of training time-frequency images in ; Perform short-time Fourier transform on the interference test signal received by the radar to obtain Test time-frequency image dataset of categories ,in, Indicates The test time-frequency image set of categories, and , express Middle Test time-frequency images; express The total number of test time-frequency images in ; Represents the total number of categories in the unknown category unknown; let The true category label is recorded as ; Step 2, construct an interference time-frequency image open set recognition network, including: the first layer is composed of a convolution layer, a RELU activation function, a batch normalization layer, a maximum pooling layer and an SA attention mechanism layer; the second to fifth layers are composed of a convolution layer, a RELU activation function and a batch normalization layer, wherein the input of the second layer is residually connected to the batch normalization layer of the third layer; the input of the fourth layer is residually connected to the batch normalization layer of the fifth layer; the sixth layer is composed of a convolution layer, a RELU activation function, a convolution layer, an SA attention mechanism layer, a flattening layer, two fully connected layers and a RELU activation function; the seventh layer is a classification layer; Step 2.1: Use the interference time-frequency image open set to identify the first layer of the network during training Perform feature extraction and obtain the Category Differentiated time-frequency features ; Step 2.2 It is processed in the subsequent second to sixth layers of the network structure and outputs Class The activation vector of the training time-frequency image ; Step 2.3 In the input classification layer, the Softmax classifier is used to output the Category Predicted category labels for interference time-frequency images ; Step 2.4, based on and Constructing a cross entropy loss function and using it to perform preliminary training on the interference time-frequency image open set recognition network, and completing the training when the cross entropy loss function tends to be stable, to obtain the interference time-frequency image open set recognition model after preliminary training; Step 2.5: During the test, replace the Softmax classifier in the classification layer of the interference time-frequency image open set recognition model after preliminary training with the OpenMax layer, so as to Process and obtain Activation values ​​belonging to the unknown class unknown ; Step 2.6: According to and calculate Belong to Probability of class and the probability of belonging to the unknown category unknown ; if and The maximum value of the two is less than the set threshold Δ or the maximum value of the two is , then It is regarded as an interference test time-frequency image in the unknown category unknown, thereby obtaining an interference test time-frequency image dataset of the unknown category unknown ; Step 3: After clustering, we get A test time-frequency image dataset of unknown categories ,in, Indicates A set of test interference time-frequency images of unknown categories; Step 4: Establish a generative adversarial network, including a generator and a discriminator, to process unknown category test time-frequency images or unknown category interference virtual samples to train the generative adversarial network and obtain a trained generative adversarial model for generating A set of interfering time-frequency images of unknown categories ,in, Indicates A set of interfering time-frequency generated images of unknown categories; Step 5: , and The training is completed when the cross entropy loss function tends to be stable, and the optimal interference time-frequency image open set recognition model after the final training is obtained for recognition. The interference time-frequency image of the class.

2. The radar interference open set identification method based on the sustainable learning algorithm according to claim 1 is characterized in that: Step 2.1 is performed as follows: Step 2.1.

1. After being processed by the convolution layer, RELU activation function, batch normalization layer, and maximum pooling layer in the first layer, the first Category Training time-frequency features ; Step 2.1.2 In the first layer of the SA attention mechanism, the Divide into time-frequency sub-features, among which Time-frequency sub-feature Divided into the first First frequency sub-feature and The second time-frequency sub-feature ; Step 2.1.3 The input channel attention branch performs global average pooling to obtain the Channel 1 statistics , and then use formula (1) to get the First channel attention features : (1) In formula (1), , Separate a 1st scaling parameter and a 1st shift parameter; is the sigmoid activation function; Step 2.1.4 The spatial attention branch in the SA attention mechanism layer of the first layer is input and group normalized to obtain the Second space statistics , and then use formula (2) to get the Second spatial attention feature : (2) In formula (2), , Separate a second scaling parameter and a second shift parameter; Step 2.1.5: The aggregation module in the SA attention mechanism layer will and After feature fusion, the fused features are shuffled by the Channel Shuffle operator to obtain the first Category The image Time-frequency fusion features , thus After performing aggregation operations along the channel dimension, we get Category Differentiated time-frequency features .

3. The radar interference open set identification method based on a sustainable learning algorithm according to claim 3 is characterized in that: Step 2.5 is performed as follows: Step 2.5.1: Input the first to sixth layers of the interference time-frequency image open set recognition model after preliminary training for processing, and obtain the Category The activation vector of the test time-frequency image ; Step 2.5.2: According to Calculate the The average activation vector of the class training time-frequency images , and calculate and The distance between The distance between the activation vector of each training time-frequency image is used to construct the Weibull-like distribution model ; Step 2.5.3, according to , calculate the Class Test time-frequency images Belong to Probability of class ; Step 2.5.4: Use formula (3) to get Belong to Test activation vector for the class : (3) In formula (3), represents the Hadamard product; Step 2.5.5: Use formula (4) to get Activation values ​​belonging to the unknown class unknown : (4)。 4. The radar interference open set identification method based on the sustainable learning algorithm according to claim 3 is characterized in that: Step 4 is performed as follows: Step 4.

1. Enter A random noise vector To the generator for processing, the output is Unknown category Interference virtual samples ; Step 4.2: Unknown category Test interference time-frequency image and Input the discriminator module, the corresponding output belong Likelihood Rating and belong Likelihood score ; Step 4.3: and A set of interfering virtual samples of unknown categories Merge into A mixed dataset of interfering time-frequency images of unknown categories, and extract the A set of interfering time-frequency interpolated images of unknown categories ; Step 4.4: Use formula (5) to construct the loss function of the discriminator : (5) In formula (5), Express expectations, , , Respectively represent the A set of unknown test interference time-frequency images , interference virtual sample set and interference time-frequency interpolation image set Samples of express belong Likelihood Rating right The Euclidean norm of the gradient of , is a hyperparameter; express The Interference time-frequency interpolation image; Step 4.5: Use formula (6) to construct the loss function of the generator : (7) Step 4.6, based on and , the weights of the generator and the discriminator are updated alternately using the RMSprop optimization method, thereby training the generative adversarial network until and When the training is stable, the training is completed and a trained generative adversarial model is obtained; Step 4.7: Set the random noise vector Input the generator in the trained generative adversarial model and output A set of interfering time-frequency images of unknown categories ,and ,in, For the Unknown category The interference time-frequency image is generated. For the The total number of interfering time-frequency generated images of unknown categories.

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the radar interference open set identification method described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the radar interference open set identification method according to any one of claims 1 to 4 are executed.

Images