A modeling method of a radar jamming recognition network and a radar jamming recognition method

Abstract

The application discloses a radar jamming recognition network modeling method and a radar jamming recognition method, wherein the radar jamming recognition network modeling method comprises the following steps: modeling a radar active jamming data set based on different active jamming signals to obtain a training set; designing a ResNet-SCA network comprising a space attention module and a coordinate attention module based on a ResNet model; wherein the space attention module and the coordinate attention module respectively select important space regions of input feature maps based on different dimensions to enhance the recognition ability of the network to significant regions in an image; training the ResNet-SCA network by using the training set, and using cross-entropy loss and contrast loss to jointly optimize the network during the training process; and using the trained network as a radar jamming recognition network. The radar jamming recognition network constructed by using the method can recognize more radar jamming categories, and has a high recognition accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar jamming recognition, and particularly relates to a modeling method of a radar jamming recognition network and a radar jamming recognition method. BACKGROUND

[0002] As an important detection and recognition technology, radar is widely used in aviation and meteorology. With the continuous development of technology, radar systems are also facing increasingly complex and diversified jamming threats. Only by timely and accurately identifying the jamming pattern can effective countermeasures be taken. Therefore, it is of great significance to study radar jamming recognition.

[0003] Traditional radar jamming recognition methods mainly rely on manual feature extraction and classification recognition through machine learning algorithms such as decision tree and support vector machine (SVM). These methods require tedious analysis and calculation to extract features, and specific feature extraction algorithms need to be designed for different jamming signals, which limits their versatility. In addition, these methods have weak generalization ability and are difficult to adapt to complex radar working environments.

[0004] In the field of radar jamming recognition, deep learning-based methods provide a new way for automatic recognition and classification of active jamming signals through their powerful feature learning ability. These methods mainly include convolutional neural networks (CNN), recursive neural networks (RNN), long short-term memory networks (LSTM), and Bayesian deep learning, which can automatically learn jamming features from raw radar data without human intervention, thereby improving recognition efficiency and accuracy.

[0005] However, existing deep learning methods still have problems such as high model training calculation cost, few types of radar jamming recognition, and low recognition accuracy in radar jamming recognition. SUMMARY

[0006] In order to solve the above problems in the prior art, the present application provides a modeling method of a radar jamming recognition network and a radar jamming recognition method, aiming to realize automatic recognition and classification of multiple radar active jammers and improve the accuracy of radar active jamming recognition. The technical problems to be solved by the present application are solved by the following technical solutions:

[0007] In a first aspect, the present application provides a modeling method of a radar jamming recognition network, comprising:

[0008] modeling a radar active jamming dataset based on different active jamming signals to obtain a training set;

[0009] designing a ResNet-SCA (Residual Network with Spatial Coordinate Attention) network including a spatial attention module and a coordinate attention module based on a ResNet (Residual Network) model; wherein the spatial attention module and the coordinate attention module respectively select important spatial regions of input feature maps based on different dimensions of adaptivity, so as to enhance the recognition ability of the network to significant regions in an image;

[0010] training the ResNet-SCA network by using the training set, and using cross-entropy loss and contrastive loss to jointly optimize the network during the training process, and taking the trained ResNet-SCA network as a radar jamming recognition network.

[0011] In an embodiment of the first aspect of the application, the radar active jamming dataset is modeled based on different active jamming signals, comprising:

[0012] simulating radar active jamming according to the modulation principles of different active jammers, and obtaining a plurality of time-frequency images under each jamming by short-time Fourier transform, thereby obtaining the radar active jamming dataset.

[0013] In an embodiment of the first aspect of the application, the ResNet-SCA network including the spatial attention module and the coordinate attention module is designed based on the ResNet model, specifically comprising:

[0014] based on the ResNet18 network, removing the max-pooling operation in the ResNet18 network, and adding a spatial attention module and a coordinate attention module at the starting position of each basic residual block to constitute the ResNet-SCA network.

[0015] In an embodiment of the first aspect of the application, the spatial attention module specifically processes the input feature map in the following manner:

[0016] respectively calculating the global average and the maximum of the input feature map to capture spatial distribution and saliency information, to obtain first statistical features and second statistical features;

[0017] splicing the first statistical features and the second statistical features in the channel dimension to obtain a first feature tensor;

[0018] The first feature tensor is further extracted through a 7*7 convolution layer; meanwhile, the input feature map is extracted through a 1*1 convolution layer to generate spatial attention weights;

[0019] The extracted spatial features and the spatial attention weights are subjected to element-wise multiplication operation to realize focusing and emphasizing on important spatial regions in the input feature map, and finally spatial attention features are output.

[0020] In an embodiment of the first aspect of the application, the coordinate attention module specifically processes the input feature map in the following manner:

[0021] Global statistical information is independently extracted along the height and width directions of the input feature map respectively by using adaptive average pooling technology, and two statistical tensors of coordinate dimensions are obtained correspondingly;

[0022] The two statistical tensors of coordinate dimensions are spliced in the channel dimension to obtain a second feature tensor;

[0023] The second feature tensor is further refined through a 1*1 convolution layer and a batch normalization layer, and then through an activation function to obtain a third feature tensor;

[0024] The third feature tensor is respectively subjected to two independent 1*1 convolution layers to generate coordinate attention weights in the height and width directions respectively; meanwhile, the input feature map is subjected to feature extraction through a 1*1 convolution layer to obtain a fourth feature tensor;

[0025] The coordinate attention weights and the fourth feature tensor are subjected to element-wise multiplication to realize focusing and emphasizing on important spatial regions in the input feature map, and finally coordinate attention features are output.

[0026] In an embodiment of the first aspect of the application, the function expression of the cross-entropy loss is:

[0027]

[0028] In the formula, L CE represents the cross-entropy loss function, N represents the number of samples in the current batch, M represents the total number of categories, c represents a category, i represents a current predicted sample, y i,c is a variable, if the category c is the true label of the sample i, then y i,c = 1, otherwise y i,c = 0; p i,c represents the probability that the network predicts the sample i to be the category c.

[0029] In an embodiment of the first aspect of the application, the function expression of the contrastive loss is:

[0030]

[0031] In the formula, L CL represents a contrast loss function, T represents a temperature parameter, T0 represents a basic temperature parameter, N represents the number of current batch samples, K represents the number of views, M ij represents a mask matrix, z ij represents a dot product score between anchor point features a i and contrast features c j ik represents a dot product score between anchor point features a i and contrast features c k .

[0032] In a second aspect, the present application provides a radar jamming recognition method, comprising:

[0033] The radar jamming recognition network is constructed by using the modeling method of the radar jamming recognition network provided in the first aspect of the present application.

[0034] The radar active jamming image to be recognized is input into the radar jamming recognition network for processing to obtain a recognition result.

[0035] The present application has the following beneficial effects:

[0036] 1. The modeling method of the radar jamming recognition network provided in the present application, on the one hand, constructs a ResNet-SCA network comprising a spatial attention module and a coordinate attention module, and the spatial attention module and the coordinate attention module can be used to adaptively select the features most useful for the current task, thereby reducing noise and irrelevant information; on the other hand, the ResNet-SCA network is jointly optimized by using a cross-entropy loss and a contrast loss, so that the features finally extracted by the network have class-in compactness and class-out separability. The radar jamming recognition network constructed by using the method can recognize more radar jamming categories and has a high recognition accuracy.

[0037] 2. The radar jamming recognition network is designed by using a network structure based on ResNet in the present application, and the radar jamming recognition network involves fewer convolutional layers, reduces the computational complexity, and reduces the computational cost of model training.

[0038] The present application will be further described in detail below in combination with the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of a modeling method of a radar jamming recognition network provided in an embodiment of the present application;

[0040] Figure 2 is a structural diagram of a ResNet-SCA network provided in an embodiment of the present application.

[0041] ​Figure 3 This is a schematic diagram of the spatial attention module provided in an embodiment of the present invention;

[0042] Figure 4 This is a principle block diagram of the coordinate attention module provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic flowchart of a radar interference identification method provided in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] A first aspect of this invention provides a modeling method for radar interference identification networks. See also... Figure 1 , Figure 1 This is a flowchart illustrating a modeling method for radar interference identification networks provided in an embodiment of the present invention. The method mainly includes the following steps:

[0046] Step 1: Model a radar active jamming dataset based on different active jamming signals to obtain a training set.

[0047] Specifically, this embodiment can simulate radar active interference based on the modulation principle of different active interferences, and obtain multiple time-frequency images under each type of interference through short-time Fourier transform, thereby obtaining a radar active interference dataset.

[0048] Optionally, as one implementation, the radar active jamming signals selected in this embodiment include 8 types of suppression jamming, 6 types of deception jamming, and 35 types of composite jamming, to form a time-frequency image dataset including 49 types of radar active jamming.

[0049] Among them, suppression interference includes noise amplitude modulation interference, noise frequency modulation interference, noise product interference, noise convolution interference, multi-point frequency interference, sine wave modulation sweep frequency interference, sawtooth wave modulation sweep frequency interference, and trapezoidal wave modulation sweep frequency interference; deception interference includes dense false target interference, intermittent sampling and forwarding interference, intermittent sampling and repeated forwarding interference, sampling pulse interference, multiple false target interference, and comb spectrum interference; composite interference is composed of the seven types of suppression interference (excluding noise amplitude modulation interference) and the five types of deception interference (excluding comb spectrum interference) added one by one in sequence.

[0050] Specifically, in the case of 0dB of the dry noise ratio, 1000 samples are randomly generated for each kind of interference, each sample is 128*128 in size, so as to obtain a radar active interference dataset including 1000*49 samples.

[0051] It can be understood that, generally speaking, the radar active interference dataset will be usually divided into a training set and a test set according to a certain proportion (for example, 7:3), the training set is used for network training, and the test set is used for network performance evaluation.

[0052] Step 2: ResNet-SCA network including a spatial attention module and a coordinate attention module is designed based on the ResNet model; wherein the spatial attention module and the coordinate attention module respectively select important spatial regions of input feature maps based on different dimensions to enhance the recognition ability of the network to the significant regions in the image.

[0053] Optionally, as an implementation manner, the radar interference recognition network architecture is constructed based on the ResNet18 network in the embodiment. Since the pooling operation is easy to lose part of the information, the maximum pooling operation in the ResNet18 is selected to be discarded in the embodiment. Meanwhile, the spatial attention module and the coordinate attention module are added at the starting position of each basic residual block of the network to enhance the recognition ability of the network to the significant regions in the image, so as to constitute the ResNet-SCA network as the radar interference recognition network to be trained.

[0054] Specifically, please refer to Figure 2 , Figure 2 is a structural schematic diagram of the ResNet-SCA network provided by the embodiment of the application. The input time-frequency diagram is input into the first convolutional layer after preprocessing to perform preliminary feature extraction, then is sequentially processed through multiple residual blocks, and after the average pooling operation, is divided into two branches, one branch is processed through the full connection layer 128 to calculate the cross-entropy loss and predict the classification category, and the other branch is processed through two full connection layers to calculate the contrast loss. It can be seen that the feature map needs to be extracted through the spatial attention module and the coordinate attention module respectively in each residual block, and then the output features of the two are combined to input the convolutional layer for processing.

[0055] The radar interference recognition network is designed based on the ResNet network structure in the application, the number of involved convolutional layers is small, the calculation complexity is reduced, and the calculation cost of model training is reduced.

[0056] Further, please refer to Figure 3 , Figure 3 is a principle block diagram of the spatial attention module provided by the embodiment of the application. In the embodiment, the spatial attention module specifically processes the input feature map in the following manner:

[0057] First, the global average and maximum values of the input feature map are calculated respectively to capture spatial distribution and saliency information, obtaining a first statistical feature and a second statistical feature.

[0058] Then, the first statistical feature and the second statistical feature are spliced in the channel dimension to obtain a first feature tensor; the tensor contains rich spatial statistical information.

[0059] Next, the first feature tensor is input into a 7x7 convolution layer with an input channel of 2, an output channel of 1, and a padding of 3 to further extract spatial features; at the same time, the input feature map is input into a 1x1 convolution layer to generate spatial attention weights.

[0060] Finally, the extracted spatial features and the spatial attention weights are element-wise multiplied to focus and emphasize important spatial regions in the input feature map, and finally output spatial attention features.

[0061] Further, please refer to Figure 4 , Figure 4 is the principle block diagram of the coordinate attention module provided by the embodiment of the application. In the embodiment, the coordinate attention module specifically processes the input feature map in the following manner:

[0062] First, the adaptive average pooling technology is used to independently extract global statistical information along the height and width directions of the input feature map, respectively, to obtain two coordinate dimension statistical tensors, which represent the distribution characteristics of the feature map in the height and width dimensions, respectively.

[0063] Then, the two coordinate dimension statistical tensors are spliced in the channel dimension to obtain a second feature tensor, which integrates rich spatial statistical information.

[0064] Next, the second feature tensor is input into a 1x1 convolution layer and a batch normalization layer, and then input into an activation function to further refine the feature representation, obtaining a third feature tensor;

[0065] Subsequently, the third feature tensor is input into two independent 1x1 convolution layers to generate coordinate attention weights in the height and width directions, respectively; at the same time, the input feature map is input into a 1x1 convolution layer to extract features, obtaining a fourth feature tensor;

[0066] Finally, the coordinate attention weights and the fourth feature tensor are element-wise multiplied to focus and emphasize important spatial regions in the input feature map, and finally output coordinate attention features.

[0067] The embodiment adds a spatial attention module and a coordinate attention module in a basic residual block of the network, focuses and emphasizes important spatial regions in a feature map, can enhance the spatial feature perception ability of the network, and thus enhances the identification ability of the network to significant regions in an image.

[0068] Step 3: The ResNet-SCA network is trained by using the training set, and the cross-entropy loss and the contrastive loss are used to jointly optimize the network during the training process. The trained ResNet-SCA network is used as the radar jamming recognition network.

[0069] The training set obtained in step 1 is input into the ResNet-SCA network constructed in step 2 for training. The processing of the input image by the ResNet-SCA network can refer to the existing ResNet18 network and the related description in step 2. In addition, during the training process, the cross-entropy loss and the contrastive loss are used to jointly optimize the ResNet-SCA network, so that the network can learn a feature space with compactness within a class and separability between classes.

[0070] Specifically, in the training phase, the input radar jamming time-frequency image is preprocessed. Specifically, the input image is copied and randomly augmented as a positive sample, and then the input and the positive sample are randomly cropped to 112x112 and randomly scaled to obtain the preprocessed image.

[0071] Then, the network constructed in step 2 is used to extract image features. For the extracted image features, the embodiment performs two kinds of processing. The first kind is to obtain a predicted class probability through a fully connected layer, and to optimize using a cross-entropy loss function. The expression of the cross-entropy loss function is:

[0072]

[0073] In the formula, L CE represents the cross-entropy loss function, N represents the number of samples in the current batch, the embodiment can set N to 128, M represents the total number of classes, c represents a class, i represents a current predicted sample, y i,c is a variable, if the class c is the true label of the sample i, then y i,c = 1, otherwise y i,c = 0; p i,c represents the probability that the network predicts the sample i to be the class c.

[0074] The second kind is to use a contrastive loss, which realizes by pulling closer the representations of samples from the same class (positive samples) and pushing away the representations of samples from different classes (negative samples). In the embodiment, all views of the same sample are used as anchor features, and all views in the same batch are used as contrast features, excluding the case of self-contrast, to obtain the contrastive loss:

[0075]

[0076] In the formula, L CL represents a contrast loss function, T represents a temperature parameter, T0 represents a basic temperature parameter, N represents a current batch sample number, that is, a batch size, K represents a view number, M ij represents a mask matrix, if the anchor feature a i is a positive sample pair (that is, from the same category), M j = 1, otherwise, M ij = 0, z ij represents a dot product score between the anchor feature a ij and the contrast feature c i , and correspondingly, z j represents a dot product score between the anchor feature a ik and the contrast feature c i . k

[0077] Optionally, as an implementation manner, T and T0 in the embodiment can be 0.07.

[0078] Finally, the total loss can be obtained as follows:

[0079] L ALL = L CE + L CL .

[0080] It should be noted that the SGD (Stochastic Gradient Descent) optimizer is used to optimize in the training process of the embodiment, the number of training times can be set to 1000 rounds, and the initial learning rate can be set to 0.01.

[0081] Through multiple rounds of training, the trained ResNet-SCA network is obtained, which can be used as a radar jamming recognition network.

[0082] The ResNet-SCA network is optimized by using the cross-entropy loss and the contrast loss in the embodiment, so that the features finally extracted by the network have the intra-class compactness and the inter-class separability, the radar jamming recognition accuracy is improved, and more radar jamming categories can be recognized.

[0083] It can be understood that after the trained ResNet-SCA network is obtained, the following steps are further included:

[0084] The performance of the network is evaluated by using the test set obtained in step 1, and the existing conventional indicators can be selected as the evaluation indicators, and the embodiment does not make a specific introduction.

[0085] ​The radar jamming recognition network modeling method provided by the application can recognize more radar jamming categories and has high recognition accuracy.

[0086] The second aspect of the application also provides a radar jamming recognition method, please refer to Figure 5 , Figure 5 is a flowchart of a radar jamming recognition method provided by an embodiment of the application, and specifically comprises:

[0087] Step one: constructing a radar jamming recognition network by using the radar jamming recognition network modeling method provided by the first aspect of the application;

[0088] Step two: inputting a radar active jamming image to be recognized into the radar jamming recognition network for processing to obtain a recognition result.

[0089] Therefore, the method can also recognize more radar jamming categories and has high recognition accuracy.

[0090] In the description of the present application, the terms "first", "second" are only used for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0091] In addition, the description of the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in the present application.

[0092] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all of them shall be deemed as falling within the protection scope of the present application.

Claims

1. A method of modeling a radar jamming recognition network, characterized by, The application relates to a radar active jamming recognition method based on a ResNet-SCA network. The application comprises the following steps: A radar active jamming dataset is modeled based on different active jamming signals to obtain a training set; A ResNet-SCA network comprising a spatial attention module and a coordinate attention module is designed based on a ResNet model; wherein the spatial attention module and the coordinate attention module respectively select important spatial regions of input feature maps based on different dimensions to enhance the recognition ability of the network to significant regions in images; 2. The method of claim 1, wherein, The ResNet-SCA network is trained by using the training set, and cross-entropy loss and contrast loss are used to optimize the network during the training process, and the trained ResNet-SCA network is used as a radar jamming recognition network. A radar active jamming dataset is modeled based on different active jamming signals, comprising the following steps:

3. The method of claim 1, wherein, Radar active jamming is simulated according to the modulation principles of different active jamming, and a plurality of time-frequency images under each kind of jamming are obtained by short-time Fourier transform, thereby obtaining a radar active jamming dataset. A ResNet-SCA network comprising a spatial attention module and a coordinate attention module is designed based on a ResNet model, comprising the following steps:

4. The method of claim 3, wherein, A ResNet18 network is taken as a basis, a maximum pooling operation in the ResNet18 network is removed, and a spatial attention module and a coordinate attention module are added at the starting position of each basic residual block to form a ResNet-SCA network. The spatial attention module processes the input feature map in the following manner: The global average value and the maximum value of the input feature map are calculated respectively to capture spatial distribution and significance information, thereby obtaining first statistical features and second statistical features; The first statistical features and the second statistical features are spliced in the channel dimension to obtain a first feature tensor; The first feature tensor is further processed by a 7*7 convolution layer to extract spatial features, and the input feature map is processed by a 1*1 convolution layer to generate spatial attention weights; 5. The method of claim 1, wherein, The extracted spatial features and the spatial attention weights are subjected to element-wise multiplication to focus on and emphasize important spatial regions in the input feature map, and finally spatial attention features are output. The coordinate attention module processes the input feature map in the following manner: Global statistical information is independently extracted along the height and width directions of the input feature map by using an adaptive average pooling technology, thereby obtaining two statistical tensors in the coordinate dimension; The two statistical tensors in the coordinate dimension are spliced in the channel dimension to obtain a second feature tensor; The second feature tensor is further refined by a 1*1 convolution layer and a batch normalization layer, and then by an activation function, thereby obtaining a third feature tensor; The third feature tensor is respectively processed by two independent 1*1 convolution layers to generate coordinate attention weights in the height and width directions; meanwhile, the input feature map is processed by a 1*1 convolution layer to obtain a fourth feature tensor; The coordinate attention weight is multiplied with the fourth feature tensor element by element to realize focusing and emphasizing on important spatial regions in the input feature map, and finally output a coordinate attention feature.

6. The method of claim 1, wherein, The function expression of the cross-entropy loss is: wherein L CE represents a cross-entropy loss function, N represents the number of samples in the current batch, M represents the total number of classes, c represents a class, i represents a current predicted sample, y i,c is a variable, y i,c = 1 if the class c is the true label of the sample i, otherwise y i,c = 0; p i,c represents the probability of the network predicting that the sample i is the class c.

7. The method of claim 1, wherein, The function expression of the contrastive loss is: In the formula, L CL represents a contrast loss function, T represents a temperature parameter, T0 represents a base temperature parameter, N represents a current batch sample number, K represents a view number, M ij represents a mask matrix, z ij represents a dot product score between anchor point features a i and contrast features c j . z ik represents a dot product score between anchor point features a i and contrast features c k .

8. A radar jamming identification method, characterized by, The method comprises the steps of: The radar jamming recognition network is constructed by using the modeling method of the radar jamming recognition network according to any one of claims 1-7. The radar active jamming image to be recognized is input into the radar jamming recognition network for processing to obtain a recognition result.

Images