A Method for Identifying Composite Interference Signals of Linear Frequency Modulated Radar Based on a Dual-Channel Gated Attention Mechanism with Large Convolutional Kernels

Abstract

This invention relates to a method for identifying composite interference signals in linear frequency modulated (LFM) radar based on a dual-channel gated attention mechanism with large convolutional kernels, implemented through an LFM radar interference identification system. The method includes: Step S1: Building the composite interference identification system and generating composite interference signals, and constructing a training database through preprocessing; Step S2: Designing and building a deep learning network based on a dual-channel gated attention mechanism with large convolutional kernels, i.e., a composite interference signal identification model, and training it using preprocessed data; Step S3: Evaluating the classification performance of the composite interference signal identification model using a test sample set to obtain the trained composite interference signal identification model; Step S4: Preprocessing the composite interference signal to be identified and then performing signal identification through step 2. The method proposed in this invention achieves the highest recognition accuracy under all JNR (Joint Noise Recognition) conditions.

Description

Technical Field

[0001] This invention relates to a method for identifying composite interference signals in linear frequency modulated radar based on a dual-channel gated attention mechanism with a large convolution kernel, and belongs to the field of linear frequency modulated radar technology. Background Technology

[0002] Radar is a core device for situational awareness of the battlefield environment in modern warfare. As the current mainstream radar technology, linear frequency modulated radar detects target information by transmitting and receiving electromagnetic waves. Due to its large time-bandwidth product, it can achieve high range resolution and velocity measurement accuracy, and is widely used in target detection, tracking guidance, and imaging.

[0003] In the modern information-based electronic warfare environment, the electromagnetic spectrum is becoming increasingly complex, and interference patterns are showing a trend towards being more composite, intelligent, and multi-sourced. Traditional single-signal interference pattern recognition methods (such as intentional or unintentional interference) are no longer sufficient to handle composite interference situations involving multiple superimposed interference signals. Furthermore, composite interference signals exhibit more complex nonlinear characteristics and multi-scale structures in the time, frequency, or time-frequency domains, posing a severe challenge to anti-interference technologies relying on traditional feature extraction and judgment, making it difficult to guarantee recognition accuracy and robustness. Therefore, designing and developing an intelligent linear frequency modulated radar interference recognition method capable of handling complex composite interference signals with high accuracy has become crucial for ensuring normal radar operation and maximizing radar effectiveness.

[0004] In recent years, deep learning technology has developed rapidly, especially convolutional neural networks (CNNs), which have demonstrated powerful feature learning capabilities in the field of complex pattern recognition. This provides a new approach to solving the problem of radar composite interference identification. Applying deep learning methods to radar interference signal identification can bypass the tedious process of manually designing interference signal features in traditional methods. By leveraging the powerful nonlinear fitting capabilities of deep learning, high-dimensional, deep, and discriminative composite features can be automatically extracted from signal data. This application helps improve the intelligence level of radar anti-jamming and its combat effectiveness in complex electromagnetic environments. However, most existing CNN-based interference signal identification methods use fixed, small-sized convolutional kernels, which are limited by a small effective receptive field. This makes it difficult to capture long-range dependencies spanning long time scales in composite interference signals, resulting in insufficient feature extraction. This hinders the maximum effectiveness of linear frequency modulated radar and can even lead to serious consequences.

[0005] Therefore, developing efficient and accurate composite interference signal identification technology for linear frequency modulated radar is crucial to ensuring the efficient and reliable use of linear frequency modulated radar. Summary of the Invention

[0006] To address the current difficulty in identifying composite interference signals in linear frequency modulated (LFM) radar, this invention proposes a method for identifying composite interference signals in LFM radar based on a dual-channel gated attention mechanism with a large convolutional kernel. This method enables in-depth extraction and efficient utilization of composite interference signal features, achieving an accuracy rate of over 95% even under low interference-to-noise ratio (-10 dB).

[0007] This invention constructs a deep learning network that integrates a large convolutional kernel dual-channel gating mechanism and a dual attention mechanism to achieve efficient identification of radar composite interference signals in complex electromagnetic environments.

[0008] First, the one-dimensional time-series composite interference signal is converted into a two-dimensional time-frequency image containing rich time-frequency information using short-time Fourier transform; Next, a parallel asymmetric dual-channel feature extraction structure is designed, which uses large-size convolutional kernels and dilated convolutions to capture the global evolution contour and local detail modulation features of composite interference signals. Furthermore, a gating mechanism is introduced to adaptively adjust the fusion weights of the dual-channel features according to the importance of the interference signal features. Furthermore, by leveraging the complementary properties of coordinate attention and efficient channel attention mechanisms, useless features are suppressed and key feature regions are focused from both spatial and channel dimensions. Finally, the type of interference signal is output through a classification network. This significantly improves the accuracy and robustness of composite interference signal identification under low interference-to-noise ratio conditions.

[0009] The technical solution of the present invention is as follows: A method for identifying composite interference signals of linear frequency modulated radar based on a dual-channel gated attention mechanism with large convolutional kernels is implemented through a linear frequency modulated radar interference identification system. The linear frequency modulated radar jamming identification system includes a target detector and a composite jamming signal identification model. The target detector includes a transmitter and a receiver for transmitting frequency modulated signals and receiving echo signals. The jamming signal originates from a jammer, which generates eight common types of jamming signals: sinusoidal frequency modulation jamming, sinusoidal amplitude modulation jamming, noise frequency modulation jamming, noise amplitude modulation jamming, linear sweep frequency jamming, logarithmic sweep frequency jamming, range deception jamming, and velocity deception jamming. For situations involving multiple jamming signals, the system combines these eight common types of jamming signals in pairs to obtain a total of twenty-eight typical composite jamming signals. The ambient noise is set to additive white Gaussian noise. The composite jamming signal identification model consists of four parts: sample input, dual-core multi-scale feature extraction, dual attention feature enhancement, and classification output. Step S1: Complete the construction of the composite interference identification system and the generation of composite interference signals, and build a training database through preprocessing; Step S2: Design and build a deep learning network based on a dual-channel gated attention mechanism with large convolutional kernels, i.e., a composite interference signal recognition model, and train it using preprocessed data: Step S3: Evaluate the classification performance of the composite interference signal recognition model using a test sample set to obtain the trained composite interference signal recognition model; Step S4: After preprocessing the composite interference signal to be identified, the signal identification is achieved through step 2.

[0010] According to a preferred embodiment of the present invention, the specific implementation process of step S1 includes: For a linear frequency modulated (LFM) radar interference identification system, the transmitter transmits a LFM signal. The expression is: ,in, The amplitude of the transmitted signal. The carrier frequency of the linear frequency modulated signal. μ For frequency modulation slope, t It is a time variable; When the linear frequency modulated signal reaches the target, it is reflected, resulting in an echo signal. The echo signal carries the target's position and velocity information. Represented as: ,in, The amplitude of the echo signal. For the two-way transmission delay between the radar and the target, r The distance between the radar and the target. c The speed of light; The signals received by the receiver include echo signals, composite jamming signals emitted by the jammer, and ambient noise. The expression is: ,in, The first received by the receiver k Such interference signals m The number of types of interference signals. n ( t This is additive white Gaussian noise; To construct a training dataset for a linear frequency modulated radar interference identification system, a composite interference signal is modeled, and the interference-to-noise ratio (JNR) and step size are set. Based on the composite interference signal, a training dataset is constructed, and after shuffling, a portion of it is selected as the training set, while the remainder is used as the validation set. Finally, a short-time Fourier transform is performed on the dataset to obtain the time-frequency distribution image of the interference signal, which is then used as the input to the composite interference signal recognition model; the corresponding time-frequency transform result is... ,in, For the input signal to be analyzed, For tThe sliding window function centered on this.

[0011] More preferably, JNR is defined as the power of the composite interference signal. With noise power The ratio, i.e. The JNR ranges from -20dB to 10dB, with a step size of 2dB.

[0012] A further preferred option is the Hamming window function.

[0013] According to a preferred embodiment of the present invention, the composite interference signal recognition model includes an input layer, a dual-core multi-scale feature extraction layer, a dual complementary attention layer, and a classification output layer; The input layer consists of a 32-channel 3×3 convolutional kernel for simple feature extraction of the input samples; The dual-core multi-scale feature extraction layer includes two parallel feature extraction branches, which process time-frequency features at different scales respectively. A channel gating module is used to weight the two feature branches to achieve adaptive fusion and dynamic selection of features at different scales. The first branch, also known as the large receptive field branch, consists of a 13×13 large convolutional kernel paired with a 3×3 small convolutional kernel. The first branch employs depthwise separable convolution and dilated convolution with an inflation factor of 3 to extract large-scale global features of the signal. The second branch, namely the mid-receptive field branch, consists of a 5×5 medium-sized convolutional kernel paired with a 3×3 small convolutional kernel. The second branch uses dilated convolution with an inflation factor of 2 to extract medium-scale local features of the signal. After processing with large / medium convolutional kernels, both the first and second branches are followed by a standard 3×3 convolutional layer to further refine the feature images and extract small-scale detail features of the interference signals. The feature maps obtained from the two branches are input into the channel gating module to learn and generate adaptive weight vectors for each of the two branches, resulting in high-dimensional features after feature fusion of the two branches; the dual complementary attention layer includes a coordinate attention mechanism module and an efficient channel attention mechanism module set in parallel. The CoordAtt module is used to capture location-sensitive information and long-range dependencies of feature maps in the spatial dimension, enabling focus on key spatial regions; The ECA module is used to capture local cross-channel interaction information along the channel dimension, enabling adaptive enhancement of highly discriminative feature channels; By splicing and fusing the outputs of CoordAtt and ECA, complementary enhancement of spatial and channel domain features is achieved, highlighting the key characteristics of composite interference signals.

[0014] According to a preferred embodiment of the present invention, the specific implementation process of step S2 includes: Step 2.1: Input layer feature mapping; First, the time-frequency distribution image is used as the input data for the composite interference signal recognition model. At the same time, one-hot encoding is performed on 28 typical composite interference signal categories to generate category labels corresponding to each time-frequency distribution image. The time-frequency distribution image and its corresponding category label are combined to form training sample pairs, which are then input into the composite interference signal recognition model for supervised training. The input sample enters the input layer, where a preliminary convolution operation is performed to expand the channel dimension and extract basic features. Step 2.2: Dual-core multi-scale feature extraction; In the first branch, the output of the 13×13 large-size dilated convolution... Represented as: ,in The basic feature map obtained in step 2.1 The size is k × k The convolution kernel weights, d The coefficient of thermal expansion; Simultaneously, features are extracted using DSC, and the output is calculated jointly by depthwise convolution (DWConv) and pointwise convolution (PWConv); the calculation result of DWConv is: ,in p It is a channel index. It is the first p The convolutional kernel weights corresponding to each channel; the calculation result of PWConv is: ,in It is the connection of the first p The input channel and the first k The weights of the 1×1 pointwise convolutional kernel for each output channel; In the second branch, a 5×5 medium-sized dilated convolution is used to focus on local texture details in the time-frequency image and extract medium-scale local features of the signal. After feature extraction, both branches use 3×3 small-size convolution kernels to further refine the feature images and extract small-scale detail features of the interference signals. A channel gating module is introduced to achieve adaptive fusion of features at different scales in the first and second branches; including: First, output feature maps for the first and second branches. and By stitching along the channel dimension, we obtain Obtained through global average pooling operation A global description; Then, a dynamic weight vector is generated using a multilayer perceptron (MLP). gThe expression is: Among them, GAP Indicates to Perform global average pooling. and This is the gate weight matrix, used to learn the nonlinear relationships between channels. and These represent the ReLU and Sigmoid activation functions, respectively. Weight vector g Divide into weighted sub-vectors uniformly along the channel dimension and The expression is: ,Right now Pick g The first half of the elements, Pick g The latter half of the elements, representing the importance weights learned by the network for the two different scale features, are used to calculate the final fused feature map: ,in This indicates multiplication by channel; Step 2.3: Enhancement of dual attention features; Parallel introduction of coordinate attention (CoordAtt) and efficient channel attention (ECA) mechanisms to the feature map output in step 2.2. Perform refined processing: CoordAtt (Coordinate Attention) accurately locates key time-frequency regions of interference signals by capturing feature encodings in both horizontal and vertical directions. The aggregation formulas for horizontal and vertical directional features are as follows: and ,in Indicates the first i Horizontal feature encoding of the row Indicates the first j Vertical feature encoding of columns, H and W Feature maps Height and width; Next, the feature codes in the horizontal and vertical directions are concatenated and transformed using a 1×1 convolution function to obtain the intermediate feature vector. ,in, This indicates a splicing operation along a spatial dimension; Then, the intermediate feature vector f It is further split into two independent feature vectors along the spatial dimension. and The final attention weights are obtained using the following formula. and : , ; in, This represents a 1×1 convolutional layer acting on the feature vectors in the height direction. It is a 1×1 convolutional layer that operates in the width direction; Then, the weights and feature maps are compared. The feature map, enhanced with coordinate attention, is obtained by multiplication using a broadcast mechanism. ; Efficient Channel Attention (ECA) achieves cross-channel interaction through 1D convolution without dimensionality reduction to enhance important channels. The weights are calculated as follows: ,in, Indicates the kernel size as k One-dimensional convolution; Next, the weights With feature map Multiplication yields the feature map after channel attention enhancement. ; Finally, the feature maps output from both CoordAtt and ECA are... and By concatenating and integrating the channels, an enhanced feature map is obtained. Where Concat represents the channel concatenation operation; Step 2.4: Classify and output; Enhanced feature map The input is a classification output layer, which includes a Dropout layer with a dropout rate of 0.1 and a convolutional layer with a kernel size of 1×1; the Dropout layer is used to suppress overfitting. This convolutional layer uses a 1×1 convolutional kernel to process the concatenated high-dimensional feature map. Dimensionality reduction is performed on the channel dimension, mapping it to a classification feature map where the number of channels equals the number of interference signal categories. Classification feature map The GAP layer then aggregates the spatial dimensions, compressing the two-dimensional feature map of each channel into a single numerical vector. Finally, a classification prediction vector with a dimension of 1×28 is output through tensor flattening operation.

[0015] According to a preferred embodiment of the present invention, during the training phase of the composite interference signal recognition model, a cross-entropy loss function is used to measure the difference between the predicted output of the composite interference signal recognition model and the true label. L Represented as: ; in, Indicates the true label, This represents the network's predicted output.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for identifying linear frequency modulated radar composite interference signals based on a dual-channel gated attention mechanism with large convolutional kernels.

[0017] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for identifying linear frequency modulated radar composite interference signals based on a dual-channel gated attention mechanism with large convolutional kernels.

[0018] The beneficial effects of this invention are as follows: The proposed method was compared with other mainstream networks, including VGG16, ResNet18, and STFTCNN. Simulation results show that the proposed method achieved the highest recognition accuracy under all JNR conditions. The advantage of the proposed method is particularly significant at extremely low JNR conditions of -20dB to -10dB. Furthermore, the proposed method achieves a recognition accuracy of over 95% at -10dB, while other mainstream comparison networks under the same conditions only achieve 70% to 90% accuracy. This demonstrates that the proposed method has stronger recognition accuracy and robustness, showing a clear advantage. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the linear frequency modulated radar composite interference signal identification method based on the large convolution kernel dual-channel gated attention mechanism of the present invention; Figure 2 This is a framework diagram of a composite interference signal identification model; Figure 3 This is a comparison chart of the overall recognition accuracy of different methods under different interference-to-noise ratios. Detailed Implementation

[0020] The present invention will be further defined below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0021] Example 1 A method for identifying composite interference signals in linear frequency modulated radar based on a dual-channel gated attention mechanism with large convolutional kernels, such as... Figure 1 As shown, this is achieved through a linear frequency modulated radar interference identification system; The linear frequency modulated radar jamming identification system includes a target detector and a composite jamming signal identification model. The target detector includes a transmitter and a receiver for transmitting frequency modulated signals and receiving echo signals. The jamming signal originates from a jammer, which generates eight common types of jamming signals: sinusoidal frequency modulation jamming, sinusoidal amplitude modulation jamming, noise frequency modulation jamming, noise amplitude modulation jamming, linear sweep frequency jamming, logarithmic sweep frequency jamming, range deception jamming, and velocity deception jamming. For situations involving multiple jamming signals, the system combines these eight common types of jamming signals in pairs to obtain a total of twenty-eight typical composite jamming signals. The ambient noise is set to additive white Gaussian noise. The composite jamming signal identification model consists of four parts: sample input, dual-core multi-scale feature extraction, dual attention feature enhancement, and classification output. Step S1: Complete the construction of the composite interference identification system and the generation of composite interference signals, and build a training database through preprocessing; Step S2: Design and build a deep learning network based on a dual-channel gated attention mechanism with large convolutional kernels, i.e., a composite interference signal recognition model, and train it using preprocessed data: Step S3: Evaluate the classification performance of the composite interference signal recognition model using a test sample set to obtain the trained composite interference signal recognition model; Step S4: After preprocessing the composite interference signal to be identified, the signal identification is achieved through step 2.

[0022] Example 2 The linear frequency modulated radar composite interference signal identification method based on a large convolutional kernel dual-channel gated attention mechanism as described in Example 1 According to a preferred embodiment of the present invention, the specific implementation process of step S1 includes: For linear frequency modulated (LFM) radar interference identification systems, the transmitter's transmitted waveform is typically a LFM signal. The expression is: ,in, The amplitude of the transmitted signal. The carrier frequency of the linear frequency modulated signal. μ For frequency modulation slope, t It is a time variable; When the linear frequency modulated signal reaches the target, it is reflected, resulting in an echo signal. The echo signal carries the target's position and velocity information. Represented as: ,in, The amplitude of the echo signal. For the two-way transmission delay between the radar and the target, r The distance between the radar and the target. c The speed of light; The signals received by the receiver include echo signals, composite jamming signals emitted by the jammer, and ambient noise. The expression is: ,in, The first received by the receiver k Such interference signals m The number of types of interference signals. n ( t This is additive white Gaussian noise; the main consideration here is... m The typical case of 2 is when the receiver is affected by composite interference containing two different interference signals.

[0023] A training dataset for a linear frequency modulated radar jamming identification system is constructed. Composite jamming signals are modeled, and the interference-to-noise ratio (JNR) and step size are set. Based on the composite jamming signals, a portion of the dataset is shuffled and used as the training set, with the remainder serving as the validation set. Based on the aforementioned jamming signal configuration, a total of 28 classes of composite jamming signals can be generated, with 500 samples generated for each class at each JNR, totaling 2.24 × 10⁻⁶ samples. For each sample, the dataset is randomly shuffled, and 80% of it is selected as the training set, while the remaining 20% ​​is used as the validation set.

[0024] Finally, a Short Time Fourier Transform (STFT) is performed on the dataset to obtain the time-frequency distribution image of the interference signal, which is then used as the input to the composite interference signal identification model; the corresponding time-frequency transform result is... ,in, The input signal to be analyzed is a one-dimensional time-series signal sequence containing composite interference signals received by the receiver. For t The sliding window function centered on this.

[0025] JNR is defined as the power of the composite interference signal. With noise power The ratio, i.e. The JNR ranges from -20dB to 10dB, with a step size of 2dB.

[0026] The sliding window function is a Hamming window function.

[0027] like Figure 2 As shown, the composite interference signal recognition model includes an input layer, a dual-core multi-scale feature extraction layer, a dual complementary attention layer, and a classification output layer; The input layer consists of a 32-channel 3×3 convolutional kernel for simple feature extraction of the input samples; The dual-core multi-scale feature extraction layer includes two parallel feature extraction branches, which process time-frequency features at different scales respectively. A channel gating module is used to weight the two feature branches to achieve adaptive fusion and dynamic selection of features at different scales. The first branch, also known as the large receptive field branch, consists of a 13×13 large convolutional kernel paired with a 3×3 small convolutional kernel. To reduce computational cost, the first branch employs depthwise separable convolution (DSC) and dilated convolution with an inflation factor of 3 to extract large-scale global features of the signal. The second branch, namely the mid-receptive field branch, consists of a 5×5 medium-sized convolutional kernel paired with a 3×3 small convolutional kernel. The second branch uses dilated convolution with a dilation factor of 2 to focus on local texture details in the time-frequency image and is used to extract medium-scale local features of the signal. After processing with large / medium convolutional kernels, both the first and second branches are followed by a standard 3×3 convolutional layer to further refine the feature images and extract small-scale detail features of the interference signals. The feature maps obtained from the two branches are input into the channel gating module to learn and generate adaptive weight vectors for each of the two branches, resulting in high-dimensional features after feature fusion of the two branches; the dual complementary attention layer includes a coordinate attention mechanism (CoordAtt) module and an efficient channel attention mechanism (ECA) module set in parallel; The CoordAtt module is used to capture location-sensitive information and long-range dependencies of feature maps in the spatial dimension, enabling focus on key spatial regions; The ECA module is used to capture local cross-channel interaction information along the channel dimension, enabling adaptive enhancement of highly discriminative feature channels; By splicing and fusing the outputs of CoordAtt and ECA, complementary enhancement of spatial and channel domain features is achieved, highlighting the key characteristics of composite interference signals.

[0028] The specific implementation process of step S2 includes: Step 2.1: Input layer feature mapping; First, the time-frequency distribution image is used as the input data for the composite interference signal recognition model. At the same time, one-hot encoding is performed on 28 typical composite interference signal categories to generate category labels corresponding to each time-frequency distribution image. The time-frequency distribution image and its corresponding category label are combined to form training sample pairs, which are then input into the composite interference signal recognition model for supervised training. The input sample enters the input layer, where a preliminary convolution operation is performed to expand the channel dimension and extract basic features. Step 2.2: Dual-core multi-scale feature extraction; In the first branch, the output of the 13×13 large-size dilated convolution... Represented as: ,in The basic feature map obtained in step 2.1 The size is k × k The convolution kernel weights, d The coefficient of thermal expansion; Simultaneously, features are extracted using DSC, and the output is calculated jointly by depthwise convolution (DWConv) and pointwise convolution (PWConv); the calculation result of DWConv is: ,in p It is a channel index. It is the first p The convolutional kernel weights corresponding to each channel; the calculation result of PWConv is: ,in It is the connection of the first p The input channel and the first k The weights of the 1×1 pointwise convolutional kernels for each output channel are used; with the help of large convolutional kernels, the network can establish a broad effective receptive field, thereby extracting large-scale global features of interference signals. By combining depthwise separable convolution with dilated convolution, the number of network parameters and computational complexity can be significantly reduced while keeping the receptive field unchanged, and long-distance time-frequency dependencies can be efficiently captured.

[0029] In the second branch, similar to the first branch, a 5×5 medium-sized dilated convolution is used to focus on local texture details in the time-frequency image and extract medium-scale local features of the signal. After feature extraction, both branches use 3×3 small-size convolution kernels to further refine the feature images and extract small-scale detail features of the interference signals. A channel gating module is introduced to achieve adaptive fusion of features at different scales in the first and second branches; including: First, output feature maps for the first and second branches. and By stitching along the channel dimension, we obtain Obtained through Global Average Pooling (GAP) operation A global description; Then, a dynamic weight vector is generated using a multilayer perceptron (MLP). g The expression is: Among them, GAP Indicates to Perform global average pooling. and This is the gate weight matrix, used to learn the nonlinear relationships between channels. and These represent the ReLU and Sigmoid activation functions, respectively. Weight vector g Divide into weighted sub-vectors uniformly along the channel dimension and The expression is: ,Right now Pick g The first half of the elements, Pick g The latter half of the elements, representing the importance weights learned by the network for the two different scale features, are used to calculate the final fused feature map: ,in This indicates multiplication by channel; The aforementioned dual-core multi-scale feature extraction layer is repeated four times in the network, and the number of output channels for each layer is set to 32, 64, 128 and 128 respectively.

[0030] Step 2.3: Enhancement of dual attention features; In the deep feature enhancement stage, coordinate attention (CoordAtt) and efficient channel attention (ECA) mechanisms are introduced in parallel to enhance the feature map output in step 2.2. Perform refined processing: CoordAtt (Coordinate Attention) accurately locates key time-frequency regions of interference signals by capturing feature encodings in both horizontal and vertical directions. The aggregation formulas for horizontal and vertical directional features are as follows: and ,in Indicates the first i Horizontal feature encoding of the row Indicates the first j Vertical feature encoding of columns, H and W Feature maps Height and width; Next, the feature codes in the horizontal and vertical directions are concatenated and transformed using a 1×1 convolution function to obtain the intermediate feature vector. ,in, This indicates a splicing operation along a spatial dimension; Then, the intermediate feature vector f It is further split into two independent feature vectors along the spatial dimension. and The final attention weights are obtained using the following formula. and : , ; in, This represents a 1×1 convolutional layer acting on the feature vectors in the height direction. It is a 1×1 convolutional layer that operates in the width direction; Then, the weights and feature maps are compared. The feature map, enhanced with coordinate attention, is obtained by multiplication using a broadcast mechanism. ; Efficient Channel Attention (ECA) achieves cross-channel interaction through 1D convolution without dimensionality reduction to enhance important channels. The weights are calculated as follows: ,in, Indicates the kernel size as k One-dimensional convolution; Next, the weights With feature map Multiplication yields the feature map after channel attention enhancement. ; Finally, the feature maps output from both CoordAtt and ECA are... and By concatenating and integrating the channels, an enhanced feature map is obtained. Concat represents the channel splicing operation; by splicing and fusing the output features of CoordAtt and ECA, complementary enhancement of spatial domain and channel domain features is achieved, highlighting the key features of composite interference signals.

[0031] Step 2.4: Classify and output; Enhanced feature map The input is a classification output layer, which includes a Dropout layer with a dropout rate of 0.1 and a convolutional layer with a kernel size of 1×1; the Dropout layer is used to suppress overfitting. This convolutional layer uses a 1×1 convolutional kernel to process the concatenated high-dimensional feature map. Dimensionality reduction is performed on the channel dimension, mapping it to a classification feature map where the number of channels equals the number of interference signal categories. Classification feature map The GAP layer then aggregates the spatial dimensions, compressing the two-dimensional feature map of each channel into a single numerical vector. Finally, a classification prediction vector with a dimension of 1×28 is output through tensor flattening operation.

[0032] During the training phase of the composite interference signal recognition model, the cross-entropy loss function is used to measure the difference between the predicted output of the composite interference signal recognition model and the true label. L Represented as: ; in, Indicates the true label, This represents the network's predicted output. The network uses the Adam optimizer to dynamically update the network weights. The training hyperparameters are set as follows: initial learning rate of 0.001, batch size of 64, and number of iterations of 50. After parameter settings, supervised training is performed on the network using the training set. The parameters are continuously optimized through backpropagation until the loss function converges, thereby achieving accurate classification of composite interference signals.

[0033] The network classification performance is evaluated using a test sample set, including: First, a test dataset independent of the training set was constructed, with the JNR variation range set to -20dB to 10dB and a step size of 2dB, thus dividing the entire test set into 16 test cases with different JNRs. Under each JNR, 100 simulated test samples were generated for each type of composite interference signal, resulting in 2.8× [samples / test cases] per JNR. The total number of samples in the test dataset is 4.48 × 10⁻⁶. A set of samples were then used. Subsequently, test samples under different JNR conditions were input into the constructed and trained deep learning network to test the network's classification ability. The network's recognition accuracy was obtained by comparing the predicted labels output by the network with the true labels of the samples.

[0034] Figure 3 This chart compares the overall recognition accuracy of different methods under different interference-to-noise ratios (JNRs). The proposed method is compared with other mainstream networks, including VGG16, ResNet18, and STFTCNN. Simulation results show that the proposed method achieves the highest recognition accuracy under all JNR conditions. The advantage of the proposed method is particularly significant at extremely low JNRs of -20dB to -10dB. Furthermore, the proposed method achieves over 95% recognition accuracy at -10dB, while other mainstream comparison networks under the same conditions only achieve 70% to 90% accuracy. This demonstrates that the proposed method has stronger recognition accuracy and robustness, showing a clear advantage.

[0035] Example 3 A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the linear frequency modulated radar composite interference signal identification method based on a large convolutional kernel dual-channel gating attention mechanism as described in Embodiment 1 or 2.

[0036] Example 4 A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the linear frequency modulated radar composite interference signal identification method based on a large convolutional kernel dual-channel gated attention mechanism as described in Embodiment 1 or 2.

Claims

1. A method for identifying composite interference signals of linear frequency modulated radar based on a dual-channel gated attention mechanism with large convolutional kernels, characterized in that, This is achieved through a linear frequency modulated radar interference identification system. The linear frequency modulated radar jamming identification system includes a target detector and a composite jamming signal identification model. The target detector includes a transmitter and a receiver for transmitting frequency modulated signals and receiving echo signals. The jamming signals originate from a jammer, which generates eight common types of jamming signals, including sinusoidal frequency modulation jamming, sinusoidal amplitude modulation jamming, noise frequency modulation jamming, noise amplitude modulation jamming, linear sweep frequency jamming, logarithmic sweep frequency jamming, range deception jamming, and velocity deception jamming. For situations where multiple jamming signals coexist, the system combines these eight common types of jamming signals in pairs to obtain a total of twenty-eight typical composite jamming signals. The ambient noise is set to additive white Gaussian noise. The composite jamming signal identification model consists of four parts: sample input, dual-core multi-scale feature extraction, dual attention feature enhancement, and classification output. include: Step S1: Complete the construction of the composite interference identification system and the generation of composite interference signals, and build a training database through preprocessing; Step S2: Design and build a deep learning network based on a dual-channel gated attention mechanism with large convolutional kernels, i.e., a composite interference signal recognition model, and train it using preprocessed data: Step S3: Evaluate the classification performance of the composite interference signal recognition model using a test sample set to obtain the trained composite interference signal recognition model; Step S4: After preprocessing the composite interference signal to be identified, the signal identification is achieved through step 2.

2. The method for identifying linear frequency modulated radar composite interference signals based on a dual-channel gated attention mechanism with large convolutional kernels according to claim 1, characterized in that, The specific implementation process of step S1 includes: For a linear frequency modulated (LFM) radar interference identification system, the transmitter transmits a LFM signal. The expression is: ,in, The amplitude of the transmitted signal. The carrier frequency of the linear frequency modulated signal. μ For frequency modulation slope, t It is a time variable; When the linear frequency modulated signal reaches the target, it is reflected, resulting in an echo signal. The echo signal carries the target's position and velocity information. Represented as: ,in, The amplitude of the echo signal. For the two-way transmission delay between the radar and the target, r The distance between the radar and the target. c The speed of light; The signals received by the receiver include echo signals, composite jamming signals emitted by the jammer, and ambient noise. The expression is: ,in, The first received by the receiver k Such interference signals m The number of types of interference signals. n ( t This is additive white Gaussian noise; To construct a training dataset for a linear frequency modulated radar interference identification system, a composite interference signal is modeled, and the interference-to-noise ratio (JNR) and step size are set. Based on the composite interference signal, a training dataset is constructed, and after shuffling, a portion of it is selected as the training set, while the remainder is used as the validation set. Finally, a short-time Fourier transform is performed on the dataset to obtain the time-frequency distribution image of the interference signal, which is then used as the input to the composite interference signal recognition model; the corresponding time-frequency transform result is... ,in, For the input signal to be analyzed, For t The sliding window function centered on this.

3. The method for identifying composite interference signals of linear frequency modulated radar based on a dual-channel gated attention mechanism with a large convolutional kernel as described in claim 2, characterized in that, JNR is defined as the power of the composite interference signal. With noise power The ratio, i.e. The JNR ranges from -20dB to 10dB, with a step size of 2dB.

4. The method for identifying linear frequency modulated radar composite interference signals based on a dual-channel gated attention mechanism with a large convolutional kernel as described in claim 2, characterized in that, The sliding window function is a Hamming window function.

5. The method for identifying linear frequency modulated radar composite interference signals based on a dual-channel gated attention mechanism with a large convolutional kernel as described in claim 1, characterized in that, The composite interference signal recognition model includes an input layer, a dual-core multi-scale feature extraction layer, a dual complementary attention layer, and a classification output layer; The input layer consists of a 32-channel 3×3 convolutional kernel for simple feature extraction of the input samples; The dual-core multi-scale feature extraction layer includes two parallel feature extraction branches, which process time-frequency features at different scales respectively. A channel gating module is used to weight the two feature branches to achieve adaptive fusion and dynamic selection of features at different scales. The first branch, also known as the large receptive field branch, consists of a 13×13 large convolutional kernel paired with a 3×3 small convolutional kernel. The first branch employs depthwise separable convolution and dilated convolution with an inflation factor of 3 to extract large-scale global features of the signal. The second branch, namely the mid-receptive field branch, consists of a 5×5 medium-sized convolutional kernel paired with a 3×3 small convolutional kernel. The second branch uses dilated convolution with an inflation factor of 2 to extract medium-scale local features of the signal. After processing with large / medium convolutional kernels, both the first and second branches are followed by a standard 3×3 convolutional layer to further refine the feature images and extract small-scale detail features of the interference signals. The feature maps obtained from the two branches are input into the channel gating module to learn and generate adaptive weight vectors for each of the two branches, resulting in high-dimensional features after feature fusion of the two branches; the dual complementary attention layer includes a coordinate attention mechanism module and an efficient channel attention mechanism module set in parallel. The CoordAtt module is used to capture location-sensitive information and long-range dependencies of feature maps in the spatial dimension, enabling focus on key spatial regions; The ECA module is used to capture local cross-channel interaction information along the channel dimension, enabling adaptive enhancement of highly discriminative feature channels; By splicing and fusing the outputs of CoordAtt and ECA, complementary enhancement of spatial and channel domain features is achieved, highlighting the key characteristics of composite interference signals.

6. The method for identifying composite interference signals of linear frequency modulated radar based on a dual-channel gated attention mechanism with a large convolutional kernel as described in claim 1, characterized in that, The specific implementation process of step S2 includes: Step 2.1: Input layer feature mapping; First, the time-frequency distribution image is used as the input data for the composite interference signal recognition model. At the same time, one-hot encoding is performed on 28 typical composite interference signal categories to generate category labels corresponding to each time-frequency distribution image. The time-frequency distribution image and its corresponding category label are combined to form training sample pairs, which are then input into the composite interference signal recognition model for supervised training. The input sample enters the input layer, where a preliminary convolution operation is performed to expand the channel dimension and extract basic features. Step 2.2: Dual-core multi-scale feature extraction; In the first branch, the output of the 13×13 large-size dilated convolution... Represented as: ,in The basic feature map obtained in step 2.1 The size is k × k The convolution kernel weights, d The coefficient of thermal expansion; Simultaneously, features are extracted using DSC, and the output is calculated jointly by depthwise convolution (DWConv) and pointwise convolution (PWConv); the calculation result of DWConv is: ,in p It is a channel index. It is the first p The convolutional kernel weights corresponding to each channel; the calculation result of PWConv is: ,in It is the connection of the first p The input channel and the first k The weights of the 1×1 pointwise convolutional kernel for each output channel; In the second branch, a 5×5 medium-sized dilated convolution is used to focus on local texture details in the time-frequency image and extract medium-scale local features of the signal. After feature extraction, both branches use 3×3 small-size convolution kernels to further refine the feature images and extract small-scale detail features of the interference signals. A channel gating module is introduced to achieve adaptive fusion of features at different scales in the first and second branches; including: First, output feature maps for the first and second branches. and By stitching along the channel dimension, we obtain Obtained through global average pooling operation A global description; Then, a dynamic weight vector is generated using a multilayer perceptron (MLP). g The expression is: Among them, GAP Indicates to Perform global average pooling. and This is the gate weight matrix, used to learn the nonlinear relationships between channels. and These represent the ReLU and Sigmoid activation functions, respectively. Weight vector g Divide into weighted sub-vectors uniformly along the channel dimension and The expression is: ,Right now Pick g The first half of the elements, Pick g The latter half of the elements, representing the importance weights learned by the network for the two different scale features, are used to calculate the final fused feature map: ,in This indicates multiplication by channel; Step 2.3: Enhancement of dual attention features; Parallel introduction of coordinate attention (CoordAtt) and efficient channel attention (ECA) mechanisms to the feature map output in step 2.

2. Perform refined processing: CoordAtt (Coordinate Attention) accurately locates key time-frequency regions of interference signals by capturing feature encodings in both horizontal and vertical directions. The aggregation formulas for horizontal and vertical directional features are as follows: and ,in Indicates the first i Horizontal feature encoding of the row Indicates the first j Vertical feature encoding of columns, H and W Feature maps Height and width; Next, the feature codes in the horizontal and vertical directions are concatenated and transformed using a 1×1 convolution function to obtain the intermediate feature vector. ,in, This indicates a splicing operation along a spatial dimension; Then, the intermediate feature vector f It is further split into two independent feature vectors along the spatial dimension. and The final attention weights are obtained using the following formula. and : ， ； in, This represents a 1×1 convolutional layer acting on the feature vectors in the height direction. It is a 1×1 convolutional layer that operates in the width direction; Then, the weights and feature maps are compared. The feature map, enhanced with coordinate attention, is obtained by multiplication using a broadcast mechanism. ; Efficient Channel Attention (ECA) achieves cross-channel interaction through 1D convolution without dimensionality reduction to enhance important channels. The weights are calculated as follows: ,in, Indicates the kernel size as k One-dimensional convolution; Next, the weights With feature map Multiplication yields the feature map after channel attention enhancement. ; Finally, the feature maps output from both CoordAtt and ECA are... and By concatenating and integrating the channels, an enhanced feature map is obtained. Where Concat represents the channel concatenation operation; Step 2.4: Classify and output; Enhanced feature map The input is a classification output layer, which includes a Dropout layer with a dropout rate of 0.1 and a convolutional layer with a kernel size of 1×1; the Dropout layer is used to suppress overfitting. This convolutional layer uses a 1×1 convolutional kernel to process the concatenated high-dimensional feature map. Dimensionality reduction is performed on the channel dimension, mapping it to a classification feature map where the number of channels equals the number of interference signal categories. Classification feature map The GAP layer then aggregates the spatial dimensions, compressing the two-dimensional feature map of each channel into a single numerical vector. Finally, a classification prediction vector with a dimension of 1×28 is output through tensor flattening operation.

7. The method for identifying composite interference signals of linear frequency modulated radar based on a dual-channel gated attention mechanism with a large convolutional kernel as described in claim 1, characterized in that, During the training phase of the composite interference signal recognition model, the cross-entropy loss function is used to measure the difference between the predicted output of the composite interference signal recognition model and the true label. L Represented as: ； in, Indicates the true label, This represents the network's predicted output.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the linear frequency modulated radar composite interference signal identification method based on a dual-channel gated attention mechanism with large convolution kernels as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the linear frequency modulated radar composite interference signal identification method based on a dual-channel gated attention mechanism with large convolutional kernels as described in any one of claims 1-7.

Images