Radar signal modulation identification method, device, electronic equipment and storage medium

Abstract

The invention discloses a method, device, electronic device and storage medium for radar signal modulation recognition, which belongs to the technical field of radar signal modulation recognition. The method comprises the following steps: step 1, obtaining radar signals and generating a variety of radar modulation signals; step 2, performing improved multiple synchronous compression transform IMSST time-frequency analysis processing on the radar modulation signals, converting them into time-frequency images and forming a data set; step 3, building a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF; step 4, using the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain a radar signal modulation recognition model. The invention not only realizes the modulation recognition of radar signals, but also improves the modulation recognition performance of radar signals, and solves the problem of difficulty in distinguishing multiple modulation signals in complex electromagnetic environments.

Description

Technical Field

[0001] The present invention relates to a method, a device, an electronic device and a storage medium for radar signal modulation recognition, and belongs to the technical field of radar signal modulation recognition. Background Art

[0002] Radar signal modulation recognition (RSMR) is one of the key technologies in modern electronic warfare systems. It involves the automatic classification or recognition of the modulation patterns of received non-cooperative radar signals and plays a vital role in electronic support systems. Accurate identification of radar signal modulation patterns helps in situational awareness and threat assessment in complex electromagnetic environments.

[0003] Radar signal modulation recognition faces many challenges in complex electromagnetic environments and low signal-to-noise ratio (SNR) conditions. Multiple interference signals are mixed with target signals, causing signal features to become blurred and difficult to extract. In a low SNR environment, the energy of the target signal is similar to the noise, further reducing the reliability of recognition. Traditional recognition methods rely on basic feature parameters such as carrier frequency, arrival time, pulse width, and pulse amplitude. These methods based on manual feature extraction are difficult to adapt to complex environments, and the recognition accuracy is easily affected by interference. At the same time, the computational complexity is high.

[0004] With the development of deep learning, more and more studies have tried to apply convolutional neural networks (CNNs) and self-attention mechanisms to the field of radar signal modulation recognition. CNN performs well in extracting time-frequency features, but it is difficult to capture long-distance dependencies when processing high-resolution and long sequence inputs. To solve these problems, the self-attention mechanism was introduced. ViT[Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929,2020.] is a typical representative of the self-attention mechanism. It realizes global feature capture through a multi-head self-attention mechanism and demonstrates its advantages through parallel computing, gradually showing its superiority in signal processing tasks. The disadvantage of self-attention is its high computational complexity, especially when processing long sequences or high-resolution images, its computational and memory consumption will increase significantly. In order to overcome the limitations of Transformer in processing long sequences, Mamba[Albert Gu and Tri Dao. Mamba: Linear-time sequence modelingwith selective state spaces. arXiv preprint arXiv:2312.00752, 2023.] comprehensively selected the qualitative space model (SSM)[Rudolph Emil Kalman. A new approach to linear filteringand prediction problems. 1960.]), which can flexibly propagate or forget information according to the changes of the input signal by dynamically adjusting parameters, which not only improves the ability to process long sequences but also maintains linear complexity. However, in the radar signal time-frequency image processing scenario, Mamba's core mechanism still has significant limitations: its scanning mechanism designed for one-dimensional sequences is difficult to effectively model the two-dimensional spatial correlation of the time-frequency graph, resulting in insufficient ability to capture key features such as the instantaneous frequency change and frequency hopping mode of the signal; the global information propagation characteristics are easily affected by noise in a low signal-to-noise ratio environment, and lack the ability to adaptively model the non-rigid deformation caused by multipath effects and Doppler frequency shift in the time-frequency graph. These problems restrict the accurate recognition of radar signals in complex electromagnetic environments. Therefore, the present invention proposes a method for radar signal modulation recognition based on compound deformable convolution and Mamba's multi-branch multi-view feature extraction network (DCMNet), which can give full play to the synergistic capabilities of dynamic deformation perception and global sequence modeling and improve the accuracy of radar signal recognition. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method, device, electronic device and storage medium for radar signal modulation recognition, which can improve the modulation recognition performance of radar signals and solve the problem of difficulty in distinguishing multiple modulated signals in complex electromagnetic environments.

[0006] The technical solution adopted by the present invention to solve the technical problem is:

[0007] In a first aspect, an embodiment of the present invention provides a method for radar signal modulation recognition, comprising the following steps:

[0008] Step 1, acquiring radar signals and generating a variety of radar modulation signals;

[0009] Step 2: Perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, convert it into a time-frequency image and form a data set;

[0010] Step 3: Build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF;

[0011] Step 4: Use the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain the radar signal modulation recognition model.

[0012] As a possible implementation of this embodiment, the step 1 of acquiring a radar signal and generating a plurality of radar modulation signals includes:

[0013] Collection time t Radar signal:

[0014] ,

[0015] in, represents the amplitude envelope of the signal, represents the instantaneous frequency, represents the instantaneous phase, is additive noise, usually modeled as zero-mean Gaussian white noise;

[0016] Use MATLAB to generate a variety of radar modulation signals, including LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank.

[0017] As a possible implementation of this embodiment, the step 2, performing an improved multiple synchronous compression transform IMSST time-frequency analysis process on the radar modulation signal to convert it into a time-frequency image, includes:

[0018] The radar modulation signal is transformed into a time-frequency image by using an improved multi-scale time-frequency analysis method. :

[0019] ,

[0020] in, and represents discrete frequency, express n The discrete short-time Fourier transform at is the Kronecker delta function, represents two rounding operations on the instantaneous frequency of the multi-synchronous compression transform, Indicates the number of iterations.

[0021] As a possible implementation of this embodiment, the step 3 is to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF, including:

[0022] The multi-branch multi-view feature extraction network DCMNet is designed as a whole to form a DCMNet overall framework structure, which includes a convolutional embedding layer, a stacked progressive residual convolution structure and a hybrid convolution-SSM module; the convolutional embedding layer, the stacked progressive residual convolution structure and the hybrid convolution-SSM module gradually reduce the feature map size through spatial downsampling and expand the number of channels;

[0023] Design a progressive residual convolution structure, which includes multiple layers of convolution and residual connection. The convolution kernel size of each convolution layer is 7, 5, 3 and 1 respectively, and each convolution layer is followed by a GeLU activation function. The residual connection transmits information through a direct path.

[0024] A composite convolution-SSM module is designed. The composite convolution-SSM module includes an inverted deformable convolution IDC and a Mamba module. The input feature map is divided into two subsets according to the channel, one subset enters the inverted deformable convolution IDC, and the other subset enters the Mamba module. The output features of the inverted deformable convolution IDC and the Mamba module are fused through the cross-gated feature fusion CGFF module.

[0025] As a possible implementation of this embodiment, the inverted deformable convolution IDC introduces a learnable offset based on the standard convolution kernel. The expression of the inverted deformable convolution IDC is:

[0026] ,

[0027] in, represents the coordinates of the output location, is the index of the convolution kernel, For the The weight of the convolution kernel; and are the input and output feature maps, respectively. For the predefined fixed offset positions; and are the dynamic offset and modulation factor respectively.

[0028] As a possible implementation of this embodiment, the Mamba introduces a mechanism for selectively processing information to selectively process input information; adopts a hardware-aware algorithm and uses a parallel scanning algorithm to perform cyclic calculations of the model; simplifies the architecture design of the SSM by merging the design of the previous SSM architecture with the MLP block of the Transformer into a single module.

[0029] As a possible implementation of this embodiment, the cross-gated feature fusion CGFF includes two convolution blocks, which process features from the inverted deformable convolution IDC and Mamba modules respectively, and are processed by convolution modules using a 1×1 convolution layer followed by batch normalization (BN) and an activation function, and then an attention map is generated through a sigmod function, and then element-by-element multiplication is performed for interactive processing, and then the interactively processed features are spliced ​​to form the final fused features.

[0030] As a possible implementation of this embodiment, the expression of the cross-gated feature fusion CGFF is as follows:

[0031] ,

[0032] in, is the feature from the inverted deformable convolution IDC, For features from the Mamba module, is the output of the inverted deformable convolution IDC after processing by the convolution block and the ReLU activation function. is the output of the Mamba module after element-by-element multiplication by the SiLU activation function. for and The fused features after splicing.

[0033] As a possible implementation of this embodiment, the step 4, using the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain a radar signal modulation recognition model, includes:

[0034] The processed time-frequency image is input into the multi-branch multi-view feature extraction network DCMNet and trained using supervised learning;

[0035] During the training process, the cross entropy loss function is used to measure the classification performance of the model, and the optimization algorithm (AdamW) is used to adjust the model parameters to minimize the classification error;

[0036] The time-frequency features are extracted through forward propagation, and the weights are updated through back propagation, so that radar signals of different modulation types can be effectively distinguished;

[0037] After the network training is completed, the radar signal modulation recognition model is obtained.

[0038] In a second aspect, an embodiment of the present invention provides a device for radar signal modulation recognition, including:

[0039] A signal acquisition module is used to acquire radar signals and generate a variety of radar modulation signals;

[0040] The data set construction module is used to perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on radar modulation signals, convert them into time-frequency images and form data sets;

[0041] The network building module is used to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF;

[0042] The network training module is used to train the multi-branch multi-view feature extraction network DCMNet using the data set to obtain the radar signal modulation recognition model.

[0043] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform the steps of any of the above-mentioned methods for radar signal modulation identification.

[0044] In a fourth aspect, an embodiment of the present invention provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned radar signal modulation identification methods are executed.

[0045] The beneficial effects of the technical solution of the embodiment of the present invention are as follows:

[0046] The multi-branch multi-view feature extraction network DCMNet based on compound deformable convolution and Mamba of the present invention forms a compound convolution-SSM structure by introducing the designed inverted deformable convolution structure (IDC) and selective state space model (SSM). This innovative design utilizes the flexibility of deformable convolution and the global information processing capability of the state space model. The inverted deformable convolution module can dynamically adjust the position of the convolution kernel according to the characteristics of the input signal to more accurately extract the signal space features; the selective state space model enhances the performance of the model in processing long sequence signals by capturing the dependency between global information and long sequence data; combined with the cross-gated feature fusion (CGFF) structure, the fusion of features and the transmission of information are further optimized to ensure the comprehensiveness and consistency of feature extraction. The present invention not only realizes the modulation recognition of radar signals, but also improves the modulation recognition performance of radar signals, and solves the problem of difficulty in distinguishing multiple modulation signals in complex electromagnetic environments.

[0047] The multi-branch multi-view feature extraction network DCMNet of the present invention has low resource consumption, so that DCMNet can still run efficiently under the condition of limited hardware resources, and is particularly suitable for scenarios with limited computing resources and storage in practical applications;

[0048] The multi-branch multi-view feature extraction network DCMNet of the present invention adopts a multi-branch and multi-view design, which enables the network to comprehensively extract signal features from multiple scales and dimensions. This feature extraction method effectively improves the recognition accuracy of the model for different radar signal modulation modes, ensuring strong robustness when facing complex signals.

[0049] The multi-branch multi-view feature extraction network DCMNet of the present invention can still achieve excellent modulation recognition effect with a small number of samples, has strong generalization ability, and can maintain high recognition performance when the number of samples is insufficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of a method for radar signal modulation recognition according to an exemplary embodiment;

[0051] Figure 2 is a schematic structural diagram of a device for radar signal modulation recognition according to an exemplary embodiment;

[0052] Figure 3 It is an overall framework diagram of a multi-branch multi-view feature extraction network DCMNet according to an exemplary embodiment;

[0053] Figure 4 is a progressive residual convolution structure framework diagram according to an exemplary embodiment;

[0054] Figure 5 is a composite convolution-SSM structure diagram according to an exemplary embodiment;

[0055] Figure 6 is a diagram of an inverted deformable convolution framework according to an exemplary embodiment;

[0056] Figure 7 is a structural diagram of Mamba according to an exemplary embodiment;

[0057] Figure 8 is a cross-gated feature fusion framework diagram according to an exemplary embodiment;

[0058] Fig. 9 is a schematic diagram showing the recognition accuracy of each signal according to an exemplary embodiment of the present invention;

[0059] Fig.10 This is a comparison chart of the total recognition rate of the data set by different methods;

[0060] Fig.11 This is a comparison chart of different methods in terms of FLOPs and GPU memory usage;

[0061] Fig.12 This is a comparison chart of the recognition accuracy of different methods under different signal-to-noise ratios;

[0062] Fig.13 It is a schematic diagram of the visualization effect of the confusion matrix of the present invention. DETAILED DESCRIPTION

[0063] In order to more clearly illustrate the technical features of the solution of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0064] like Figure 1 As shown, a method for radar signal modulation recognition provided by an embodiment of the present invention includes the following steps:

[0065] Step 1, acquiring radar signals and generating a variety of radar modulation signals;

[0066] Step 2: Perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, convert it into a time-frequency image and form a data set;

[0067] Step 3: Build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF;

[0068] Step 4: Use the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain the radar signal modulation recognition model.

[0069] As a possible implementation of this embodiment, the step 1 of acquiring a radar signal and generating a plurality of radar modulation signals includes:

[0070] Collection time t Radar signal:

[0071] ,

[0072] in, represents the amplitude envelope of the signal, represents the instantaneous frequency, represents the instantaneous phase, is additive noise, usually modeled as zero-mean Gaussian white noise;

[0073] Use MATLAB to generate a variety of radar modulation signals, including LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank.

[0074] Firstly, a radar signal dataset containing ten different modulation types was constructed. These modulation types cover a variety of typical radar signal modulation methods, aiming to provide diverse signal samples for subsequent modulation identification.

[0075] As a possible implementation of this embodiment, the step 2, performing an improved multiple synchronous compression transform IMSST time-frequency analysis process on the radar modulation signal to convert it into a time-frequency image, includes:

[0076] The radar modulation signal is transformed into a time-frequency image by using an improved multi-scale time-frequency analysis method. :

[0077] ,

[0078] in, and represents discrete frequency, express n The discrete short-time Fourier transform at is the Kronecker delta function, represents two rounding operations on the instantaneous frequency of the multi-synchronous compression transform, Indicates the number of iterations.

[0079] In order to enhance the model's ability to identify the characteristics of different modulation signals, the data set was subjected to an improved multiple synchronous compression transform (IMSST) time-frequency analysis. This analysis method combines compression and synchronization technology to extract the time-frequency characteristics of radar signals and generate corresponding time-frequency images, thereby retaining the main time domain and frequency domain information of the signal. After IMSST processing, the time-frequency image can not only show the changes in the signal on the time axis and frequency axis, but also effectively display the instantaneous characteristics and spectral characteristics of the signal, providing rich input data for subsequent feature extraction and classification.

[0080] As a possible implementation of this embodiment, the step 3 is to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF, including:

[0081] The multi-branch multi-view feature extraction network DCMNet is designed as a whole to form a DCMNet overall framework structure, which includes a convolutional embedding layer, a stacked progressive residual convolution structure and a hybrid convolution-SSM module; the convolutional embedding layer, the stacked progressive residual convolution structure and the hybrid convolution-SSM module gradually reduce the feature map size through spatial downsampling and expand the number of channels;

[0082] Design a progressive residual convolution structure, which includes multiple layers of convolution and residual connection. The convolution kernel size of each convolution layer is 7, 5, 3 and 1 respectively, and each convolution layer is followed by a GeLU activation function. The residual connection transmits information through a direct path.

[0083] A composite convolution-SSM module is designed. The composite convolution-SSM module includes an inverted deformable convolution IDC and a Mamba module. The input feature map is divided into two subsets according to the channel, one subset enters the inverted deformable convolution IDC, and the other subset enters the Mamba module. The output features of the inverted deformable convolution IDC and the Mamba module are fused through the cross-gated feature fusion CGFF module.

[0084] As a possible implementation of this embodiment, the inverted deformable convolution IDC introduces a learnable offset based on the standard convolution kernel. The expression of the inverted deformable convolution IDC is:

[0085] ,

[0086] in, represents the coordinates of the output location, is the index of the convolution kernel, For the The weight of the convolution kernel; and are the input and output feature maps, respectively. For the predefined fixed offset positions; and are the dynamic offset and modulation factor respectively.

[0087] As a possible implementation of this embodiment, the Mamba introduces a mechanism for selectively processing information to selectively process input information; adopts a hardware-aware algorithm and uses a parallel scanning algorithm to perform cyclic calculations of the model; simplifies the architecture design of the SSM by merging the design of the previous SSM architecture with the MLP block of the Transformer into a single module.

[0088] As a possible implementation of this embodiment, the cross-gated feature fusion CGFF includes two convolution blocks, which process features from the inverted deformable convolution IDC and Mamba modules respectively, and are processed by convolution modules using a 1×1 convolution layer followed by batch normalization (BN) and an activation function, and then an attention map is generated through a sigmod function, and then element-by-element multiplication is performed for interactive processing, and then the interactively processed features are spliced ​​to form the final fused features.

[0089] As a possible implementation of this embodiment, the expression of the cross-gated feature fusion CGFF is as follows:

[0090] ,

[0091] in, is the feature from the inverted deformable convolution IDC, For features from the Mamba module, is the output of the inverted deformable convolution IDC after processing by the convolution block and the ReLU activation function. is the output of the Mamba module after element-by-element multiplication by the SiLU activation function. for and The fused features after splicing.

[0092] The present invention designs and builds the entire DCMNet network, which can efficiently extract multi-level and multi-angle feature information from the time-frequency image of the radar signal through a series of innovative convolution structures and feature fusion mechanisms. The design process of DCMNet starts from the basic convolution layer, and gradually adds a progressive residual convolution structure to ensure that information can be effectively transmitted in the deep network to avoid information loss. The progressive residual structure further improves the performance of the network by optimizing and strengthening feature learning layer by layer, especially when processing complex signals, it can accurately capture the slight changes in the signal. In order to more accurately extract spatial and frequency domain information, the network also designs a composite convolution-SSM structure. This structure first performs channel expansion to obtain more dimensional information later. The adaptability of the network to changes in signal morphology is further enhanced by the deformable convolution module. The deformable convolution can adaptively adjust the shape of the convolution kernel according to the characteristics of the input signal, thereby more effectively capturing the local changes and irregular structures of the signal. In addition, the SSM scanning mechanism can dynamically adjust the scanning strategy of the convolution kernel, focus on learning important areas in the signal, and further improve the accuracy of feature extraction. In terms of feature fusion, the present invention also designs a cross-gated feature fusion mechanism (CGFF), through which the network can efficiently fuse features between different levels and branches, and select the most representative features for merging. This design enables the network to fully understand all levels of the signal from multiple perspectives, thereby effectively improving the accuracy and robustness of radar signal modulation recognition. The entire network structure gradually extracts key information from the signal through multi-level convolution, scanning and fusion mechanisms, and ultimately uses this information for accurate classification of modulation types.

[0093] As a possible implementation of this embodiment, the step 4, using the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain a radar signal modulation recognition model, includes:

[0094] The processed time-frequency image is input into the multi-branch multi-view feature extraction network DCMNet and trained using supervised learning;

[0095] During the training process, the cross entropy loss function is used to measure the classification performance of the model, and the optimization algorithm (AdamW) is used to adjust the model parameters to minimize the classification error;

[0096] The time-frequency features are extracted through forward propagation, and the weights are updated through back propagation, so that radar signals of different modulation types can be effectively distinguished;

[0097] After the network training is completed, the radar signal modulation recognition model is obtained.

[0098] Based on the constructed radar signal data set, the DCMNet network was systematically trained. During the training process, the data set was divided into a training set, a validation set, and a test set to ensure that the model can generalize and learn on different signal samples. During the training process, an adaptive optimization algorithm was used to accelerate convergence, and the network parameters were adjusted through multiple rounds of iterations to improve the modulation recognition performance of the network. During the training process, DCMNet was able to gradually optimize its feature extraction and classification capabilities, and accurately distinguish different radar signal modulation types through multi-level convolution and feature fusion techniques. Finally, the fully trained network was able to accurately identify the modulation of the input radar signal, not only showing good recognition effects on the training data, but also demonstrating the effectiveness and robustness of the method in practical applications on the test data. Through this training process, the present invention successfully achieved efficient and accurate radar signal modulation recognition, providing reliable technical support for subsequent radar applications.

[0099] like Figure 2 As shown, an embodiment of the present invention provides a radar signal modulation recognition device, including:

[0100] A signal acquisition module is used to acquire radar signals and generate a variety of radar modulation signals;

[0101] The data set construction module is used to perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on radar modulation signals, convert them into time-frequency images and form data sets;

[0102] The network building module is used to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF;

[0103] The network training module is used to train the multi-branch multi-view feature extraction network DCMNet using the data set to obtain the radar signal modulation recognition model.

[0104] The specific implementation process of radar signal modulation identification of the present invention is as follows.

[0105] Step 1: Acquire radar signals and generate multiple radar modulation signals.

[0106] t The radar signal received at each moment can usually be expressed as a complex signal in the following form:

[0107] ,

[0108] in, represents the amplitude envelope of the signal, represents the instantaneous frequency, represents the instantaneous phase, is additive noise, usually modeled as zero-mean Gaussian white noise.

[0109] Ten radar modulation signals, including LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank, are generated using MATLAB 2021b as simulation data to verify the effectiveness of the model. Here, [·] represents a random parameter set, f s , f c , N s , B , T p , N bc , cpp , M and N p They represent sampling frequency, carrier frequency, number of samples, bandwidth, symbol width, Barker code length, number of codes per cycle, frequency step, and code period, respectively. All signals are simulated according to different signal-to-noise ratios (SNRs), and the SNR range is set from -16dB to 4dB with an interval of 2dB to ensure data diversity and fairness of the simulation experiment. The specific parameters of the signal are shown in Table 1.

[0110] Table 1 Signal parameters

[0111]

[0112] Step 2: Perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, convert it into a time-frequency image and form a data set.

[0113] In order to accurately extract the time-frequency characteristics of radar signals, the improved multi-scale time-frequency analysis method (IMSST) is used to perform time-frequency transformation on the original signal. The IMSST method effectively improves the resolution and clarity of time-frequency representation through multi-scale transformation and improved synchronous squeezing transformation. The core idea is to decompose the signal using multi-scale analysis and concentrate the time-frequency energy on the actual time-frequency trajectory through synchronous squeezing. Finally, the obtained time-frequency image size is 1×64×64, which can be used as the input of the deep learning network. IMSST can be described by the following formula:

[0114] ,

[0115] Among them, the variable and represents discrete frequency, express n The discrete short-time Fourier transform at is the Kronecker delta function; represents two rounding operations on the instantaneous frequency of the multi-synchronous compression transform, Indicates the number of iterations.

[0116] Step 3: Build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF.

[0117] 3.1 Overall design of DCMNet.

[0118] Figure 3 This is the overall framework of DCMNet proposed in this invention. The framework consists of multiple key modules, including convolutional embedding layers (using 7×7 large kernel convolution), stacked progressive residual convolution structures, and hybrid convolution-SSM modules. The feature map size is gradually reduced by spatial downsampling between layers, and the number of channels is expanded. Finally, the extracted features are flattened and linearly classified to achieve the recognition of radar signal modulation patterns. In addition, DCMNet focuses on simplicity and lightness in design, using only one block per layer and only 0.46M parameters, which significantly reduces computational complexity and resource consumption while maintaining good performance.

[0119] 3.2 Design of progressive residual convolution structure.

[0120] Figure 4 The overall framework structure of the progressive residual convolution proposed in the present invention. The structure includes multiple layers of convolution and residual connections. The convolution kernel sizes of each convolution layer are 7, 5, 3 and 1 respectively. Through this layer-by-layer refined multi-scale convolution kernel design, more and more refined feature information can be gradually extracted to ensure the detail of feature extraction. At the same time, each convolution layer is connected to the GeLU activation function. This activation function helps to improve the feature expression ability and training stability of the network through its smooth nonlinear characteristics. In addition, in order to alleviate the gradient disappearance problem that may occur in deep networks, a large number of residual connections are introduced. The residual connection transmits information through a direct path to ensure that the gradient can be smoothly back-propagated, thereby maintaining the stability of model training. Through the design of layer-by-layer convolution and residual connection, the progressive residual convolution can significantly reduce the size of the feature map while extracting rich features, improve computational efficiency, and adapt to practical application environments with limited resources.

[0121] 3.3 Design of hybrid convolution-SSM module.

[0122] Figure 5The composite convolution-SSM structure proposed in the present invention. This structure combines the advantages of convolution operation and Mamba through a dual-branch structure. First, the input feature map is divided into two subsets according to the channel, one subset enters the convolution branch, and the other subset enters the Mamba branch. The advantage of doing this is that it reduces the computational burden of each branch, while allowing different feature processing methods to work independently on different channels. After completing the feature extraction, the output features of the two branches are fused through a cross-gated feature fusion structure. Afterwards, a channel shuffle operation is introduced to rearrange the fused features so that the features between different channels are more evenly distributed, effectively avoiding information redundancy and imbalance problems.

[0123] Figure 6 This is the structure diagram of the inverted deformable convolution (IDC) proposed in the present invention. This structure combines deformable convolution and channel interaction mechanism to improve the spatial and channel feature extraction capabilities of radar signals. The original intention of the design of IDC is to improve the flexibility and accuracy of feature extraction by dynamically adjusting the position of the convolution kernel and enhancing the information interaction between channels. IDC first expands the channel of the input feature through a convolution block with an expansion rate of 2. This design enables subsequent convolution operations to process richer feature information. Deformable convolution is a technology that enhances the flexibility of convolutional neural networks. Unlike traditional convolution, deformable convolution introduces a learnable offset based on the standard convolution kernel, so that the convolution kernel can dynamically adjust its sampling position according to the geometric structure of the input feature, thereby better capturing the complex spatial information in the feature map. The formula for deformable convolution is as follows:

[0124] ,

[0125] in, represents the coordinates of the output location, is the index of the convolution kernel, For the The weight of the convolution kernel. and are the input and output feature maps, respectively. For the predefined fixed offset positions. and They are the dynamic offset and modulation factor, which are learned through the convolution layer and used to adjust the sampling position of the convolution kernel. By introducing these two learned offsets, the convolution kernel can flexibly select the best sampling position according to the input content, thereby capturing more refined local features. This method significantly enhances the adaptability of the model when facing complex and irregular inputs, and effectively improves the accuracy and robustness of feature extraction.

[0126] IDC's channel interaction mechanism, by enhancing information transfer between channels, enables the convolution kernel to perform multi-level information fusion in the spatial and channel dimensions, further improving the performance of feature extraction. Through this design, IDC can more comprehensively understand the different levels of features in radar signals, thereby improving the performance of signal classification and recognition tasks.

[0127] Figure 7 The Mamba structure in the present invention is improved on the basis of the state space model (SSM), and its main contributions include three aspects: selective processing of information, hardware-aware algorithms, and a simpler SSM architecture. These improvements enable Mamba to perform well in the field of radar signal modulation recognition. First, Mamba introduces a mechanism for selective processing of information (Selection Mechanism), which can selectively process input information. This mechanism enables the model to extract and utilize key information more effectively, thereby improving the accuracy of feature extraction. Secondly, Mamba adopts a hardware-aware algorithm. The algorithm uses a parallel scanning algorithm instead of traditional convolution operations to perform cyclic calculations of the model. This method not only realizes parallel training of the model, but also effectively reduces IO access between different levels in the GPU memory hierarchy. This is inspired by S5 (Simplified State Space Layers for Sequence Modeling), which optimizes the efficiency of hardware resource utilization by avoiding the concretization of extended states. Finally, Mamba simplifies the architectural design of SSM. By merging the design of the previous SSM architecture with the MLP block of Transformer into a single module, Mamba creates a simpler deep sequence model architecture.

[0128] Figure 8 This is the structural diagram of the cross-gated feature fusion (CGFF) proposed in the present invention. This module aims to achieve efficient fusion of multi-branch features, rather than just simple feature splicing. The CGFF structure contains two convolution blocks, which process features from the deformable convolution branch and features from the Mamba branch respectively. These inputs are processed by each convolution block and the Sigmoid activation function, and are multiplied element-by-element with the features on the other side. This element-by-element multiplication operation ensures that features from two different branches can modulate and interact with each other, which can enhance the expressive power of the features. Subsequently, these interactively processed features are spliced ​​to form the final fused features. In this way, the information diversity and sufficiency of the fused features across the entire channel range are ensured. The expression is as follows:

[0129] ,

[0130] in, is the feature from the inverted deformable convolution IDC, For features from the Mamba module, is the output of the inverted deformable convolution IDC after processing by the convolution block and the ReLU activation function. is the output of the Mamba module after element-by-element multiplication by the SiLU activation function. for and The fused features after splicing.

[0131] Step 4: Use the data set to train the constructed DCMNet network, and finally realize the modulation recognition of radar signals.

[0132] The processed time-frequency images are input into the DCMNet model and trained using supervised learning. During the training process, the cross entropy loss function is used to measure the classification performance of the model, and the optimization algorithm (AdamW) is used to adjust the model parameters to minimize the classification error. The model extracts time-frequency features through forward propagation and updates the weights through back propagation, so that it can effectively distinguish radar signals of different modulation types. After the training is completed, the final model for radar signal modulation recognition is obtained.

[0133] The effect of the present invention is further illustrated by the following comparative experiments.

[0134] 1 Experimental environment configuration.

[0135] During training, all experiments were conducted on a Linux server with the following configuration: Intel(R) Xeon(R) Gold 6122 CPU, 64GB RAM, Nvidia RTX 3090 GPU, 24GB RAM. The software environment includes Python 3.8, PyTorch 1.12.1, and Nvidia CUDA 11.6 and CuDNN 8.4. The number of training rounds is 100 and the batch size is 32. The base learning rate is set to 0.001. The warm-up learning rate is set to 1×10⁻4 and gradually increases to the base learning rate in the early stage of training. In addition, the cosine annealing strategy is used to smoothly adjust the learning rate during training to further optimize the model, and random horizontal flipping and random vertical inversion are uniformly used to enhance data diversity. The AdamW optimizer is used to update the network model.

[0136] 2 Analysis and evaluation of experimental results.

[0137] 2.1 Signal recognition performance of the method of the present invention.

[0138] like Fig. 9As shown, LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS and Frank are radar modulation signals. Fig. 9 The radar signal modulation pattern recognition performance of the proposed method under different SNR conditions is demonstrated. Experiments show that the proposed method can maintain a high recognition accuracy under most signal-to-noise ratio conditions. With the increase of SNR, the recognition performance of each category of signals is significantly improved. For example, the accuracy of the 4FSK signal is only 37.50% when SNR=-16dB, but as the SNR increases to -10dB, the accuracy quickly increases to 72.50%, and stabilizes at 97.50% when SNR≥0dB. Similarly, the accuracy of the BPSK signal is low under low signal-to-noise ratio, but it is significantly improved to 57.50% when SNR=-10dB, and reaches 100% when SNR≥0dB. The EQFM signal shows extremely strong robustness under all SNR conditions, and the accuracy is always maintained above 96.25%. Other signal categories such as FSK, LFM, LFM-BPSK, NS, and SFM also show significant accuracy improvement with the increase of SNR, especially under high SNR conditions, the accuracy of almost all categories reaches 100%.

[0139] 2.2 Comparison of different methods.

[0140] In order to verify the performance advantage of the DCMNet proposed in this paper in the radar signal modulation recognition task, several representative neural network architectures were compared and evaluated. These comparative methods include ResNet[He K, Zhang conference oncomputer vision and pattern recognition. 2018: 4510-4520.], ConvNeXt[Liu Z,Mao H, Wu CY, et al. S, Rastegari M. Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer[J]. arXiv preprint arXiv:2110.02178,2021.], VAN[Guo MH, Lu CZ, Liu ZN, et al. Visual attention network[J]. Computational Visual Media, 2023, 9(4): 733-752.], LPINet[Huynh-The T, Doan VS, Hua CH, et al.Accurate LPI radar waveform recognition with CWD-TFA fordeep convolutional network[J]. IEEE Wireless Communications Letters, 2021, 10(8): 1638-1642.] and MAPNet[Chen K, Zhang J, Chen S, et al. Deep metric learningfor robust radar signal recognition[J]. Digital Signal Processing, 2023, 137:104017.]. These methods are widely used in the fields of computer vision and radar signal processing, each with its own characteristics, and can represent different model design ideas and characteristics. .

[0141] Figure 10-12Among them, ResNet-18 is a lightweight variant of the deep residual network (Residual Neural Network, ResNet) series proposed by Microsoft Research. It contains 18 layers of convolution and fully connected layers. Its core idea is to alleviate the gradient vanishing problem through residual connections and improve the training efficiency of deep networks; MobileNetV2 is a lightweight convolutional neural network proposed by Google, which is mainly optimized for environments with limited computing resources such as mobile devices. Its core design includes deep separable convolution and linear bottleneck structure to reduce computational complexity and improve reasoning efficiency; ConvNext is a ConvNext series model proposed by Meta AI. Based on the ResNet structure, it improves the performance of convolutional networks in visual tasks while retaining the computational efficiency of convolutional networks by borrowing optimization strategies such as normalization in Vision Transformer (ViT); MobileViT is a combination of convolutional networks and Transformer The model adopts a global self-attention mechanism to enhance the information interaction of different spatial scales, so as to maintain a strong feature expression ability under low FLOPs conditions; VAN is a visual attention network VAN (Visual Attention Network), which is a solution based on deep convolution, point convolution, and diffusion convolution to replace large kernel convolution, which improves the performance of visual tasks while reducing the time complexity of the model; LPINet is a classic radar signal feature extraction network, which uses convolutions of different scales for stacking, which is lightweight and effectively improves the recognition effect; MAPNet is a lightweight neural network architecture, whose design goal is usually to improve computational efficiency while maintaining a high feature extraction capability. Its core module MAPB (Multi-scale Attention-based Processing Block) combines multi-scale feature extraction and attention mechanism to enhance the expression ability of the model.

[0142] Fig.10The overall accuracy of these methods on the radar signal modulation recognition dataset is shown. The method proposed in the present invention is significantly better than other methods in terms of accuracy and achieves the highest recognition rate. Specifically, ResNet, as a classic convolutional neural network model that first introduced residual links, achieved an accuracy of 82.99% in this experiment. MobileNetV2 has a slightly improved accuracy while taking into account lightweight, reaching 83.63%. The accuracy of ConvNext is 79.45%, which is relatively weak. MobileViT combines convolution and self-attention mechanisms to achieve good results, with an accuracy of 83.99%. VAN replaces large kernel convolution with deep separable convolution, diffusion convolution and point convolution, reducing complexity while performing well, with an accuracy of 83.78%. The accuracy of LPINet is 78.58%, the lowest among all the comparison methods, which may be due to the fact that its network structure does not fully extract specific features. The accuracy of the multi-scale feature fusion network MAPNet is 82.16%, which shows that multi-scale features play a certain role in improving recognition performance. The DCMNet method proposed in this paper significantly improves the recognition accuracy of the model to 84.90% by introducing multi-view context information. This shows that considering the multi-dimensional feature information of the input signal is an effective way to improve the modulation recognition performance of radar signals.

[0143] After comparing the overall accuracy of each method, we further analyzed the performance of different models in terms of FLOPs (floating point operations) and GPU memory usage. Fig.11 As shown in the figure, the FLOPs of the proposed model is 95.79M, and the GPU memory occupancy is only 17.24MB. In comparison, the FLOPs of ResNet-18 is 142.44M, and the GPU memory occupancy is 58.09MB; the FLOPs of ConvNext is 362.88M, and the GPU memory occupancy is 129.14MB. Although MobileNetV2 and MobileViT are more economical in FLOPs parameters, their accuracy rates are 83.49% and 84.07%, respectively, which are lower than the proposed model. Overall, the proposed method performs best in terms of GPU memory occupancy. Although the FLOPs are not the smallest, a good balance is achieved between the amount of computation, storage burden and accuracy. This balance enables the proposed model to significantly reduce the demand for computing and storage resources while improving recognition accuracy.

[0144] In order to highlight the performance of the proposed method in radar signal modulation recognition under different signal-to-noise ratios, the present invention compares it with a variety of existing methods. Fig.12The recognition accuracy of each method under different SNR conditions is shown. The DCMNet method proposed in this invention performs well under all SNR conditions, especially under low SNR conditions, and its recognition accuracy is significantly higher than other methods. When the SNR is -16dB and -14dB, the recognition accuracy of DCMNet is 43.75% and 58.2% respectively, which is significantly better than other methods. This advantage is mainly attributed to the fact that the selective state space model can effectively filter out noise interference and retain useful signals in a high noise environment, while the deformable convolution can be more flexibly processed according to the characteristics of the signal. These technologies can more effectively capture the subtle features of the signal in a low signal-to-noise ratio environment. In contrast, the recognition accuracy of ConvNext at -16dB and -14dB is only 35.37% and 48.0%, which is because the pure convolution layer structure has insufficient feature extraction capabilities in a high noise environment, and MobileNetV2 and VAN are not as good as DCMNet at these two signal-to-noise ratios, with recognition accuracies of 40.13%, 41.88% and 56.5%, 55.76% respectively. As the SNR increases, the recognition accuracy of DCMNet increases rapidly. When the SNR is -12dB and -8dB, its recognition accuracy reaches 72.34% and 85.47% respectively, and reaches 98.63% at -4dB, which is also higher than methods such as ConvNext, VAN and MobileNetV2. Under high SNR conditions, the recognition accuracy of DCMNet is almost 100%. For example, when the SNR is 0dB, its recognition accuracy is 99.88%. Compared with LPINet and MAPNet, DCMNet has significant advantages in structural design. LPINet uses multiple layers of convolution and pooling layers. Although it performs well under certain SNR conditions, its recognition accuracy is low in high noise environments. MAPNet uses a multi-layer perceptron structure. Although it improves some performance, its recognition accuracy under low SNR conditions is still lower than that of DCMNet. By combining multi-scale convolution, deformable convolution and selective state space model, DCMNet performs well in feature extraction of high-resolution images and maintains stable recognition performance under low SNR conditions.

[0145] 2.3 Visualization of the confusion matrix of the proposed method.

[0146] In order to evaluate the performance of the model in the feature classification task, the recognition results were visualized and analyzed in combination with the confusion matrix. Fig.13The confusion matrix of the proposed model on the entire test set is shown. It can be seen that the model has achieved high recognition accuracy in most categories, and the dark blocks on the main diagonal indicate that the vast majority of samples are correctly classified. This shows that the DCMNet model has high accuracy in extracting and classifying radar signal features. There are some misclassifications, especially between LFM-BPSK and BPSK, SFM-BPSK and SFM, and 4FSK and FSK. The errors between LFM-BPSK and BPSK and between SFM-BPSK and SFM are more obvious. Since SFM-BPSK and LFM-BPSK are both composite signals, they have great similarity and overlap with BPSK in the feature space when the signal-to-noise ratio is low, making it difficult for the model to accurately identify them. There are also some classification errors between 4FSK and FSK, which may be due to their similarity in modulation mode, so that the features of the time-frequency images are relatively close. Despite some misrecognition between similar modulation modes, the DCMNet model performs well in most categories and can effectively extract and classify radar signal features.

[0147] Compared with the prior art, the present invention has the following advantages:

[0148] (1) The DCMNet model has only 0.46M parameters, 95.79M computational load (FLOPs), and only 17.24MB of GPU memory. It has low resource consumption, allowing the model to run efficiently even with limited hardware resources. It is particularly suitable for scenarios with limited computing resources and storage in practical applications.

[0149] (2) DCMNet adopts a multi-branch, multi-perspective design, which enables the network to comprehensively extract signal features from multiple scales and dimensions. This feature extraction method effectively improves the model's recognition accuracy for different radar signal modulation modes, ensuring strong robustness when facing complex signals.

[0150] (3) Despite the complex network design, DCMNet can still achieve excellent modulation recognition results with a small number of samples; this shows that the method has strong generalization ability and can maintain high recognition performance even when the number of samples is insufficient.

[0151] An electronic device provided by an embodiment of the present invention includes a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory through the bus, and the processor executes the machine-readable instructions to perform the steps of any of the above-mentioned methods for radar signal modulation recognition.

[0152] Specifically, the above-mentioned memory and processor can be general-purpose memory and processor, which are not specifically limited here. When the processor runs the computer program stored in the memory, the above-mentioned radar signal modulation identification method can be executed.

[0153] Those skilled in the art will appreciate that the structure of the computer device does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or split certain components, or arrange the components differently.

[0154] In some embodiments, the computer device may also include a touch screen that can be used to display a graphical user interface (e.g., a startup interface of an application) and receive user operations on the graphical user interface (e.g., startup operations on an application). The specific touch screen may include a display panel and a touch panel. The display panel may be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc. The touch panel may collect the user's contact or non-contact operations on or near it, and generate pre-set operation instructions, for example, the user uses any suitable object such as a finger, a stylus, or an accessory on or near the touch panel. In addition, the touch panel may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and posture, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into information that the processor can process, and then sends it to the processor, and can receive and execute commands from the processor. In addition, the touch panel can be implemented by various types such as resistive, capacitive, infrared and surface acoustic wave, and any technology developed in the future can also be used to implement the touch panel. Further, the touch panel can cover the display panel, and the user can operate on or near the touch panel covered on the display panel according to the graphical user interface displayed on the display panel. After the touch panel detects the operation on or near it, it is transmitted to the processor to determine the user input, and then the processor provides corresponding visual output on the display panel in response to the user input. In addition, the touch panel and the display panel can be implemented as two independent components or integrated.

[0155] Corresponding to the method for starting the above-mentioned application, an embodiment of the present invention further provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods for radar signal modulation recognition are executed.

[0156] The application startup device provided in the embodiment of the present application can be specific hardware on the device or software or firmware installed on the device. The device provided in the embodiment of the present application, its implementation principle and the technical effect produced are the same as those in the aforementioned method embodiment. For the sake of brief description, the parts not mentioned in the device embodiment can refer to the corresponding contents in the aforementioned method embodiment. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can all refer to the corresponding processes in the aforementioned method embodiment, and will not be repeated here.

[0157] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0158] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0159] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0160] In addition, each functional module in the embodiments provided in the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0161] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for radar signal modulation recognition, characterized in that: The steps include: Step 1, acquiring radar signals and generating a variety of radar modulation signals; Step 2: Perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, convert it into a time-frequency image and form a data set; Step 3: Build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF; Step 4: Use the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain the radar signal modulation recognition model.

2. The method for radar signal modulation recognition according to claim 1, characterized in that: The step 1, obtaining a radar signal and generating a plurality of radar modulation signals, includes: Collection time t Radar signal: ， in, represents the amplitude envelope of the signal, represents the instantaneous frequency, represents the instantaneous phase, is additive noise, usually modeled as zero-mean Gaussian white noise; Use MATLAB to generate a variety of radar modulation signals, including LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank.

3. The method for radar signal modulation recognition according to claim 2, characterized in that: The step 2, performing an improved multiple synchronous compression transform IMSST time-frequency analysis process on the radar modulation signal to convert it into a time-frequency image, includes: The radar modulation signal is transformed into a time-frequency image by using an improved multi-scale time-frequency analysis method. : ， in, and represents discrete frequency, express n The discrete short-time Fourier transform at is the Kronecker delta function, represents two rounding operations on the instantaneous frequency of the multi-synchronous compression transform, Indicates the number of iterations.

4. The method for radar signal modulation recognition according to claim 1, characterized in that: The step 3 is to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF, including: The multi-branch multi-view feature extraction network DCMNet is designed as a whole to form a DCMNet overall framework structure, which includes a convolutional embedding layer, a stacked progressive residual convolution structure and a hybrid convolution-SSM module; the convolutional embedding layer, the stacked progressive residual convolution structure and the hybrid convolution-SSM module gradually reduce the feature map size through spatial downsampling and expand the number of channels; Design a progressive residual convolution structure, which includes multiple layers of convolution and residual connection. The convolution kernel size of each convolution layer is 7, 5, 3 and 1 respectively, and each convolution layer is followed by a GeLU activation function. The residual connection transmits information through a direct path. A hybrid convolution-SSM module is designed. The hybrid convolution-SSM module includes an inverted deformable convolution IDC and a Mamba module. The input feature map is divided into two subsets according to the channel, one subset enters the inverted deformable convolution IDC, and the other subset enters the Mamba module. The output features of the inverted deformable convolution IDC and the Mamba module are fused through the cross-gated feature fusion CGFF module.

5. The method for radar signal modulation recognition according to claim 4, characterized in that: The inverted deformable convolution IDC introduces a learnable offset based on the standard convolution kernel. The expression of the inverted deformable convolution IDC is: ， in, represents the coordinates of the output location, is the index of the convolution kernel, For the The weight of the convolution kernel; and are the input and output feature maps, respectively. For the predefined fixed offset positions; and are the dynamic offset and modulation factor respectively.

6. The method for radar signal modulation recognition according to claim 4, characterized in that: The Mamba introduces a mechanism for selectively processing information to selectively process input information; adopts a hardware-aware algorithm and uses a parallel scanning algorithm to perform cyclic calculations of the model; simplifies the architecture design of the SSM by merging the design of the previous SSM architecture with the MLP block of the Transformer into a single module; The cross-gated feature fusion CGFF includes two convolution blocks, which process the features from the inverted deformable convolution IDC and Mamba modules respectively, and are processed by convolution modules using 1×1 convolution layers followed by batch normalization and activation functions, and then the attention map is generated by the sigmod function and then element-by-element multiplication is performed for interactive processing, and then the interactively processed features are spliced ​​to form the final fused features.

7. The method for radar signal modulation recognition according to any one of claims 1 to 6, characterized in that: The step 4, using the data set to train the multi-branch multi-view feature extraction network DCMNet to obtain a radar signal modulation recognition model, includes: The processed time-frequency image is input into the multi-branch multi-view feature extraction network DCMNet and trained using supervised learning; During the training process, the cross entropy loss function is used to measure the classification performance of the model, and the AdamW optimization algorithm is used to adjust the model parameters to minimize the classification error; The time-frequency features are extracted through forward propagation, and the weights are updated through back propagation, so that radar signals of different modulation types can be effectively distinguished; After the network training is completed, the radar signal modulation recognition model is obtained.

8. A device for radar signal modulation recognition, characterized in that: include: A signal acquisition module is used to acquire radar signals and generate a variety of radar modulation signals; The data set construction module is used to perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on radar modulation signals, convert them into time-frequency images and form data sets; The network building module is used to build a multi-branch multi-view feature extraction network DCMNet based on inverted deformable convolution IDC, Mamba and cross-gated feature fusion CGFF; The network training module is used to train the multi-branch multi-view feature extraction network DCMNet using the data set to obtain the radar signal modulation recognition model.

9. An electronic device, characterized in that: The electronic device comprises a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the method for radar signal modulation recognition as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, executes the steps of the method for radar signal modulation recognition as claimed in any one of claims 1 to 7.

Images