A method for recognizing an indefinite number of overlapping radar signal modulations based on semantic segmentation

Abstract

The application discloses a kind of to the identification method of overlapping radar signal modulation type for indefinite signal quantity, applied to electronic reconnaissance field.Aiming at the difficulty of identification caused by signal overlapping caused by high pulse density in space, the application proposes an AttSegOverlap-Net (ASO-Net) network model based on semantic segmentation.The characteristics are that the network generalization for the incomplete signal intercepted is enhanced by simulating the loss of part of the signal through the incomplete mask module;The network feature extraction capability is improved by the multi-branch attention feature enhancement module;The feature screening capability of network is improved by the cross-layer semantic fusion module, so as to realize the accurate identification of overlapping signal, and have certain robustness.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing, and is particularly applicable to radar radiation source signal identification in electronic warfare. Specifically, it provides a method for identifying the modulation of an indefinite number of overlapping radar signals based on semantic segmentation. Background Technology

[0002] Radar signal modulation type identification aims to obtain the modulation information of signals intercepted by reconnaissance receivers, and then combine it with other reconnaissance information to determine the relevant characteristic attributes of enemy radar. It is an important basis for sorting and radiation source identification and has become one of the important research contents in the field of electromagnetic reconnaissance.

[0003] Currently, with the increasing number of radiation source devices in space, the complexity and uncertainty of the electromagnetic environment are constantly increasing. These radiation sources send a large number of signals simultaneously within the same frequency range. Electronic reconnaissance systems are likely to intercept two or even more signals at the same time. These signals overlap in the time and frequency domains, forming overlapping radar signals, which brings new challenges to the modulation and identification of radar signals.

[0004] To address the challenges of signal overlap caused by high pulse density in space, research on modulation recognition of overlapping signals has yielded some results. Overlapping radar signal modulation recognition can be mainly categorized into multi-class learning (MCL), multi-label learning (MLL), and multi-instance multi-label learning (MIML). The paper "Active learning for large multi-class problems" by P. Jain and A. Kapoor uses multi-class learning, treating each overlap pattern as a new signal type. As the number of signal types increases, the model size and recognition difficulty increase exponentially. The paper "Automatic modulation recognition of compound signals using a deep multi-label classifier: A casestudy with radar jamming signals" by M. Zhu, Y. Li, Z. Pan, and J. Yang uses multi-label learning, treating the signal as a whole, which is less adaptable to dynamic environments. The paper "Automatic waveform recognition of overlapping LPI radar signals based on multi-instance multi-label learning" by Z. Pan, S. Wang, M. Zhu, and Y. Li addresses overlapping signal recognition by using multi-instance multi-label learning, leveraging the correlation between class labels to process multiple semantic labels. The above-mentioned research methods have achieved modulation recognition of overlapping radar signals to a certain extent. However, the diversity and complexity of signal overlap scenarios lead to insufficient adaptability of existing recognition methods. Identifying overlapping signals requires a network model with stronger generalization capabilities. Modulation recognition of overlapping signals still needs further research. Summary of the Invention

[0005] To address the problem of identifying the modulation type of overlapping radar signals, a semantic segmentation-based network structure for identifying the modulation type of overlapping radar signals with an indefinite number of signals is proposed, which enables effective identification of the modulation type of overlapping radar signals.

[0006] The technical solution adopted in this invention is: an overlapping signal modulation recognition method based on semantic segmentation, comprising the following steps:

[0007] Step 1: Obtain the original simulated radar signal and modulation parameters, and then randomly overlap the simulated radar signal to generate an overlapped radar signal;

[0008] Step 2: Extract the time-frequency domain features of the overlapping radar signals to obtain the signal time-frequency image;

[0009] Step 3: Generate a random mask matrix, and multiply the mask matrix element-wise with the generated time-frequency image of the signal to obtain a simulated incomplete signal representation, which is used as the training set data;

[0010] Step 4: Construct the ASO-Net network, which includes an encoder and a decoder. The encoder consists of three multi-branch attention feature enhancement modules and two downsampling modules cascaded alternately. The decoder consists of three multi-branch attention feature enhancement modules, two upsampling modules, and two cross-layer semantic fusion modules cascaded alternately. The multi-branch attention feature enhancement module at the end of the encoder is connected to the multi-branch attention feature enhancement module at the beginning of the decoder. There are cross-layer connections between the multi-branch attention feature enhancement modules of the encoder and the cross-layer semantic fusion modules of the decoder in the same layer. The incomplete signal representation is input into the ASO-Net network, and the output is a pixel-level classification result of the time-frequency map.

[0011] Step 5: Train and test the ASO-Net network;

[0012] Step 6: Use the trained ASO-Net network to identify the modulation type of overlapping radar signals, obtain the pixel-level classification results of the time-frequency map, and determine the existing signal categories by comparing them with the threshold.

[0013] The multi-branch attention feature enhancement module includes a feature extraction part and a feature fusion part;

[0014] The time-frequency feature map of the given sample is input into the feature extraction part, and then passed through two convolutional kernels with sizes of [sizes to be filled in]. and Convolutional layers, in parallel , and Three depthwise separable convolutional layers and one The convolutional layer adds the input time-frequency feature map to obtain the output of the feature extraction part;

[0015] The output features of the feature extraction part are input to the feature fusion part, and three attention branches are input in parallel: channel attention, simple pixel attention, and pixel attention. The outputs of the three attention branches are concatenated according to the channel dimension to obtain the output result of the multi-branch attention feature enhancement module.

[0016] The channel attention is that it undergoes global average pooling in sequence, Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the output of the channel attention;

[0017] The simplified pixel attention has two parallel branches, one of which involves sequentially processing the input feature map. Convolutional layers and The convolutional layer, another branch, inputs the feature map and passes through it sequentially. The convolutional layer and sigmoid activation function are combined, and finally the outputs of the two branches are multiplied to obtain the output of the simplified pixel attention.

[0018] The pixel attention passes through sequentially Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the pixel attention output.

[0019] The cross-layer semantic fusion module takes the outputs of the multi-branch attention feature enhancement module and the upsampling module connected across layers as inputs, adds them together, and then performs global average pooling and multilayer perceptron to obtain weights. After the weights are normalized by Softmax, they are multiplied by the two inputs of this module to obtain weighted feature maps. Finally, the two weighted feature maps are added together and fused to obtain the output result of the cross-layer semantic fusion module.

[0020] The beneficial effects of this invention are as follows:

[0021] This invention addresses the problem that existing methods are inadequate for identifying overlapping signals due to the increasing number of radar radiation sources and the significantly higher probability of intercepting overlapping radar signals. It proposes a semantic segmentation-based network structure for overlapping signal modulation recognition. To enable the network to independently learn the features of each signal type even when signals overlap, pixel-level classification of the time-frequency graphs of overlapping signals is performed. To enhance the network's generalization ability to intercept incomplete signals, a partial signal loss module is used to simulate missing signals. Furthermore, due to the complexity of overlapping signal features and severe interference from background and cross-terms, a multi-branch attention feature enhancement module is used to improve the network's feature extraction capability, and a cross-layer semantic fusion module is used to enhance the network's feature selection capability, thus achieving accurate identification of incomplete overlapping signals. Attached Figure Description

[0022] Figure 1 ASO-Net network structure diagram provided for an example of the present invention;

[0023] Figure 2 A schematic diagram of the incomplete mask module provided for an example of the present invention;

[0024] Figure 3 The multi-branch attention feature enhancement module structure diagram provided for the example of the present invention is shown in the figure, where (a) is the feature extraction part and (b) is the feature fusion part;

[0025] Figure 4 A structural diagram of a cross-layer semantic fusion module provided for an example of the present invention;

[0026] Figure 5 The image shows the result of overlapping radar signal modulation identification as an example of the present invention. Detailed Implementation

[0027] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0028] This invention provides a method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation (AttSegOverlap-Net, ASO-Net network), comprising the following steps:

[0029] Step 1: Obtain the original simulated radar signal and modulation parameters. The original radar signal is the time domain signal, and the modulation parameters include bandwidth, pulse width, sampling frequency, etc. Then, randomly overlap the simulated radar signal to generate an overlapped radar signal.

[0030] Table 1 shows the specific dataset parameter settings for this invention example. This invention simulates seven typical radar signal modulation types, including Binary Phase Shift Keying (BPSK), Linear Frequency Modulation (LFM), Triangular Frequency Modulation (TFM), Cosine Frequency Modulation (SFM), S-shaped Nonlinear Frequency Modulation (NLFM), Binary Frequency Shift Keying (Costas), and Polyphase Code (Frank). All simulated signals are baseband signals, with sampling frequencies... for The sampling points are set to 1024, U represents uniform distribution, and the signal-to-noise ratio varies within a certain range. .

[0031] Table 1

[0032] Modulation type parameter scope BPSK Code length {7,11,13} BPSK carrier frequency <![CDATA[U(f s / 8,f s / 4)]]> LFM, TFM, SFM, NLFM bandwidth <![CDATA[U(f s / 16,f s / 8)]]> LFM, TFM, SFM, NLFM carrier frequency <![CDATA[U(f s / 8,f s / 4)]]> Costas Frequency coding length {4,5,6} Costas Basic frequency <![CDATA[U(f s / 30,f s / 24)]]> Costas initial frequency <![CDATA[U(f s / 8,f s / 4)]]> Frank Encoded phase number {8,9,10,11,12} Frank carrier frequency <![CDATA[U(f s / 8,f s / 4)]]>

[0033] Step 2: Extract the time-frequency domain features of overlapping radar signals. Specifically, for a given training set, use the Smooth Pseudo-Wigner Distribution (SPWVD) algorithm to extract time-frequency features. The Wigner-Willi distribution (WVD) calculation formula is as follows:

[0034]

[0035] in, It is time. It's frequency. It is a time lag. It is a time-domain signal. It is the complex conjugate of the time domain. It is the imaginary unit.

[0036]

[0037] in, and It is a window function in both the time and frequency domains. It is the integral variable of time. It is a time-domain signal. It is the complex conjugate of the time domain.

[0038] When overlapping signals completely overlap in the time domain, the way their frequency changes over time can still serve as a distinguishing feature. Time-frequency analysis, through the joint characterization of time and frequency, can effectively reflect the relationship between signal frequency and time, making it suitable for radar signal identification and classification. Therefore, the time-frequency characteristics of a signal are suitable as input to a network.

[0039] Step 3: Construct a missing mask module and generate a random mask matrix. Specifically, to better adapt the network to situations where some signals are missing in the actual captured signals, a missing mask module is used during the training phase to simulate incomplete signals generated from complete time-frequency images. Based on the signal distribution characteristics in time-frequency images, signal fragment loss manifests as vertical stripe-like information gaps in the time-frequency domain. Therefore, the design of the missing mask module first generates a mask matrix of all 1s with the same dimension as the time-frequency representation, and then randomly assigns zero values ​​to regions of different widths to obtain the mask matrix. For example... Figure 2 This is a schematic diagram of the incomplete mask module. By multiplying the mask matrix element-wise with the generated time-frequency image of the signal, the simulated incomplete signal representation can be obtained, which serves as the training set data.

[0040] Step 4: Construct the ASO-Net network, which includes an encoder and a decoder. The encoder consists of three multi-branch attention feature enhancement modules and two downsampling modules cascaded alternately, i.e., multi-branch attention feature enhancement module, downsampling module, multi-branch attention feature enhancement module, downsampling module, multi-branch attention feature enhancement module. The decoder consists of three multi-branch attention feature enhancement modules, two upsampling modules, and two cross-layer semantic fusion modules cascaded alternately, i.e., multi-branch attention feature enhancement module, upsampling module, cross-layer semantic fusion module, multi-branch attention feature enhancement module, upsampling module, cross-layer semantic fusion module, multi-branch attention feature enhancement module. The multi-branch attention feature enhancement module at the end of the encoder is connected to the multi-branch attention feature enhancement module at the beginning of the decoder. There are cross-layer connections between the multi-branch attention feature enhancement modules of the encoder and the cross-layer semantic fusion modules of the decoder in the same layer. Specifically, the first multi-branch attention feature enhancement module of the encoder is connected to the second cross-layer semantic fusion module of the decoder, and the second multi-branch attention feature enhancement module of the encoder is connected to the first cross-layer semantic fusion module of the decoder, forming the ASO-Net network. The incomplete signal representation is input into the ASO-Net network, and the output is a pixel-level classification result of the time-frequency graph.

[0041] like Figure 1 The diagram shows the specific ASO-Net network structure of this invention. The original radar signal time series is processed by the overlay and SPWVD algorithm to obtain a two-dimensional time-frequency map. After processing by the incomplete mask module, a simulated incomplete overlay signal time-frequency map is obtained. Inputting the incomplete overlay signal time-frequency map into the ASO-Net network yields pixel-level classification results. The number of pixels for each signal category can be statistically determined, and compared with a threshold to obtain the predicted signal category. The following sections further explain each module of the network.

[0042] The multi-branch attention feature enhancement module includes a feature extraction part and a feature fusion part. The difficulty in recognition caused by signal overlap is partly due to the mutual influence between signals, resulting in complex features and difficulties in extraction. Furthermore, because the overlapping signals are unevenly distributed in the time-frequency image, and different types of signals have different coverage areas, the same convolution cannot extract complete features of the overlapping signals and is easily affected by cross terms. To better extract effective information from the time-frequency image, a multi-branch attention feature enhancement module is used. The feature extraction part uses parallel depthwise separable convolutional kernels of different sizes to capture features at different scales. Small convolutional kernels can focus on features in the overlapping signal regions, while large convolutional kernels can better capture the overall signal features, improving the model's receptive field while preserving the ability to extract detailed features. Secondly, due to the influence of background information and overlapping terms on network feature learning, the feature fusion part adjusts the importance of features learned by the network through parallel attention. In order to enable the network to better classify pixels, two types of pixel attention are used in parallel: pixel attention obtains fine pixel weights by compressing feature channels, while simple pixel attention enhances the perception of local features through 3×3 convolution. Finally, parallel channel attention preserves the network's perception of global features.

[0043] like Figure 3 The diagram shows the specific structure of the multi-branch attention feature enhancement module in this invention, where Conv represents a convolutional layer, DWConv represents a depthwise separable convolutional layer, GAP represents global average pooling, GELU represents the GELU activation function, and Sigmoid represents the Sigmoid activation function. The detailed processing procedure is as follows:

[0044] First, the time-frequency feature map of the given sample is input into the feature extraction part, such as... Figure 3 As shown in (a), the data is processed sequentially through two convolutional kernels with sizes of [sizes not specified]. and Convolutional layers, in parallel , and Three depthwise separable convolutional layers and one The convolutional layer adds the input time-frequency feature map to obtain the output of the feature extraction part;

[0045] The output features from the feature extraction part are input to the feature fusion part, along with three parallel attention branches: channel attention, simple pixel attention, and pixel attention. The outputs of these three attention branches are concatenated along the channel dimension to obtain the output of the multi-branch attention feature enhancement module, as shown in the following example. Figure 3 As shown in (b) of the diagram.

[0046] The channel attention part sequentially undergoes global average pooling, Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the output of the channel attention.

[0047] The simplified pixel attention mechanism has two parallel branches, one of which processes the input feature map sequentially. Convolutional layers and The convolutional layer, another branch, inputs the feature map and passes through it sequentially. The convolutional layer and sigmoid activation function are combined, and finally the outputs of the two branches are multiplied to obtain the output of the simplified pixel attention.

[0048] The pixel attention part passes through sequentially Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the pixel attention output.

[0049] The downsampling module is specifically... Convolutional layers adjust feature map size; the upsampling module specifically comprises... Convolutional layers, adjusting feature map size.

[0050] The cross-layer semantic fusion module adopts a structure of global average pooling and multi-layer perceptron, and uses the outputs of the cross-layer connected multi-branch attention feature enhancement module and the upsampling module as inputs to achieve cross-layer feature fusion.

[0051] The time-frequency features of overlapping signals are complex, and severe signal overlap leads to significant cross-terms, making it difficult for the network to learn the complete features of the signal. Therefore, this method employs a cross-layer connection between the encoder and decoder, enabling the network to utilize signal features from both shallow and deep layers simultaneously. Since the semantic differences between shallow encoder features and deep decoder features are significant during cross-layer connections, direct addition cannot utilize all detailed information. Therefore, to better fuse the features at the cross-layer connection between the encoder and decoder and retain more effective signal features, global average pooling is performed on the original encoder and decoder feature maps at the cross-layer connection to obtain compressed spatial information. Then, a multilayer perceptron is used to learn the compressed features and obtain channel weights. Finally, the weighted feature maps are added according to the weights to obtain the selected features, thus preserving effective features from different layers.

[0052] like Figure 4 The diagram shows the specific cross-layer semantic fusion module structure of an example of the present invention, where GAP represents global average pooling, MLP represents a multilayer perceptron, and Softmax represents the Softmax activation function. The detailed processing procedure is as follows:

[0053] First, the outputs of the multi-branch attention feature enhancement module and the upsampling module at the cross-layer connection are added together, and then global average pooling is applied to compress spatial information. Next, the signal is passed through a multilayer perceptron, including... Convolutional layer, ReLU activation function The convolutional layer produces weights. These weights are then normalized using Softmax. The normalized weights are multiplied by the two inputs of this module to obtain weighted feature maps. Finally, the two weighted feature maps are added together and fused to obtain the output of the cross-layer semantic fusion module.

[0054] Step 5: Train and test the ASO-Net network. The training process is as follows: For each incomplete and overlapping radar signal time-frequency map input, the constructed ASO-Net network undergoes forward propagation, and the cost function value is calculated; the parameters of the ASO-Net network are updated using a gradient descent-based backpropagation algorithm; the backpropagation process is iterated until the cost function converges, thus obtaining the trained ASO-Net network. Specifically, this includes the following steps:

[0055] Step 5.1: Forward propagation;

[0056] Step 5.2: Calculate the cost function value. In this example, the focus loss function is used as the cost function. The calculation method is as follows:

[0057]

[0058] in, This represents the loss weight corresponding to the true category. This represents the predicted probability of the true class. This indicates the degree of weight decay, which is set to 2 in this embodiment.

[0059] Step 5.3: Update the network parameters using the gradient descent-based backpropagation algorithm.

[0060] Step 6: Use the trained ASO-Net network to identify the modulation type of overlapping radar signals, obtain the pixel-level classification results of the time-frequency map, and determine the existing signal categories by comparing them with the threshold.

[0061] To verify the recognition performance of the proposed method, we compare it with four other methods in this field: SE-Inception (an overlapping signal recognition method), and three semantic segmentation networks: UNet, SegNet, and DeeplabV3. All methods are trained on datasets with two overlapping signals, and the recognition accuracy is tested under the conditions of two, three, and four overlapping signals.

[0062] from Figure 5It can be seen that ASO-Net achieves higher overall recognition accuracy compared to the other four methods when two signals overlap. It achieves over 83% accuracy at -12dB, over 93% accuracy at -10dB (exceeding the accuracy of other algorithms by 3%), and nearly 100% accuracy at -4dB. When the number of overlapping signals in the training and test sets is the same, all four algorithms except UNet exhibit good recognition performance. Figure 5 It can be seen that when the number of overlapping signals in the test set is greater than that in the training set, the ASO-Net proposed in this invention exhibits superiority. (a) shows two-signal overlap, (b) shows three-signal overlap, and (c) shows four-signal overlap. With three-signal overlap, the proposed algorithm outperforms the four comparative methods in all test signal-to-noise ratios, achieving an accuracy exceeding 70% at -12dB, exceeding 84% at -10dB, and surpassing other algorithms by 4% accuracy at -8dB, demonstrating a significant advantage. With four-signal overlap, the proposed algorithm shows better robustness and generalization. At an SNR higher than -4dB, its accuracy is the only one among the methods to exceed 90%. The SE-Inception algorithm in this field is significantly affected by the number of signals, with a marked decrease in accuracy with four-signal overlap. Other semantic segmentation methods also show a significant performance decrease with four-signal overlap. These results demonstrate that the method proposed in this invention performs better with a greater number of overlapping signals, exhibiting better robustness and generalization.

[0063] Based on the above-mentioned network structures, the results of identifying a large number of overlapping radar signals demonstrate that the proposed algorithm can effectively separate incomplete overlapping signals. This is achieved by using a partial mask module to simulate incomplete signals intercepted in the actual environment, a multi-branch attention feature enhancement module to improve the network's ability to extract features of complex overlapping signals, and a cross-layer semantic fusion module to fuse features at different levels.

[0064] Those skilled in the art will recognize that the examples described herein are intended to help the reader understand the principles of the invention and should be understood as, but not limited to, the specific statements and embodiments described. Various modifications and variations can be made to this invention by those skilled in the art. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention are included within the scope of the claims of this invention.

Claims

1. A method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation, characterized in that, Includes the following steps: Step 1: Obtain the simulated original radar signal and modulation parameters, and then randomly overlap the simulated radar signal to generate an overlapped radar signal; Step 2: Extract the time-frequency domain features of the overlapping radar signals to obtain the signal time-frequency image; Step 3: Generate a random mask matrix, and multiply the mask matrix element-wise with the generated time-frequency image of the signal to obtain a simulated incomplete signal representation, which is used as the training set data; Step 4: Construct the ASO-Net network, which includes an encoder and a decoder. The encoder consists of three multi-branch attention feature enhancement modules and two downsampling modules cascaded alternately. The decoder consists of three multi-branch attention feature enhancement modules, two upsampling modules, and two cross-layer semantic fusion modules cascaded alternately. The multi-branch attention feature enhancement module at the end of the encoder is connected to the multi-branch attention feature enhancement module at the beginning of the decoder. There are cross-layer connections between the multi-branch attention feature enhancement modules of the encoder and the cross-layer semantic fusion modules of the decoder in the same layer. The incomplete signal representation is input into the ASO-Net network, and the output is a pixel-level classification result of the time-frequency map. Step 5: Train and test the ASO-Net network; Step 6: Use the trained ASO-Net network to identify the modulation type of overlapping radar signals, obtain the pixel-level classification results of the time-frequency map, and determine the existing signal categories by comparing them with the threshold.

2. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 1, characterized in that, The multi-branch attention feature enhancement module includes a feature extraction part and a feature fusion part; The time-frequency feature map of a given sample is input into the feature extraction part, and then passed through two convolutional layers, three parallel depthwise separable convolutional layers, and another convolutional layer. Finally, it is added to the input time-frequency feature map to obtain the output of the feature extraction part. The output features of the feature extraction part are input to the feature fusion part, and three attention branches are input in parallel: channel attention, simple pixel attention, and pixel attention. The outputs of the three attention branches are concatenated according to the channel dimension to obtain the output result of the multi-branch attention feature enhancement module.

3. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 2, characterized in that, The channel attention is that it undergoes global average pooling in sequence, Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the output of the channel attention; The simplified pixel attention has two parallel branches, one of which involves sequentially processing the input feature map. Convolutional layers and The convolutional layer, another branch, inputs the feature map and passes through it sequentially. The convolutional layer and the sigmoid activation function are combined, and finally the outputs of the two branches are multiplied to obtain the output of the simplified pixel attention. The pixel attention passes through sequentially Convolutional layers, GELU activation function, The convolutional layer and sigmoid activation function multiply the processed feature map weights with the feature map of the input channel attention to obtain the pixel attention output.

4. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 3, characterized in that, The cross-layer semantic fusion module takes the outputs of the multi-branch attention feature enhancement module and the upsampling module connected across layers as inputs, adds them together, and then performs global average pooling and multilayer perceptron to obtain weights. After the weights are normalized by Softmax, they are multiplied by the two inputs of this module to obtain weighted feature maps. Finally, the two weighted feature maps are added together and fused to obtain the output result of the cross-layer semantic fusion module.

5. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 4, characterized in that, During training, the focus loss function is used as the cost function, and the calculation method is as follows: ； in, This represents the loss weight corresponding to the true category. This represents the predicted probability of the true class. This indicates the degree of weight decay.

6. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 5, characterized in that, The original radar signal is a time-domain signal, and the modulation parameters include bandwidth, pulse width, and sampling frequency.

7. The method for identifying modulation of an indefinite number of overlapping radar signals based on semantic segmentation according to claim 6, characterized in that, The mask matrix is ​​generated as follows: First, generate a mask matrix with all 1s that has the same dimension as the time-frequency representation. Then, randomly assign zero values ​​to regions of different widths to obtain the mask matrix.

Images