A method for spectral fusion and recognition of LPI radar signals

Abstract

This invention discloses an LPI radar signal spectrum fusion and recognition method, comprising the following steps: S1, constructing a dataset, the dataset including a time-frequency graph, and time-phase and time-amplitude sequences after Fourier transform; S2, extracting features from the time-frequency data using the SKNet module; S3, sending the output of SKNet to the Transformer module for global feature association; S4, performing a concat operation on the time-phase and amplitude sequences after passing through the Transformer module, and then fusing the concat sequence with the processed time-frequency data using a multi-head attention mechanism; S5, classifying and outputting the data using a fully connected network; S6, training the network; S7, testing the data. This invention provides an LPI radar signal spectrum fusion and recognition method to address the problem of low detection accuracy in current algorithms for detecting non-cooperative LPI radar signals under low signal-to-noise ratio conditions.

Description

Technical Field

[0001] This invention relates to the field of LPI radar signal recognition technology, and in particular to an LPI radar signal spectrum fusion and recognition method. Background Technology

[0002] Intra-pulse modulation (IPM) of radar signals is a crucial component of modern electronic warfare. In electronic environments filled with noise and low signal-to-noise ratios, the identification of IPM signals becomes a critical challenge. The widespread application of low probability of intercept (LPI) radar signal identification in complex warfare environments has also made it a key focus of combat.

[0003] Traditional LPI radar modulation signal recognition primarily relies on manual feature extraction methods, which are ineffective and often fail to uncover subtle correlations between signals. Traditional modulation recognition methods depend on five conventional parameters: direction of arrival and time of arrival (DOA / TOA), radio frequency (RF), pulse width (PA), and pulse repetition interval (PRI). However, these methods struggle to represent the characteristics of complex signals and are therefore inadequate for current needs. Now, with the development of deep learning, existing results demonstrate that deep learning technology, utilizing massive amounts of data, can replace manual feature extraction techniques, overcome the effects of noise, and achieve higher accuracy. For example, Convolutional Neural Networks (CNNs) are widely used for feature extraction, and Recurrent Neural Networks (RNNs) are widely used for natural language processing. With further research, Google proposed the Transformer model in its paper "Attention is All you need," further highlighting the advantages of parallel network computing. Utilizing deep learning technology, the detection and recognition of LPI radar signals has entered a new stage. Zhang Ming performed CWD transformation on radar signals to convert them into time-frequency domain information, and used Elman Neural Network (ENN) and CNN for recognition. The former achieved an overall recognition accuracy of 94.7% at an SNR of -2dB, while the latter achieved an overall accuracy of 96.1% at -6dB. ER Zilberman used Choi-Williams distribution to perform time-frequency processing on radar signals, extracting effective features from the images, and achieved the recognition of five waveforms, including BPSK, FRANK, P4 code, FMCW, and T1 code. STFT transformation was performed on the polyphase code signals, and using a 9-layer deep convolutional network, the recognition accuracy of the five polyphase code signals reached 91.8% at -8dB. However, there is still significant room for improvement in the recognition rate of LPI radar signals under low signal-to-noise ratio conditions. Summary of the Invention

[0004] The purpose of this invention is to provide an LPI radar signal spectrum fusion and recognition method to solve the problem of low detection accuracy of current algorithms for detecting non-cooperative LPI radar signals under low signal-to-noise ratio conditions.

[0005] The technical solution adopted in the LPI radar signal spectrum fusion and identification method disclosed in this invention is as follows: A method for fusion and identification of LPI radar signal spectra includes the following steps: S1. Construct the dataset; generate 12 different LPI radar signals using MATLAB simulation, including COSTAS, LFM, Frank, BPSK, P1-P4, and T1-T4 codes. Perform time-frequency transformation and Fourier transform on the signals to obtain the time-frequency plot, time phase, and time amplitude sequence. Each sample includes three data points: time-frequency plot, time phase, and time amplitude. The samples are divided into training and test sets in an 8:2 ratio. Median filtering is performed on all sample data for preprocessing. S2, using the SKNet module to extract features from the preprocessed time-frequency image data, extracting local features of the spectrum, obtaining the importance of each channel and representing it, and outputting a feature image with multiple scales; S3. For the feature images extracted by the SKNet module, the output features processed by the channel attention mechanism module are processed by the Transformer to make up for the problem that SKNet has a weak ability to extract local features and model globally. S4 takes the two-dimensional time phase and time amplitude sequences after Fourier transform and passes them through the Transformer network to form two one-dimensional sequences. These sequences are then input into the Multi-HeadAttention submodule of the Transformer network in S3 through the Contact operation for fusion processing. S5, after the fusion processing of the fused Transformer network, 12 one-dimensional outputs are obtained through the fully connected layer to classify 12 kinds of signals; S6, trains a radar radiation source identification network model using training set data; S7. Feed the test set into the model that has been trained in step S6 for testing.

[0006] As a preferred embodiment, the detailed steps of S1 are as follows: S1.1 Using MATLAB simulation, 12 different LPI radar signals were generated. The in-phase component (I-channel) and quadrature component (Q-channel) were extracted. The I-channel and Q-channel data were saved as sampled signals of length N, ranging from 600 to 1200. Choi-Williams distribution time-frequency processing was applied to the signals to obtain a time-frequency image dataset. Fourier transform was applied to the signals to obtain two-dimensional time-phase and two-dimensional time-amplitude sequences. Image segmentation techniques were used to obtain time-frequency images with a size of 256. 256. Gaussian white noise was added during the simulation. The signal-to-noise ratio was -10dB to 10dB, with 2dB intervals. The number of samples generated for each signal under each signal-to-noise ratio was 1500. S1.2, Establish training and test sets based on the number of signal samples: maintain a training set to test set ratio of 7:3, randomly shuffle the data and labels accordingly, and denote the training dataset as... ,in Indicates the first One sample, Indicates that the sample belongs to the first The class collected a total of c types of targets. Indicates the total number of samples; S1.3, for all samples, one of which contains a time-frequency plot, a time-phase sequence, and a time-amplitude sequence, a median filtering preprocessing method is used for the time-frequency plot, with a window size set to 3. 3. Scan the image and pad the image boundaries with 0. An image of size [m, k] will become [m+2k, n+2k] after padding, where 2k+1 is the window size. After median filtering, the preprocessed image is obtained.

[0007] As a preferred embodiment, the detailed steps of S2 are as follows: The preprocessed time-frequency images are passed through a convolutional layer with parameters set to 1 input channel, 64 output channels, and a kernel size of 3. 3. Set the stride to 1 and padding to 0, then perform a two-stage SKNet operation. The first stage contains three SK units. SK unit 1 first passes through a convolutional layer with 128 output channels, stride, and padding of 1, to obtain output Z1. Z1 is then decomposed and passed through two convolutional layers, selecting two convolutional kernels, one with a 3x3 kJ / m² kernel. 3, one is 5 5. All output channels are set to 64. The first convolutional layer has a stride of 1 and padding of 1, while the second convolutional layer has a stride of 1 and padding of 2. The two outputs are superimposed, and the channel mean is calculated. A fully connected (FC) operation is then performed on each output, producing one-dimensional data with half the number of channels. This data is then expanded by FC to the same number of channels. The data is passed through a softmax function to obtain the final weight values ​​for the two channels. These weight values ​​are multiplied by the outputs of the two convolutional layers. The final outputs of the two convolutional layers are then superimposed, and finally passed through a single convolutional layer, resulting in an output with 64 channels. 2. Output Z2 is obtained and added to Z1 to obtain the output of SK unit1. SK unit2 performs the same operation as SK unit1, but both the output and input channels are set to 128. If the input and output channels are equal, no addition operation is performed. SK unit3 performs the same operation as SK unit2. Each SK unit is followed by a ReLU activation function and batch normalization. The second stage uses the output of the first stage as input, sets the channels to 256, and performs the same operation as the first stage to obtain the output of SKNet. Batch normalization uses: in, This indicates the image before batch normalization in the output of the convolutional layer. The first in the channel One element, That is, the image data after batch normalization. and For trainable corresponding to the first Parameters for each channel, It is a very small number; to prevent division by zero, its size is 10e-8. To perform the mean operation, This indicates the operation of calculating variance; The activation function uses the ReLU function: For input, The output is the ReLU response, where n represents the number of convolutional layers. After four convolutional modules, the time-frequency plot of a sample is output. , recorded as .

[0008] As a preferred embodiment, the detailed steps of S3 are as follows: S3.1, In the data p1 after SKNet operation, although the importance of each channel after the convolutional layer can be extracted, the existing model cannot effectively capture the correlation characteristics between local features. We introduce Transformer and apply Transformer to image operations. We use Transformer as the global feature correlation part and perform signal classification processing. For the output of SKNet Flatten the `to_patch` method to obtain the patch. Through linear layer mapping, Mapped to the low-dimensional data dim, the final flattened output is , where (H,W) is the resolution after feature extraction, (P,P) is the resolution of each patch, dim is the linear layer mapping size, and N is the effective sequence length of the Transformer.

[0009] As a preferred embodiment, the detailed steps of S4 are as follows: S4.1, Two-dimensional time-amplitude-phase sequence after Fourier transform 256 2. Through the Transformer network, according to 16 The data is split into 2 parts, resulting in N patches, where N is 16. These 16 patches are then mapped using a linear layer mapping. 2 is mapped to the low-dimensional data dim, where dim is 128, and the final flattened output is: C represents the number of channels, which is 1. S4.2, globally effective features are extracted through a multi-head attention mechanism. The multi-head attention mechanism divides the 128-dimensional output after segmentation and positional encoding into 8 heads, and multiplies each feature dimension by three randomly initialized matrices. , , We obtain three matrices, Q, K, and V. Self-attention is performed on each group, and then the results from the eight groups are concatenated. Each... Represented as: in To input the number of dimensions, then... arrive Concatenate column-wise, then multiply by a randomly initialized matrix. (Learnable parameters) can then be used to obtain the final globally effective features. , recorded as ; S4.3, the obtained global effective features are input into the feedforward neural network, and through linear transformation, the high-dimensional to low-dimensional transformation is achieved, further preserving the effective features; then, after passing through the ReLU activation function, the non-linear result is output. S4.4: Normalize the output of the feedforward neural network; the final output is... ; S4.5: The time phase and time amplitude are obtained separately. Perform the Contact operation to obtain the final output. ; S4.6: The output of step S4.5 The output from step S3 is then fused together with the output of the time-frequency graph Transformer in the Multi-Head Attention module.

[0010] As a preferred embodiment, the detailed steps of S5 are as follows: S5.1, the output of S5.1 is normalized through the output of the feedforward neural network, and the final output is... ; S5.2, Output The input sequence is processed through a fully connected layer. The normalized token sequence is mapped to a low-dimensional feature `dim`, where each term is calculated using the following formula: For the token sequence, the first One neuron, The weight values ​​for each neuron, This is the deviation value. The response output of the fully connected layer; S5.3, through the softmax layer The mathematical model for softmax, after classification, is expressed as follows: in Let m represent the j-th element, and m be the number of categories. This is the response of softmax.

[0011] As a preferred embodiment, the detailed steps of S6 are as follows: The preprocessed training set samples are input into the radar radiation source identification network to train the network. The network weights are updated using the Adam algorithm; the Adam algorithm is as follows: in Represented as loss function The gradient; Represented as iterative weights; Represents the gradient operator; Indicates initialization to 0 The first moment estimate; Indicated as initialized to 0 The second moment estimate; The exponential decay rate estimated by the first moment is 0.9. The exponential decay rate estimated by the second moment is 0.9. Indicates the transpose operation; The learning rate is initially set to 0.001. To ensure smoothness and prevent division by zero, the value is 10e-8; A cross-entropy loss function is used to avoid overfitting and prevent a decrease in the network's generalization ability. The cross-entropy loss function is expressed as follows: in Represents the cross-entropy loss function; This represents the true distribution of the sample; This represents the distribution predicted by the model; the smaller the cross-entropy loss function, the closer the true distribution of the samples is to the distribution predicted by the model; an early stopping mechanism is introduced, and learning rate decay is introduced with test accuracy as the standard, setting the minimum learning rate to 0; the maximum number of training rounds is set to 100 rounds, and the batch size is 32; the network model with the highest recognition accuracy is saved based on test accuracy.

[0012] As a preferred embodiment, the detailed steps of S7 are as follows: the divided test set is placed into the trained network to obtain the final result.

[0013] The beneficial effects of the LPI radar signal spectrum fusion and recognition method disclosed in this invention are as follows: 1. By using the SKNet network, which represents the channel weights after convolution, to extract features from the time-frequency map, the importance of each channel in the total features after feature extraction is reflected, further enhancing the detailed feature representation of the signal. Furthermore, by combining it with the Transformer network, local features and global features can be combined, taking a more comprehensive view. The network can learn more noteworthy features, making subsequent classification tasks more accurate and enhancing its noise resistance.

[0014] 2. By using sequence fusion, the time phase and time amplitude sequences obtained by Fourier transform are fused with the time-frequency map features extracted by SKNet through a multi-head self-attention mechanism via a Transformer network. This enhances the network's representation ability and solves the problems of unclear features in single time-frequency maps and difficulty in recognizing time-frequency maps with phase modulation. Without sacrificing significant computational costs, this approach considers the multifaceted features of different LPI radar signals and improves the network's recognition performance. Attached Figure Description

[0015] Figure 1 This is a diagram of the LPI radar signal recognition structure based on the multi-Transformer model fusion of the LPI radar signal spectrum fusion and recognition method of the present invention.

[0016] Figure 2 This is a schematic diagram of the time-frequency map feature extraction (SKNet) of the LPI radar signal spectrum fusion and recognition method of the present invention. Detailed Implementation

[0017] The present invention will be further described and illustrated below with reference to specific embodiments and the accompanying drawings: Please refer to Figure 1 and Figure 2 A method for fusion and identification of LPI radar signal spectrum includes the following steps: S1. Construct the dataset; generate 12 different LPI radar signals using MATLAB simulation, including COSTAS, LFM, Frank, BPSK, P1-P4, and T1-T4 codes. Perform time-frequency transformation and Fourier transform on the signals to obtain the time-frequency plot, time phase, and time amplitude sequence. Each sample includes three data points: time-frequency plot, time phase, and time amplitude. The samples are divided into training and test sets in an 8:2 ratio. Median filtering is then performed on all sample data for preprocessing.

[0018] The specific steps are as follows: S1.1 Using MATLAB simulation, 12 different LPI radar signals were generated, including COSTAS, LFM, Frank, BPSK, P1-P4, and T1-T4 codes. The in-phase component (I-channel) and quadrature component (Q-channel) were extracted. The I-channel and Q-channel data were each saved as sampled signals of length N, where N ranges from 600 to 1200. The signal parameters are as follows: Next, the signal was subjected to Choi-Williams distribution time-frequency processing to obtain a time-frequency graph dataset; a Fourier transform was then performed on the signal to obtain a spectrum dataset and a phase spectrum dataset. Image cropping techniques were used to obtain images with a size of 256. 256. Gaussian white noise was added during the simulation. The signal-to-noise ratio was -10dB to -10dB, with 2dB intervals. The number of samples generated for each signal under each signal-to-noise ratio was 1500.

[0019] S1.2, Establish training and test sets based on the number of signal samples: maintain the ratio of training to test sets at 8:2, randomly shuffle the data and labels accordingly, and denote the training dataset as... ,in Indicates the first One sample, Indicates that the sample belongs to the first The class collected a total of c types of targets. Indicates the total number of samples; S1.3, for all samples, one of which contains a time-frequency plot, a time-phase sequence, and a time-amplitude sequence, a median filtering preprocessing method is used, with a window size set to 3. 3. Scan the image, for example, the data {40, 107, 5, 198, 226, 223, 37, 68, 193}. After sorting, the median of 107 is obtained and set as the final value of this point. The image boundaries are padded with 0. An image of size [m, k] becomes [m+2k, n+2k] after padding, where 2k+1 is the window size. After median filtering, the preprocessed image is obtained.

[0020] S2 uses the SKNet module to extract features from the preprocessed time-frequency image data, extracts local features from the spectrogram, obtains and characterizes the importance of each channel, and outputs a feature image with multiple scales.

[0021] The specific steps are as follows: First, a convolutional layer is used, with parameters set to 1 input channel, 64 output channels, and a kernel size of 3. 3. Set the stride to 1 and padding to 0. Then, perform a two-stage SKNet operation. The first stage contains three SK units. SK unit 1 first passes through a convolutional layer with 128 output channels, and the stride and padding are both 1, to obtain the output Z1. Z1 is then decomposed through two convolutional layers. We select two convolutional kernels, one with a 3x3 kernel. 3, one is 5 5. All output channels are set to 64. The stride and padding of the first convolutional layer are set to 1, and the stride and padding of the second convolutional layer are set to 2.

[0022] The two outputs are superimposed, and the channel mean is calculated. A fully connected (FC) operation is then performed, and each output receives one-dimensional data with half the number of channels. This data is then expanded by FC again to the same number of channels. The data is passed through a softmax function to obtain the final weight values ​​for the two channels. These weights are multiplied by the outputs of the two convolutional layers. The final two convolutional layer outputs are then superimposed, and finally passed through a single convolutional layer, resulting in a 64-channel output. 2. Obtain output Z2, and add it to Z1 to get the output of SK unit1. SK unit2 performs the same operation as SK unit1, but both the output and input channels are set to 128. If the input and output channels are equal, no addition operation is performed. SK unit3 performs the same operation as SK unit2.

[0023] Each SK unit is followed by a ReLU activation function and batch normalization. The second stage takes the output of the first stage as input, sets the channels to 256, and performs the same operations as the first stage to obtain the SKNet output. Batch normalization is performed using: in, This indicates the image before batch normalization in the output of the convolutional layer. The first in the channel One element, That is, the image data after batch normalization. and For trainable corresponding to the first Parameters for each channel, It is a very small number; to prevent division by zero, its size is 10e-8. To perform the mean operation, This indicates the operation of calculating variance; The activation function uses the ReLU function: For input, This is the ReLU response output, where n represents the number of convolutional layers.

[0024] S3. For the feature images extracted by the SKNet module, the output features processed by the channel attention mechanism module are processed by the Transformer to make up for the problem that SKNet has a weak ability to extract local features and model globally.

[0025] The detailed steps of S3 are as follows: S3.1, In the data p1 after SKNet operation, although the importance of each channel after the convolutional layer can be extracted, the existing model cannot effectively capture the correlation characteristics between local features. We introduce Transformer and apply Transformer to image operations. We use Transformer as the global feature correlation part and perform signal classification processing. For the output of SKNet Flatten the `to_patch` method to obtain the patch. Through linear layer mapping, Mapped to the low-dimensional data dim, the final flattened output is , where (H,W) is the resolution after feature extraction, (P,P) is the resolution of each patch, dim is the linear layer mapping size, and N is the effective sequence length of the Transformer.

[0026] S4 takes the two-dimensional sequences of time phase and time amplitude after Fourier transform and passes them through the Transformer network to form two one-dimensional sequences. These sequences are then input into the Multi-HeadAttention submodule of the Transformer network in S3 through the Contact operation for fusion processing.

[0027] The detailed steps of S4 are as follows: S4.1, Two-dimensional time-amplitude-phase sequence after Fourier transform 256 2. Through the Transformer network, according to 16 The data is split into 2 parts, resulting in N patches, where N is 16. These 16 patches are then mapped using a linear layer mapping. 2 is mapped to the low-dimensional data dim, where dim is 128, and the final flattened output is: C represents the number of channels, which is 1. S4.2, globally effective features are extracted through a multi-head attention mechanism. The multi-head attention mechanism divides the 128-dimensional output after segmentation and positional encoding into 8 heads, and multiplies each feature dimension by three randomly initialized matrices. , , We obtain three matrices, Q, K, and V. Self-attention is performed on each group, and then the results from the eight groups are concatenated. Each... Represented as: in To input the number of dimensions, then... arrive Concatenate column-wise, then multiply by a randomly initialized matrix. (Learnable parameters) can then be used to obtain the final globally effective features. , recorded as ; S4.3, the obtained global effective features are input into the feedforward neural network, and through linear transformation, the high-dimensional to low-dimensional transformation is achieved, further preserving the effective features; then, after passing through the ReLU activation function, the non-linear result is output. S4.4: Normalize the output of the feedforward neural network; the final output is... ; S4.5: The time phase and time amplitude are obtained separately. Perform the Contact operation to obtain the final output. ; S4.6: The output of step S4.5 Together with the output of step S3, it is fused and processed in the Multi-Head Attention module of the time-frequency graph Transformer. S5, after the fusion processing of the merged Transformer network, obtains 12 one-dimensional outputs through the fully connected layer, and performs classification of 12 types of signals.

[0028] The detailed steps of S5 are as follows: S5.1, the output of S5.1 is normalized through the output of the feedforward neural network, and the final output is... ; S5.2, Output The input sequence is processed through a fully connected layer. The normalized token sequence is mapped to a low-dimensional feature `dim`, where each term is calculated using the following formula: For the token sequence, the first One neuron, The weight values ​​for each neuron, This is the deviation value. The response output of the fully connected layer; S5.3, through the softmax layer The mathematical model for softmax, after classification, is expressed as follows: in Let m represent the j-th element, and m be the number of categories. For the softmax response S6 trains a radar radiation source identification network model using training set data.

[0029] The detailed steps of S6 are as follows: The preprocessed training set samples are input into the radar radiation source identification network to train the network. The network weights are updated using the Adam algorithm; the Adam algorithm is as follows: in Represented as loss function The gradient; Represented as iterative weights; Represents the gradient operator; Indicates initialization to 0 The first moment estimate; Indicated as initialized to 0 The second moment estimate; The exponential decay rate estimated by the first moment is 0.9. The exponential decay rate estimated by the second moment is 0.9. Indicates the transpose operation; The learning rate is initially set to 0.001. To ensure smoothness and prevent division by zero, the value is 10e-8; A cross-entropy loss function is used to avoid overfitting and prevent a decrease in the network's generalization ability. The cross-entropy loss function is expressed as follows: in Represents the cross-entropy loss function; This represents the true distribution of the sample; This represents the distribution predicted by the model; the smaller the cross-entropy loss function, the closer the true distribution of the samples is to the distribution predicted by the model; an early stopping mechanism is introduced, using test accuracy as the standard, and a learning rate decay is introduced, with a minimum learning rate of 0; the maximum number of training epochs is set to 100, and the batch size is 32; based on test accuracy, the network model with the highest recognition accuracy is saved. S7. The test set is fed into the model that has been trained in step S6 for testing. The detailed steps of S7 are: put the divided test set into the trained network to obtain the final result.

[0030] This invention provides an LPI radar signal spectrum fusion and recognition method. 1. By using the SKNet network, which represents the channel weights after convolution, to extract features from the time-frequency map, the importance of each channel in the total features after feature extraction is reflected, further enhancing the detailed feature representation of the signal. Furthermore, by combining it with the Transformer network, local features and global features can be combined, taking a more comprehensive view. The network can learn more noteworthy features, making subsequent classification tasks more accurate and enhancing noise resistance.

[0031] 2. This invention uses a sequence fusion method to fuse the time phase and time amplitude sequences obtained by Fourier transform with the time-frequency map features extracted by SKNet through a multi-head self-attention mechanism. This enhances the network's representation ability and solves the problems of unclear features in a single time-frequency map and difficulty in recognizing time-frequency maps with phase modulation. Without sacrificing significant computational costs, this invention considers the multifaceted features of different LPI radar signals and improves the network's recognition performance.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for fusion and identification of LPI radar signal spectra, characterized in that, Includes the following steps: S1. Construct the dataset; generate 12 different LPI radar signals using MATLAB simulation, including COSTAS, LFM, Frank, BPSK, P1-P4, and T1-T4 codes. Perform time-frequency transformation and Fourier transform on the signals to obtain the time-frequency plot, time phase, and time amplitude sequence. Each sample includes three data points: time-frequency plot, time phase, and time amplitude. The samples are divided into training and test sets in an 8:2 ratio. Median filtering is performed on all sample data for preprocessing. S2, using the SKNet module to extract features from the preprocessed time-frequency image data, extracting local features of the spectrum, obtaining the importance of each channel and representing it, and outputting a feature image with multiple scales; S3. For the feature images extracted by the SKNet module, the output features processed by the channel attention mechanism module are processed by the Transformer to make up for the problem that SKNet has a weak ability to extract local features and model globally. S4 takes the two-dimensional time phase and time amplitude sequences after Fourier transform and passes them through the Transformer network to form two one-dimensional sequences. These sequences are then input into the Multi-HeadAttention submodule of the Transformer network in S3 through the Contact operation for fusion processing. S5, after the fusion processing of the fused Transformer network, 12 one-dimensional outputs are obtained through the fully connected layer to classify 12 kinds of signals; S6, trains a radar radiation source identification network model using training set data; S7. Feed the test set into the model that has been trained in step S6 for testing.

2. The LPI radar signal spectrum fusion and identification method as described in claim 1, characterized in that, The detailed steps of S1 are as follows: S1.1 Using MATLAB simulation, 12 different LPI radar signals were generated. The in-phase component (I-channel) and quadrature component (Q-channel) were extracted. The I-channel and Q-channel data were saved as sampled signals of length N, ranging from 600 to 1200. Choi-Williams distribution time-frequency processing was applied to the signals to obtain a time-frequency image dataset. Fourier transform was applied to the signals to obtain two-dimensional time-phase and two-dimensional time-amplitude sequences. Image segmentation techniques were used to obtain time-frequency images with a size of 256.

256. Gaussian white noise was added during the simulation. The signal-to-noise ratio was -10dB to 10dB, with 2dB intervals. The number of samples generated for each signal under each signal-to-noise ratio was 1500. S1.2, Establish training and test sets based on the number of signal samples: maintain a training set to test set ratio of 7:3, randomly shuffle the data and labels accordingly, and denote the training dataset as... ,in Indicates the first One sample, Indicates that the sample belongs to the first The class collected a total of c types of targets. Indicates the total number of samples; S1.3, for all samples, one of which contains a time-frequency plot, a time-phase sequence, and a time-amplitude sequence, a median filtering preprocessing method is used for the time-frequency plot, with a window size set to 3.

3. Scan the image and pad the image boundaries with 0. An image of size [m, k] will become [m+2k, n+2k] after padding, where 2k+1 is the window size. After median filtering, the preprocessed image is obtained.

3. The LPI radar signal spectrum fusion and identification method as described in claim 2, characterized in that, The detailed steps of S2 are as follows: The preprocessed time-frequency images are passed through a convolutional layer with parameters set to 1 input channel, 64 output channels, and a kernel size of 3.

3. Set the stride to 1 and padding to 0, then perform a two-stage SKNet operation. The first stage contains three SK units. SK unit 1 first passes through a convolutional layer with 128 output channels, stride, and padding of 1, to obtain output Z1. Z1 is then decomposed and passed through two convolutional layers, selecting two convolutional kernels, one with a 3x3 kJ / m² kernel. 3, one is 5 5. All output channels are set to 64. The first convolutional layer has a stride of 1 and padding of 1, while the second convolutional layer has a stride of 1 and padding of 2. The two outputs are superimposed, and the channel mean is calculated. A fully connected (FC) operation is then performed on each output, producing one-dimensional data with half the number of channels. This data is then expanded by FC to the same number of channels. The data is passed through a softmax function to obtain the final weight values ​​for the two channels. These weight values ​​are multiplied by the outputs of the two convolutional layers. The final outputs of the two convolutional layers are then superimposed, and finally passed through a single convolutional layer, resulting in an output with 64 channels.

2. Output Z2 is obtained and added to Z1 to obtain the output of SK unit1. SK unit2 performs the same operation as SK unit1, but both the output and input channels are set to 128. If the input and output channels are equal, no addition operation is performed. SK unit3 performs the same operation as SK unit2. Each SK unit is followed by a ReLU activation function and batch normalization. The second stage uses the output of the first stage as input, sets the channels to 256, and performs the same operation as the first stage to obtain the output of SKNet. Batch normalization uses: in, This indicates the image before batch normalization in the output of the convolutional layer. The first in the channel One element, That is, the image data after batch normalization. and For trainable corresponding to the first Parameters for each channel, It is a very small number; to prevent division by zero, its size is 10e-8. To perform the mean operation, This indicates the operation of calculating variance; The activation function uses the ReLU function: For input, The output is the ReLU response, where n represents the number of convolutional layers. After four convolutional modules, the time-frequency plot of a sample is output. , recorded as .

4. The LPI radar signal spectrum fusion and identification method as described in claim 3, characterized in that, The detailed steps of S3 are as follows: S3.1, In the data p1 after SKNet operation, although the importance of each channel after the convolutional layer can be extracted, the existing model cannot effectively capture the correlation characteristics between local features. We introduce Transformer and apply Transformer to image operations. We use Transformer as the global feature correlation part and perform signal classification processing. For the output of SKNet Flatten the `to_patch` method to obtain the patch. Through linear layer mapping, Mapped to the low-dimensional data dim, the final flattened output is , where (H,W) is the resolution after feature extraction, (P,P) is the resolution of each patch, dim is the linear layer mapping size, and N is the effective sequence length of the Transformer.

5. The LPI radar signal spectrum fusion and identification method as described in claim 4, characterized in that, The detailed steps of S4 are as follows: S4.1, Two-dimensional time-amplitude-phase sequence after Fourier transform 256 2. Through the Transformer network, according to 16 The data is split into 2 parts, resulting in N patches, where N is 16. These 16 patches are then mapped using a linear layer mapping. 2 is mapped to the low-dimensional data dim, where dim is 128, and the final flattened output is: C represents the number of channels, which is 1. S4.2, globally effective features are extracted through a multi-head attention mechanism. The multi-head attention mechanism divides the 128-dimensional output after segmentation and positional encoding into 8 heads, and multiplies each feature dimension by three randomly initialized matrices. , , We obtain three matrices: Q, K, and V. Self-attention is performed on each group, and then the results from the eight groups are concatenated. Each of these matrices... Represented as: in To input the number of dimensions, then... arrive Concatenate column-wise, then multiply by a randomly initialized matrix. This will yield the final globally effective features. , recorded as ; S4.3, the obtained global effective features are input into the feedforward neural network, and through linear transformation, the high-dimensional to low-dimensional transformation is achieved, further preserving the effective features; then, after passing through the ReLU activation function, the non-linear result is output. S4.4: Normalize the output of the feedforward neural network; the final output is... ; S4.5: The time phase and time amplitude are obtained separately. Perform the Contact operation to obtain the final output. ; S4.6: The output of step S4.5 The output from step S3 is then fused together with the output of the time-frequency graph Transformer in the Multi-Head Attention module.

6. The LPI radar signal spectrum fusion and identification method as described in claim 5, characterized in that, The detailed steps of S5 are as follows: S5.1, the output of S5.1 is normalized through the output of the feedforward neural network, and the final output is... ; S5.2, Output The input sequence is processed through a fully connected layer. The normalized token sequence is mapped to a low-dimensional feature `dim`, where each term is calculated using the following formula: For the token sequence, the first One neuron, The weight values ​​for each neuron, This is the deviation value. The response output of the fully connected layer; S5.3, through the softmax layer Based on the classification, the mathematical model of softmax is expressed as follows: in Let m represent the j-th element, and m be the number of categories. This is the response of softmax.

7. The LPI radar signal spectrum fusion and identification method as described in claim 6, characterized in that, The detailed steps of S6 are as follows: The preprocessed training set samples are input into the radar radiation source identification network to train the network. The network weights are updated using the Adam algorithm; the Adam algorithm is as follows: in Represented as loss function The gradient; Represented as iterative weights; Represents the gradient operator; Indicates initialization to 0 The first moment estimate; Indicated as initialized to 0 The second moment estimate; The exponential decay rate estimated by the first moment is 0.

9. The exponential decay rate estimated by the second moment is 0.

9. Indicates the transpose operation; The learning rate is initially set to 0.

001. To ensure smoothness and prevent division by zero, the value is 10e-8; A cross-entropy loss function is used to avoid overfitting and prevent a decrease in the network's generalization ability. The cross-entropy loss function is expressed as follows: in Represents the cross-entropy loss function; This represents the true distribution of the sample; This represents the distribution predicted by the model; the smaller the cross-entropy loss function, the closer the true distribution of the samples is to the distribution predicted by the model; an early stopping mechanism is introduced, and learning rate decay is introduced with test accuracy as the standard, setting the minimum learning rate to 0; the maximum number of training rounds is set to 100 rounds, and the batch size is 32; the network model with the highest recognition accuracy is saved based on test accuracy.

8. The LPI radar signal spectrum fusion and identification method as described in claim 7, characterized in that, The detailed steps of S7 are as follows: put the divided test set into the trained network to obtain the final result.

Images