An ELMNet-based low probability of intercept radar signal intra-pulse modulation recognition method, system, device and program product

Abstract

The application discloses a low probability of intercept radar signal intra-pulse modulation recognition method, system, device and program product based on ELMNet, the method comprises the following steps: performing smooth pseudo-Wigner-Ville distribution time-frequency analysis on a low probability of intercept radar signal to generate a time-frequency image set; preprocessing the time-frequency image set, dividing a training set and a test set; constructing an ELMNet model, training the model using the training set, using a label smoothing cross-entropy loss function, setting the model hyperparameters and the optimizer; testing the trained model using the test set to evaluate the recognition accuracy and clustering effect. The application optimizes the feature extraction effect, expands the feature perception range and improves the feature expression ability, effectively alleviates the overfitting problem, and improves the model generalization ability; solves the problem of insufficient recognition accuracy of low probability of intercept radar signals in a low signal-to-noise ratio environment, significantly improves the recognition performance, maintains high computing efficiency, and is suitable for practical engineering applications.

Description

Technical Field

[0001] This invention belongs to the fields of radar signal processing and artificial intelligence technology, specifically relating to a method, system, device, and program product for identifying intra-pulse modulation of low probability of intercept radar signals based on ELMNet. Background Technology

[0002] Low probability of intercept (LPI) radar signals play a crucial role in modern electronic warfare due to their low power, complex modulation characteristics, and strong anti-jamming capabilities. Traditional radar signal identification methods primarily rely on manual feature extraction combined with classical classification algorithms. However, these methods have the following limitations: traditional methods require experts to design feature extraction rules, making them difficult to adapt to the complex and ever-changing nature of LPI radar signals; for signals with complex modulation types and low signal-to-noise ratios, the identification accuracy of traditional methods decreases significantly; and traditional methods have low computational efficiency when processing large-scale signal data, making it difficult to meet real-time requirements.

[0003] In recent years, deep learning technology has achieved remarkable results in fields such as image recognition and speech processing, providing new solutions for radar signal recognition. However, existing deep learning models are typically computationally complex and have a large number of parameters, making them difficult to apply directly to resource-constrained embedded devices or scenarios with high real-time requirements. Therefore, designing a lightweight and efficient deep learning model to achieve accurate recognition of radar signals with low probability of intercept has become a current research hotspot and challenge. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a low probability of intercept (POC) radar signal intra-pulse modulation identification method based on ELMNet, which significantly reduces the computational complexity and number of parameters of the model while ensuring identification accuracy, making it suitable for practical application scenarios. A system, electronic device, and software product are also provided.

[0005] Technical solution: The identification method of the present invention includes the following steps:

[0006] Smoothed pseudo-Wigner-Willy distribution time-frequency analysis is performed on low probability of intercept radar signals to generate a time-frequency image set; the time-frequency image set is then preprocessed, including grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping; the preprocessed time-frequency image set is then divided into a training set and a test set according to a preset ratio;

[0007] An ELMNet model is constructed and trained using a training set. The label-smoothed cross-entropy loss function is employed, and model hyperparameters and optimizers are set. The ELMNet model consists of a first convolutional layer, at least one EM unit structure, and a second convolutional layer. The preprocessed time-frequency image undergoes initial feature extraction through the first convolutional layer. Next, it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency. Then, the second convolutional layer further adjusts the number of channels in the feature map. Finally, a linear transformation layer outputs an unnormalized score, which is then converted to a probability distribution using the Softmax function. The EM unit structure consists of a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the ECA module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the final output.

[0008] The trained ELMNet model was tested using a test set to evaluate recognition accuracy and clustering performance.

[0009] Furthermore, the expression for the smoothed pseudo-Wigner-Villi distribution is:

[0010] ;

[0011] in, To smooth the pseudo-Wigner-Willy distribution function, For the time domain, a window function. For the frequency domain, (This is a window function) For low probability of intercept radar signals In time The value at that location, For low probability of intercept radar signals In time The complex conjugate of the value at that position. For time variables, For time delay variables, For angular frequency variables, This is the time integration variable.

[0012] Furthermore, grayscale processing is used to remove interference from color information; Wiener filtering adjusts the filter parameters based on the statistical characteristics of the input signal to minimize the mean square error of the output signal, and the smoothed time-frequency image is obtained based on the mean and variance; the calculation formula is:

[0013] ;

[0014] in, This represents the mean of the neighborhood of each pixel. This represents the variance of the neighborhood of each pixel. Indicates the noise variance. After processing, in coordinates Pixel value at the location, For the original coordinates The pixel value at the location, where , Represents the row and column indices of image pixels in a two-dimensional plane.

[0015] Furthermore, principal component extraction includes: First, determining the effective signal region through two threshold detections: The first threshold detection selects the center value of the highest classification probability as the threshold using a histogram, and the second threshold detection distinguishes signal and noise regions based on the energy values ​​of frequency components that are continuously above or below the threshold; Then, determining the start and end positions of the signal to reflect the main energy distribution of the image, thereby extracting key features.

[0016] Furthermore, adaptive cropping includes: using a bilinear interpolation algorithm to adjust the size of the time-frequency image to a preset size.

[0017] Furthermore, the label smoothing cross-entropy loss function for:

[0018] ;

[0019] ;

[0020] in, for The actual label after label smoothing The label smoothing coefficient. The original true label value, The total number of all possible categories in the classification task. The number of samples input to the model for each training session. Predict samples for the model Category The probability, For sample index, For category indexing.

[0021] Furthermore, the hyperparameters of the ELMNet model include batch size, optimizer, and learning rate. During the training process of the ELMNet model, the gradient of the label smooth cross-entropy loss value with respect to the model parameters is calculated through the backpropagation algorithm, and the model parameters are updated using the Adam optimizer to minimize the loss function.

[0022] The identification system corresponding to the method includes:

[0023] The time-frequency image set construction module is used to perform smooth pseudo-Wigner-Villier distribution time-frequency analysis on low probability of intercept radar signals and generate time-frequency image sets.

[0024] The preprocessing module is used to preprocess the time-frequency image set, including grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping.

[0025] The dataset partitioning module is used to divide the preprocessed time-frequency image set into a training set and a test set according to a preset ratio;

[0026] The ELMNet model construction and training module is used to build the ELMNet model, train the ELMNet model using the training set, employ the label-smoothed cross-entropy loss function, and set the model hyperparameters and optimizer. The ELMNet model consists of a first convolutional layer, at least one EM unit structure, and a second convolutional layer. The preprocessed time-frequency image is passed through the first convolutional layer for initial feature extraction. Then, it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency. Then, it passes through the second convolutional layer to further adjust the number of channels in the feature map. Finally, it passes through a linear transformation layer to output an unnormalized score, which is then converted into a probability distribution by the Softmax function. The EM unit structure consists of a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the efficient channel attention (ECA) module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the output.

[0027] The model testing module is used to test the trained ELMNet model using a test set to evaluate recognition accuracy and clustering performance.

[0028] An electronic device for storing and executing the method includes a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor, wherein the computer program / instructions, when executed by the processor, implement the steps of the ELMNet-based low probability of intercept radar signal intra-pulse modulation identification method.

[0029] A computer program product for storing and executing the method includes a computer program / instruction that, when executed by a processor, implements the steps of the ELMNet-based low probability of intercept radar signal intra-pulse modulation identification method.

[0030] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows: It improves the low probability of intercept radar signal recognition technology in three main aspects: (1) By image enhancement preprocessing, noise interference is reduced, redundant frequency bands are removed, and the dual time-frequency features of the signal are effectively utilized; (2) Hybrid dilated convolution and efficient channel attention module are introduced into MobileNetV2 to enhance feature extraction capability; (3) Label smoothing strategy is adopted to avoid overfitting and improve the robustness and recognition accuracy of the model. Experimental results show that compared with the existing methods, the recognition method proposed in this invention has improved performance, with a recognition accuracy of 98.98% for 12 low probability of intercept radar signals at a signal-to-noise ratio of −8 dB, and is superior to the existing network in terms of the number of parameters. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the preprocessing module in this invention;

[0033] Figure 3 The following are the preprocessing results of 12 types of radar signals in this invention: (a) is the preprocessing result of binary phase shift keying (BPSK), (b) is the preprocessing result of frequency-coded signal Costas, (c) is the preprocessing result of linear frequency modulation (LFM), (d) is the preprocessing result of nonlinear frequency modulation (NLFM), (e) to (h) are the preprocessing results of polyphase codes P1, P2, P3, and P4, respectively, and (i) to (l) are the preprocessing results of multi-time codes T1, T2, T3, and T4, respectively.

[0034] Figure 4 This is a schematic diagram of the high-efficiency MobileNetV2 unit structure in this invention;

[0035] Figure 5 This is a schematic diagram of the overall structure of the ELMNet model in this invention;

[0036] Figure 6 This is a schematic diagram comparing the recognition accuracy of different networks in a simulation experiment;

[0037] Figure 7 This is a schematic diagram comparing the recognition accuracy of different preprocessing methods in a simulation experiment.

[0038] Figure 8 This is a schematic diagram comparing heatmaps of different attention mechanisms in a simulation experiment;

[0039] Figure 9 The diagram shows the clustering of input samples and output samples trained with different loss functions in the simulation experiment. (a) is the clustering diagram of the input samples, and (b) is the clustering diagram based on... Function clustering graph, (c) is based on Function clustering graph. Detailed Implementation

[0040] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, the low probability of intercept radar intra-pulse modulation identification method based on EfficientLightweight MobileNetV2 (ELMNet) of the present invention includes the following steps:

[0042] S1. Perform smoothed pseudo-Wigner-Ville distribution (SPWVD) time-frequency analysis on low probability of intercept (LPI) radar signals to generate a time-frequency image set. LPI radar signals include binary phase-shift keying (BPSK), frequency-coded signals (Costas), linear frequency modulation (LFM), nonlinear frequency modulation (NLFM), polyphase codes (P1, P2, P3, P4), and multiple time codes (T1, T2, T3, T4). The expression for the smoothed pseudo-Wigner-Ville distribution is:

[0043] (1)

[0044] in, To smooth the pseudo-Wigner-Willy distribution function, For the time domain, a window function. For the frequency domain, (This is a window function) For low probability of intercept radar signals In time The value at that location, For low probability of intercept radar signals In time The complex conjugate of the value at that position. For time variables, For time delay variables, For angular frequency variables, This is the time integration variable.

[0045] S2. Perform grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping on the time-frequency image set. For example... Figure 2As shown, the specific steps are as follows: The original time-frequency image is converted into a grayscale image. Grayscale conversion removes interference from color information while preserving the texture and structure of the image. Wiener filtering is used to smooth background noise to obtain a Wiener grayscale image. Principal component extraction is used to identify the effective regions in the time-frequency image. Threshold detection is used to determine the start and end positions of the signal and extract key features to remove redundant information and obtain a principal component image. Adaptive cropping is performed based on the input signal-to-noise ratio parameters. A bilinear interpolation algorithm is used to adjust the height and width of the cropped time-frequency image to a preset size (224×224 in this embodiment) to obtain an adaptively cropped image that contains more effective information and reduces noise.

[0046] The preprocessed time-frequency image set was divided into a training set and a test set for model training, validation, and testing. The training set was generated within a signal-to-noise ratio (SNR) range of -8 dB to 8 dB, and the test set was generated within a SNR range of -10 dB to 10 dB.

[0047] Specifically, such as Figure 3 As shown, it includes the following steps:

[0048] S21. Image grayscale conversion is the process of converting a color time-frequency image into a grayscale image that contains only brightness information. The grayscale conversion formula is:

[0049] (2)

[0050] in, The grayscale value of the signal. This represents the time-frequency energy value of the signal. , Indices for different positions of the signal in the time domain. This represents the time-frequency energy value of the signal. This represents the number of points in the time domain of the signal. Image grayscale conversion removes interference from color information while preserving the texture and structure of the image.

[0051] S22. Wiener filtering adjusts the filter parameters based on the statistical characteristics of the input signal to minimize the mean square error of the output signal, obtaining a smoothed time-frequency image based on the mean and variance. Its calculation formula is:

[0052] (3)

[0053] (4)

[0054] (5)

[0055] in, The mean of the neighborhood of each pixel. The variance of the neighborhood of each pixel, The smoothed time-frequency image, For noise variance, This represents the number of pixels in the horizontal direction. This represents the number of pixels in the vertical direction. For the defined pixel neighborhood range, For Wiener filtering, in coordinates The pixel value at the specified location. In image processing, Wiener filtering is often used to remove noise and preserve useful signals.

[0056] S23. Principal Component Extraction (PCE) aims to extract effective regions and remove redundant parts from time-frequency images. First, the image is preprocessed to obtain the frequency distribution. Effective signal regions are then determined through two threshold detection methods. The first threshold detection uses a histogram to select the center value of the highest classification probability as the threshold. The second threshold detection distinguishes signal and noise regions based on the energy values ​​of frequency components that are consecutively above or below the threshold. Finally, the start and end positions of the signal are determined, reflecting the main energy distribution of the image, thereby extracting key features. The two-stage threshold detection method accurately determines the modulation energy region.

[0057] S24. The effective range of the signal has been determined through two threshold detections. Based on the input signal-to-noise ratio parameter, the height and width of the cropped time-frequency image are adjusted to 224×224 using a bilinear interpolation algorithm to include more effective information and reduce noise.

[0058] In the embodiments, Figure 2 The flowchart below shows the preprocessing steps for enhancing time-frequency feature images. These steps include the original time-frequency image, grayscale conversion, Wiener filtering, adaptive image cropping, and image resizing. Overall, the time-frequency image after the above preprocessing better reflects the subtle time-frequency features of the signal.

[0059] In this embodiment, time-frequency analysis was performed on 12 types of radar signals, followed by time-frequency image enhancement preprocessing. Figure 3 Figures (a) to (l) present the image preprocessing results for 12 types of radar signals at a signal-to-noise ratio of 10 dB. The 12 types of radar signals are: Binary Phase Shift Keying (BPSK), Frequency Coded Signal (Costas), Linear Frequency Modulation (LFM), Nonlinear Frequency Modulation (NLFM), Polyphase Code (P1, P2, P3, P4), and Multi-Time Code (T1, T2, T3, T4).

[0060] S25. Dataset partitioning:

[0061] The preprocessed time-frequency image dataset was divided into training and test sets according to a preset ratio: the training set was generated within a signal-to-noise ratio (SNR) range of -8 dB to 8 dB, in 2 dB increments, resulting in 27,648 time-frequency images (256 signals / class × 12 classes × 9 SNR increments). The test set was generated within a SNR range of -10 dB to 10 dB, in 2 dB increments, resulting in 11,880 time-frequency images (90 signals / class × 12 classes × 11 SNR increments). For ease of reference, the dataset of signal time-frequency images without SNR enhancement was named TFA-Original, and the dataset with enhancement was named TFA-IE.

[0062] S3. Construct the ELMNet model, using the label smooth cross-entropy loss function, and set the ELMNet model hyperparameters and optimizer. The model hyperparameters include batch size, optimizer, and learning rate.

[0063] S31. Constructing the ELMNet model:

[0064] Based on the MobileNetV2 architecture, and combining Hybrid Dilated Convolution (HDC) and Efficient Channel Attention (ECA) modules, an efficient MobileNetV2 (EM) unit structure suitable for 12 types of low probability of intercept radar signal recognition tasks is designed.

[0065] a) EM unit structure, refer to Figure 4 :

[0066] like Figure 4 As shown, the EM unit structure includes convolution, hybrid dilated convolution (HDC), inverted residual block, and efficient channel attention (ECA) module. The input feature map of the EM unit structure is X, and the output feature map is Y. X is input to the convolution module, and the output of the convolution module is connected to the input of the HDC. The output of the HDC is connected to the input of the inverted residual block, and the output of the inverted residual block is connected to the input of the ECA module. The output of the ECA module is multiplied by the output of the inverted residual block and then added to the input X to obtain the output feature map Y. The convolution operation consists of convolution, batch normalization (BN), and ReLU. The attention mechanism helps the network focus on important features and suppress unimportant features, thereby improving feature representation ability. Introducing hybrid dilated convolution into the network can effectively reduce information loss and improve the network's recognition ability.

[0067] b) Overall structure of the ELMNet model (refer to) Figure 5 :

[0068] The ELMNet model comprises a first convolutional layer, at least one EM unit structure (7 EM unit structures are used in this embodiment), and a second convolutional layer. The preprocessed time-frequency image undergoes preliminary feature extraction through the first convolutional layer; then it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency; then it passes through the second convolutional layer to further adjust the number of channels in the feature map; finally, it outputs an unnormalized score through a linear transformation layer, which is then converted into a probability distribution through the Softmax function. The EM unit structure includes a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the efficient channel attention (ECA) module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the output.

[0069] like Figure 5 As shown, the ELMNet model constructed in this invention is a highly efficient deep learning model. First, the input is a preprocessed time-frequency image with a size of 1×224×224, which passes through a convolutional layer. The purpose of this stage is to perform preliminary feature extraction on the input time-frequency image. Next, it passes through seven EM unit structures to reduce the number of parameters and improve computational efficiency. After all the EM unit structures, the model includes a final convolutional layer to further adjust the number of channels in the feature map. Finally, the model outputs an unnormalized score through a linear layer, which is then converted into a probability distribution using a Softmax function.

[0070] S32. Determine the label smoothing cross-entropy loss function:

[0071] To enhance the generalization ability of the ELMNet model, a cross-entropy loss function is introduced. In multi-class classification tasks in deep learning, the cross-entropy loss function... () is a standard method for measuring the difference between predicted probabilities and true labels. The actual labels are used, and the label smooth cross-entropy loss is applied. The label used is a smoothed label. A smoothed label is created by multiplying the actual label by a label smoothing factor less than 1. And multiply the probability value of the true label by (1 - label smoothing coefficient). The obtained values ​​are distributed equally among the other categories. This prevents the model from becoming overconfident, thereby improving its generalization ability. This invention employs a label smoothing strategy to reduce the risk of overfitting during network training. The core idea is to introduce a label smoothing coefficient. Used for smoothing one-hot code labels The function is as follows:

[0072] (6)

[0073] (7)

[0074] in, for The actual label after label smoothing The label smoothing coefficient. The original true label value, The total number of all possible categories in the classification task. The number of samples input to the model for each training session. Predict samples for the model Category The probability, For sample index, For category indexing.

[0075] S33. Model hyperparameter settings:

[0076] This invention is based on the PyTorch deep learning framework. During training, the maximum number of iterations for network training is 100. The optimization algorithm uses adaptive momentum estimation, with an initial learning rate set at 0.01, dynamically adjusted based on changes in the test loss. When the test loss fails to improve after three consecutive iterations, the learning rate is reduced to 30% of its original value to further enhance the model's optimization performance. To prevent overfitting and improve training efficiency, an early stopping strategy is employed. These settings effectively balance the model's training efficiency and generalization ability, ensuring the reliability of the experimental results.

[0077] This invention combines hybrid dilated convolution and an efficient channel attention mechanism to significantly reduce the computational complexity and number of parameters of the model while ensuring recognition accuracy, making it suitable for practical application scenarios.

[0078] S4. Train the ELMNet model using the training set, specifically including the following steps:

[0079] Initialize the parameters of the ELMNet model, including the weights and biases of the convolutional layers, the parameters of the batch normalization layers, etc.; initialize the parameters of the optimizer, including the momentum coefficient and the weight decay coefficient.

[0080] The time-frequency images in the training set are input into the ELMNet model, which is processed through convolution, EM units, linear layers, and the Softmax function to extract features and output prediction results. The label smoothing cross-entropy loss value between the prediction results and the true labels is calculated.

[0081] The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated using the Adam optimizer to minimize the loss function.

[0082] The initial learning rate was set to 0.01. When the test loss failed to improve after three consecutive tests, the learning rate was reduced to 30% of the original value to accelerate model convergence and avoid overfitting.

[0083] Record the loss value and recognition accuracy for each training cycle for subsequent analysis and model optimization.

[0084] S5. For the trained ELMNet model, test it using the test set and evaluate its recognition accuracy and clustering effect; specifically including the following steps:

[0085] Load the best model parameters saved during training.

[0086] Input the time-frequency images in the test set into the ELMNet model to obtain the model's prediction results; calculate the model's recognition accuracy on the test set, which is the ratio of the number of correctly classified samples to the total number of samples.

[0087] Ninety samples were randomly selected from each class of signal in the test set, and the high-dimensional features extracted by the model were reduced to two dimensions using the t-distributed random neighborhood embedding (t-SNE) algorithm. The feature distribution after dimensionality reduction was visualized, and the clustering effect of different classes of signals was analyzed.

[0088] The model's recognition accuracy under different signal-to-noise ratio conditions is calculated, and recognition accuracy curves are plotted. The confusion matrix of the model is calculated, and the classification error of various signals is analyzed.

[0089] The impact of different network structures, preprocessing methods, and loss functions on model performance is compared; the robustness and generalization ability of the model under low signal-to-noise ratio conditions are analyzed.

[0090] In summary, by performing time-frequency analysis and image enhancement preprocessing on 12 types of low probability of intercept radar signals, the ELMNet model was trained, and the label smoothing cross-entropy loss function was used to achieve better clustering results in the training structure, thereby improving the accuracy and computational efficiency of radar signal classification.

[0091] This invention acquires time-frequency feature images of radar signals through time-frequency analysis and designs a preprocessing module to optimize feature extraction. It improves upon the lightweight MobileNetV2 model by introducing an efficient channel attention module (ECA) and hybrid dilated convolution (HDC) to construct an ELMNet model, expanding the feature perception range and enhancing feature representation capabilities. Simultaneously, a label smoothing strategy is employed to optimize the network training process, effectively mitigating overfitting and improving model generalization ability. This invention addresses the problem of insufficient accuracy in recognizing low-probability radar signals under low signal-to-noise ratio environments, significantly improving recognition performance while maintaining high computational efficiency, making it suitable for practical engineering applications.

[0092] This invention also provides a low probability of intercept radar signal intra-pulse modulation identification system based on ELMNet, comprising:

[0093] The time-frequency image set construction module is used to perform smooth pseudo-Wigner-Villier distribution time-frequency analysis on low probability of intercept radar signals and generate time-frequency image sets.

[0094] The preprocessing module is used to preprocess the time-frequency image set, including grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping.

[0095] The dataset partitioning module is used to divide the preprocessed time-frequency image set into a training set and a test set according to a preset ratio;

[0096] The ELMNet model construction and training module is used to build the ELMNet model, train the ELMNet model using the training set, employ the label-smoothed cross-entropy loss function, and set the model hyperparameters and optimizer. The ELMNet model consists of a first convolutional layer, at least one EM unit structure, and a second convolutional layer. The preprocessed time-frequency image is passed through the first convolutional layer for initial feature extraction. Then, it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency. Then, it passes through the second convolutional layer to further adjust the number of channels in the feature map. Finally, it passes through a linear transformation layer to output an unnormalized score, which is then converted into a probability distribution by the Softmax function. The EM unit structure consists of a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the efficient channel attention (ECA) module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the output.

[0097] The model testing module is used to test the trained ELMNet model using a test set to evaluate recognition accuracy and clustering performance.

[0098] In another embodiment of the present invention, an electronic device includes a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor, characterized in that the computer program / instructions, when executed by the processor, implement the steps of the ELMNet-based low probability of intercept radar signal intra-pulse modulation identification method.

[0099] In another embodiment of the present invention, a computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the ELMNet-based low probability of intercept radar signal intra-pulse modulation identification method.

[0100] To verify the effectiveness of the present invention, the following simulation experimental data further illustrate the following:

[0101] a) Simulation parameters and simulation conditions:

[0102] The simulation involved 12 typical low probability of intercept (LOC) radar signals, including binary phase shift keying (BPSK), frequency-coded signals (Costas), linear frequency modulation (LFM), nonlinear frequency modulation (NLFM), polyphase codes (P1, P2, P3, P4), and multi-time codes (T1, T2, T3, T4). Detailed parameters are shown in Table 1. This indicates that the parameter is uniformly and randomly selected within the interval; This indicates that the parameter can take any integer value within the range; This indicates that the parameter can only take specific discrete values.

[0103] Table 1 Signal Simulation Parameter Table

[0104]

[0105] in, This represents the signal sampling rate.

[0106] b) Simulation Experiment 1: Comparing the impact of different networks on recognition accuracy. To investigate the impact of using different networks to process low probability of intercept (LPRC) radar signals on network recognition performance, five different networks—MobileNetV2, ResNet18, ResNet34, ResNet50, and GoogLeNet—were trained using the enhanced smoothed pseudo-Wigner-Villi distribution SPWVD-IE dataset. The recognition accuracy of the five networks for 12 classes of LPRC radar signals under different signal-to-noise ratios was obtained as follows: Figure 6 As shown in the line graph, MobileNetV2 performs best under low signal-to-noise ratio conditions.

[0107] c) Simulation Experiment 2: Comparing the impact of different preprocessing methods on recognition accuracy. To study the impact of different preprocessing methods on the recognition performance of low probability of intercept radar signal networks, this embodiment selected two time-frequency analysis techniques: smoothed pseudo-Wigner-Villi distribution (SPWVD) and short-time Fourier transform (STFT). Based on these two time-frequency analysis methods, corresponding image enhancement datasets were constructed: the enhanced smoothed pseudo-Wigner-Villi distribution SPWVD-IE dataset and the enhanced short-time Fourier transform (STFT-IE) dataset. Figure 7 This demonstrates the recognition accuracy of ELMNet on datasets using four different time-frequency analysis methods. (From the broken line...) Figure 7 As can be seen from the data, the SPWVD-IE dataset with enhanced smooth pseudo-Wigner-Villi distribution has the highest recognition accuracy under low signal-to-noise ratio conditions.

[0108] d) To analyze the impact of the ECA module on network recognition performance, this experiment investigated the contributions of three attention mechanisms—ECA, Coordinate Attention (CA), and Convolutional Block Attention Module (CBAM)—to time-frequency feature extraction in the network. Based on the Grad-CAM method, the results of feature extraction for 12 types of low probability of intercept radar signals after passing through networks containing ECA, CA, and CBAM attention mechanisms were visualized, and heatmaps of the three attention mechanisms in recognizing the 12 types of signals were plotted.

[0109] like Figure 8 As shown, when the network identifies the test samples mentioned above, they are easily affected by background noise, resulting in scattered and incomplete attention regions. The ECA attention mechanism exhibits more obvious attention regions in the heatmaps of the 12 signals, especially in the central region of the image, indicating that it is better able to identify and emphasize these regions. In contrast, the heatmaps of CA and CBAM show lower attention in these regions, which may mean that their attention mechanisms perform worse than those of the ECA attention mechanism.

[0110] e) Simulation Experiment 5: Effectiveness Analysis of Label Smoothing Loss. To verify the effect of label smoothing loss on network recognition performance, two loss functions were used to optimize the network proposed in this invention during the ELMNet model training process. and The function. Simultaneously, an ergonomic experiment was conducted to select the label smoothing coefficient. During the two training processes, the consistency of other factors such as dataset partitioning, model, and optimizer was controlled, and the differences in recognition performance brought about by different losses were directly compared. From the enhanced smoothed pseudo-Wigner-Villi distribution SPWVD-IE test dataset, 90 samples of each signal class were randomly selected. Under different loss functions, the t-distributed stochastic neighborhood embedding (t-SNE) method was used to project the input samples and the features extracted by the network onto a two-dimensional plane, respectively. The resulting sample clustering effect is as follows: Figure 9 As shown in (a) to (c).

[0111] Figure 9 In the diagram (a), the clustering graph of the input samples is shown. Figure 9 (b) and Figure 9 In the middle (c), it represents the cross-entropy loss function. Label smoothing cross-entropy loss function Below is a clustering diagram after dimensionality reduction of the features extracted by the network. As shown in the diagram, the signals from different categories of the input samples overlap, and the dispersion of each category is relatively large. When only cross-entropy loss is used to train the network, the time-frequency features exhibit instances where individual samples of a signal are mixed with signals from other categories. After introducing label smoothing loss, the inter-class distance of each signal clearly increases, and the clustering of signals of the same category is better, indicating that label smoothing effectively increases the separation between different signals, thereby improving the network's recognition accuracy.

[0112] This invention utilizes deep learning technology, combined with hybrid dilated convolution and efficient channel attention mechanisms, to significantly reduce the computational complexity and number of parameters of the model while ensuring recognition accuracy. It enables automatic identification and classification of radar signals with low probability of intercept, and is applicable to fields such as electronic reconnaissance, signal analysis, and military defense. It solves the problems of traditional methods relying on manual feature extraction, high computational complexity of deep learning models, and decreased recognition performance in low signal-to-noise ratio environments.

Claims

1. A method for identifying intra-pulse modulation of low probability of intercept radar signals based on ELMNet, characterized in that, Includes the following steps: Smoothed pseudo-Wigner-Villier distribution time-frequency analysis is performed on low probability of intercept radar signals to generate a time-frequency image set; The time-frequency image set is preprocessed, including grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping. The preprocessed time-frequency image set is divided into a training set and a test set according to a preset ratio; An ELMNet model is constructed and trained using a training set. The label-smoothed cross-entropy loss function is employed, and model hyperparameters and optimizers are set. The ELMNet model consists of a first convolutional layer, at least one EM unit structure, and a second convolutional layer. The preprocessed time-frequency image undergoes initial feature extraction through the first convolutional layer. Next, it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency. Then, the second convolutional layer further adjusts the number of channels in the feature map. Finally, a linear transformation layer outputs an unnormalized score, which is then converted to a probability distribution using the Softmax function. The EM unit structure consists of a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the ECA module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the final output. The trained ELMNet model was tested using a test set to evaluate recognition accuracy and clustering performance.

2. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, The expression for the smoothed pseudo-Wigner-Willy distribution is: ； in, To smooth the pseudo-Wigner-Willy distribution function, For the time domain, a window function. For the frequency domain, (This is a window function) For low probability of intercept radar signals In time The value at that location, For low probability of intercept radar signals In time The complex conjugate of the value at that position. For time variables, For time delay variables, For angular frequency variables, This is the time integration variable.

3. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, Grayscale processing is used to remove interference from color information; Wiener filtering adjusts the filter parameters based on the statistical characteristics of the input signal to minimize the mean square error of the output signal, and obtains the smoothed time-frequency image based on the mean and variance; the calculation formula is: ； in, This represents the mean of the neighborhood of each pixel. This represents the variance of the neighborhood of each pixel. Indicates the noise variance. After processing, in coordinates Pixel value at the location, For the original coordinates The pixel value at the location, where , Represents the row and column indices of image pixels in a two-dimensional plane.

4. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, Principal component extraction includes: First, determining the effective signal region through two threshold detections: the first threshold detection selects the center value of the highest classification probability as the threshold using a histogram, and the second threshold detection distinguishes signal and noise regions based on the energy values ​​of frequency components that are continuously above or below the threshold; then, determining the start and end positions of the signal to reflect the main energy distribution of the image, thereby extracting key features.

5. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, Adaptive cropping includes: using a bilinear interpolation algorithm to adjust the size of the time-frequency image to a preset size.

6. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, Label smooth cross-entropy loss function for: ； ； in, for The actual label after label smoothing The label smoothing coefficient. The original true label value, The total number of all possible categories in the classification task. The number of samples input to the model for each training session. Predict samples for the model Category The probability, For sample index, For category indexing.

7. The low probability of intercept radar signal intra-pulse modulation identification method based on ELMNet according to claim 1, characterized in that, The hyperparameters of the ELMNet model include batch size, optimizer, and learning rate. During the training process of the ELMNet model, the gradient of the label smooth cross-entropy loss value with respect to the model parameters is calculated through the backpropagation algorithm, and the model parameters are updated using the Adam optimizer to minimize the loss function.

8. A low probability of intercept radar signal intra-pulse modulation identification system based on ELMNet, characterized in that, include: The time-frequency image set construction module is used to perform smooth pseudo-Wigner-Villier distribution time-frequency analysis on low probability of intercept radar signals and generate a time-frequency image set. The preprocessing module is used to preprocess the time-frequency image set, including grayscale conversion, Wiener filtering for noise reduction, principal component extraction, and adaptive cropping. The dataset partitioning module is used to divide the preprocessed time-frequency image set into a training set and a test set according to a preset ratio; The ELMNet model construction and training module is used to build the ELMNet model, train the ELMNet model using the training set, employ the label-smoothed cross-entropy loss function, and set the model hyperparameters and optimizer. The ELMNet model consists of a first convolutional layer, at least one EM unit structure, and a second convolutional layer. The preprocessed time-frequency image is passed through the first convolutional layer for initial feature extraction. Then, it passes through at least one EM unit structure to reduce the number of parameters and improve computational efficiency. Then, it passes through the second convolutional layer to further adjust the number of channels in the feature map. Finally, it passes through a linear transformation layer to output an unnormalized score, which is then converted into a probability distribution by the Softmax function. The EM unit structure consists of a third convolutional layer, a hybrid dilated convolution (HDC), an inverted residual block, and an efficient channel attention (ECA) module connected in sequence. The output of the efficient channel attention (ECA) module is multiplied by the output of the inverted residual block and then added to the input of the EM unit structure to obtain the output. The model testing module is used to test the trained ELMNet model using a test set to evaluate recognition accuracy and clustering performance.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor, characterized in that, when the computer program / instruction is executed by the processor, it implements the steps of the low probability of intercept radar signal intra-pulse modulation identification method based on any one of claims 1-7.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the low probability of intercept radar signal intra-pulse modulation identification method based on any one of claims 1-7.

Images