Radar interference identification method based on WC3Dnet network

Abstract

The invention discloses a radar interference identification method based on a WC3Dnet wavelet convolution fusion network, and belongs to the technical field of deep learning and radar signal processing, and the method comprises the steps: 1, obtaining a radar interference time-frequency image data set, and carrying out the graying, size standardization and tensor conversion preprocessing; 2, a WC3Dnet network is constructed, the core of the WC3Dnet network is a multilayer wavelet transform fusion convolution module, image multi-scale frequency features are extracted through secondary wavelet decomposition, and feature relations are learned through convolution operation and are fused step by step; 3, training the network by adopting a cross entropy loss function and an AdamW optimizer, and only storing the model with the highest test accuracy; and 4, performing interference type classification on the test set by using the trained model. According to the method, the multi-scale analysis capability of wavelet transform and the feature extraction capability of the convolutional neural network are fused, so that the accuracy and noise immunity of image classification can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and radar signal processing technology, and is a radar interference identification method based on WC3D_net network. Background Technology

[0002] Traditional radar jamming identification methods suffer from drawbacks such as poor feature transferability and insufficient generalization ability in complex electromagnetic environments. To address this issue, the current main approach is to leverage the differences in feature parameters (e.g., frequency domain, time domain, time-frequency domain) of various jamming signals, transforming it into a pattern recognition problem, and then applying classification algorithms such as support vector machines, tree models, and logistic regression to identify the jamming patterns. However, these methods require manual extraction of highly discriminative features, exhibit poor transferability, and are ill-suited to complex and ever-changing adversarial environments.

[0003] While existing deep learning solutions can automatically learn features, conventional convolutional neural networks have limited ability to capture high-frequency details, and wavelet transform, although it has multi-scale analysis characteristics, lacks a feature learning mechanism, resulting in a significant drop in recognition rate under low signal-to-noise ratio conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, a radar interference identification method based on the WC3D_net network is proposed. This method aims to enhance the model's ability to perceive the global structure and local disturbances of interference signals, thereby significantly improving the accuracy and noise resistance of image classification and solving the problem of a significant decrease in radar interference identification rate under low signal-to-noise ratio conditions.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The radar interference identification method based on WC3D_net network of the present invention is characterized by the following steps; Step 1: Obtain the preprocessed standardized radar jamming image I; let the true class label of I be denoted as . ,and 1,2,…, }, Indicates the total number of categories; Step 2: Construct a radar interference classification network, including: M stacked units, a convolutional feature extraction layer, an average pooling layer, a Dropout layer, a convolutional layer, a linear classification layer, and a softmax layer, and then... After processing, the predicted class probability set of I is obtained. ; Step 3: Construct the cross-entropy loss function L using equation (7): (7) In equation (6), for The one-hot encoded value of the i-th category; Step 4: Train the radar jamming classification network using the AdamW optimizer and minimize the cross-entropy loss function L to update the parameters until the cross-entropy loss function L converges, thereby obtaining the optimal radar jamming classification model, which is used to predict the category of the input standardized radar jamming image.

[0006] The radar interference identification method based on WC3D_net network described in this invention is also characterized in that step 2 includes the following steps; Step 2.1: Construct M stacked units, each stacked unit including: a convolution module, a wavelet decomposition module, and a wavelet convolution fusion module; and process I to obtain the Mth secondary fusion feature. ; When m=1, I is input into the m-th stacked unit, and the convolution module obtains the m-th convolutional feature map using equation (1). ; (1) In equation (1), Conv is the convolution operation, BN is batch normalization, and ReLU is the activation function; The wavelet decomposition module in the m-th stacked unit performs first-level Haar wavelet decomposition on I to obtain the m-th first-level low-frequency sub-image LL. m,1 The m-th first-order vertical high-frequency sub-image LH m,1 The m-th first-level high-frequency sub-image HL m,1 The m-th first-order diagonal high-frequency sub-image HH m,1 ; For LL m,1 Perform second-order Haar wavelet decomposition to obtain the m-th second-order low-frequency sub-image LL. m,2 The m-th second-order vertical high-frequency sub-image LH m,2 The m-th secondary level high-frequency sub-image HL m,2 The m-th second-order diagonal high-frequency sub-image HH m,2 ; The fusion module in the m-th stacking unit uses equation (2) to obtain the m-th first-level low-frequency enhancement feature. and the m-th second-order low-frequency enhancement feature : (2) The fusion module in the m-th stacking unit uses equation (3) to obtain the m-th first-level fusion feature. and the m-th secondary fusion feature : (3) In equation (2), This indicates that bilinear interpolation is used; This indicates element-wise addition; When m = 2, 3, ..., M, the (m-1)th secondary fusion feature will be... The m-th stacked unit is processed to obtain the m-th secondary fusion feature. Thus, the Mth stacking unit outputs the Mth secondary fusion feature. ; Step 2.2, the convolutional feature extraction layer pair The convolutional activation features are obtained by sequentially performing convolution, normalization, and activation function processing. ; Step 2.3, the average pooling layer pair Average pooling is performed on each channel to obtain the pooled feature vector. ; Step 2.4: The Dropout layer randomly discards... Any element in the array yields a sparsity feature. ; Step 2.5: The convolutional layer uses equation (4) to obtain the final features. ; (4) Step 2.6, the linear classification layer will After mapping to the class space, the Softmax activation function is used to transform the mapped feature vector into a probability distribution, which is then used to classify I, resulting in a set of predicted class probabilities. .

[0007] Furthermore, step 2.6 includes the following steps; Step 2.6.1, using formula (5) to... Mapping to a C-dimensional category space, we obtain the score set y=( , … … ),in, This represents the score for the i-th category: y=W× +b (5) In equation (4), W is the weight matrix and b is the bias term; Step 2.6.2: Use equation (6) to obtain the predicted probability of I in the i-th category. Thus, the set of predicted category probabilities is obtained. ; = / (6).

[0008] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0009] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The training network of this invention uses CNN, which has a powerful feature learning capability. It can automatically learn identifiable features from the time-frequency images of interference signals without the need for manual extraction, and has strong noise resistance.

[0011] 2. This invention uses wavelet transform, which extracts low-frequency contour and high-frequency detail features through two-level wavelet decomposition. It can capture the frequency and position information of the image, which is beneficial to the recovery of image features. It has multi-level decomposition characteristics, and after learning by convolutional layers, it is upsampled and fused step by step, which can bring multi-scale information to the model. Due to its reversibility, it will not cause information loss during the transformation process, and can increase the receptive field to a certain extent. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the complete processing flow of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention; Figure 3 This is the output flowchart of the Mth stacking unit of the present invention (taking M=1 as an example); Figure 4 The present invention obtains the predicted category probability set from the Mth secondary fusion feature. flow chart; Figure 5 This is a flowchart illustrating the model training, optimization, and testing process of this invention. Detailed Implementation

[0013] In this embodiment, a radar interference identification method based on the WC3D_net wavelet convolutional fusion network is described, such as... Figure 1 As shown, the method includes: image dataset preprocessing, construction of the WC3D_net feature extraction network, construction of the complete classification network, network training and optimization, and model performance evaluation. The method comprises the following steps: Step 1: Data Collection and Labeling Step 1.1. Using a pulse compression radar (transmitting a linear frequency modulated waveform), the acquired data includes the mixed time-frequency characteristics of target echoes and interference signals. In this embodiment, eight typical interference signals are acquired, including: noise frequency modulation interference, range drag interference, velocity drag interference, angle deception interference, combined interference, comb spectrum interference, intermittent sampling interference, and noise suppression interference.

[0014] Step 1.2. Convert the acquired interference signal into a time-frequency image using short-time Fourier transform (STFT). The image resolution is 1920×1080 pixels, and a total of 20,000 time-frequency images are generated.

[0015] Step 1.3. Divide the time-frequency image dataset into a training set (14,000 images) and a test set (6,000 images) in a 7:3 ratio, and label each time-frequency image with the corresponding interference type ID (0-7).

[0016] Step 1.4. Use equation (1) to convert each time-frequency image to grayscale to obtain a single-channel grayscale image: (1) In equation (1), These are the original pixel values ​​of the time-frequency image.

[0017] Step 1.5. Use bilinear interpolation to adjust the resolution of the grayscale image to 256×256 pixels to ensure uniform input size.

[0018] Step 1.6. Normalize the pixel values ​​of the grayscale image to... The interval is converted into a PyTorch tensor, and the output shape is... The preprocessed standardized radar jamming image I; let the true category label of I be denoted as ,and 1,2,…, }, This indicates the total number of categories; the complete preprocessing workflow is as follows: Figure 2 As shown, even after the original radar signal is converted into a video image by STFT, there are still problems such as inconsistent size and channel redundancy. Therefore, a unified preprocessing process is needed to ensure the stability and reproducibility of network training.

[0019] Step 2: Construct a radar interference classification network, including: M stacked units, a convolutional feature extraction layer, an average pooling layer, a Dropout layer, a convolutional layer, a linear classification layer, and a softmax layer, and then... After processing, the predicted class probability set of I is obtained. ; Figure 3 The flowchart of the output of the Mth stacking unit when M=1 is shown, as follows: Figure 4As shown, the Mth stacking unit outputs a second-level fusion feature. The process continues through convolutional feature extraction layers, average pooling layers, dropout layers, convolutional layers, linear classification layers, and softmax layers to obtain the predicted class probability set P of radar interference categories. WC3D_net adopts a structural design of "multi-level wavelet + deep convolutional layer-by-layer fusion", which, compared with traditional single-path CNNs, can better characterize the structural differences of radar interference at multiple scales.

[0020] Step 2.1: Construct M stacked units, each stacked unit including: a convolution module, a wavelet decomposition module, and a wavelet convolution fusion module; and process I to obtain the Mth secondary fusion feature. The stacked structure enables the network to enhance its expressive power in the depth direction step by step, while each level introduces wavelet frequency band information to avoid excessive smoothing of high-frequency discrimination information in deep networks.

[0021] When m=1, I is input into the m-th stacked unit, and the convolution module obtains the m-th convolutional feature map using equation (1). ; (1) In Equation (1), Conv is the convolution operation, BN is batch normalization, and ReLU is the activation function. This branch is used to extract the local spatial texture features of the original time-frequency image, which is the "benchmark feature" for subsequent multi-scale fusion.

[0022] The wavelet decomposition module in the m-th stacked unit performs first-level Haar wavelet decomposition on I to obtain the m-th first-level low-frequency sub-image LL. m,1 The m-th first-order vertical high-frequency sub-image LH m,1 The m-th first-level high-frequency sub-image HL m,1 The m-th first-order diagonal high-frequency sub-image HH m,1 ; For LL m,1 Perform second-order Haar wavelet decomposition to obtain the m-th second-order low-frequency sub-image LL. m,2 The m-th second-order vertical high-frequency sub-image LH m,2 The m-th secondary level high-frequency sub-image HL m,2 The m-th second-order diagonal high-frequency sub-image HH m,2 The Haar wavelet has the advantages of simple computation, good orthogonality, and sensitivity to abrupt changes, making it suitable for describing sudden frequency jumps and energy edge structures in radar jamming. Only the LL component is selected for subsequent fusion because the low-frequency components retain the main structure of the jamming.

[0023] The fusion module in the m-th stacking unit uses equation (2) to obtain the m-th first-level low-frequency enhancement feature. and the m-th second-order low-frequency enhancement feature : (2) By performing feature mapping on low-frequency sub-images at different scales through convolution, the features are made consistent with the original features in the channel dimension, creating conditions for element-wise fusion.

[0024] The fusion module in the m-th stacking unit uses equation (3) to obtain the m-th first-level fusion feature. and the m-th secondary fusion feature : (3) In equation (2), This indicates that bilinear interpolation is used; This indicates element-wise addition; just as Figure 3 As shown, the two-level fusion achieves cross-wavelet scale information fusion and complementary fusion of wavelet domain and original spatial domain features, respectively. It involves element-wise addition rather than splicing.

[0025] When m = 2, 3, ..., M, the (m-1)th secondary fusion feature will be... The m-th stacked unit is processed to obtain the m-th secondary fusion feature. Thus, the Mth stacking unit outputs the Mth secondary fusion feature. This recursive stacking structure is equivalent to a "multi-scale residual enhancement network" that maintains gradient stability while increasing network depth.

[0026] Step 2.2, Convolutional Feature Extraction Layer The convolutional activation features are obtained by sequentially performing convolution, normalization, and activation function processing. This layer is used to further integrate the feature responses after multi-level fusion features to improve class separability.

[0027] Step 2.3, Average pooling layer Each channel in the image is subjected to average pooling. The average pooling layer compresses the feature map to 1×1×128 to obtain the pooled feature vector. Global average pooling can reduce the number of parameters and enhance the model's robustness to translation and local distortion.

[0028] Step 2.4: The Dropout layer sets the drop probability to 0.5 and randomly drops items. Any element in the array yields a sparsity feature. Overfitting is effectively prevented by randomly discarding data.

[0029] Step 2.5: Compress the number of channels to 8 (corresponding to 8 types of interference) using a convolutional layer. The convolutional layer then uses equation (4) to obtain the final features. ; (4) Step 2.6, the linear classification layer will After mapping to the class space, the Softmax activation function is used to transform the mapped feature vector into a probability distribution, which is then used to classify I, resulting in a set of predicted class probabilities. .

[0030] Step 2.6.1, using formula (5) to... Mapping to a C-dimensional category space, we obtain the score set y=( , … … ),in, This represents the score for the i-th category: y=W× +b (5) In equation (4), W is the weight matrix and b is the bias term.

[0031] Step 2.6.2: Use equation (6) to obtain the predicted probability of I in the i-th category. Thus, the set of predicted category probabilities is obtained. ; = / (6) Step 3: Construct the cross-entropy loss function L using equation (7): (7) In equation (6), for The one-hot encoded value of the i-th category.

[0032] Step 4: In this embodiment, the initial learning rate is set to 0.001, the weight decay is 0.01, the batch size is 64, the training period is 1000 epochs, and the learning rate is decayed to 0.1 times the original rate every 200 epochs. The AdamW optimizer is then used to train the radar interference classification network, minimizing the cross-entropy loss function L to update the parameters. Training terminates when the cross-entropy loss function L converges or the accuracy does not improve after 50 consecutive epochs. Model parameters are saved only when the test set accuracy exceeds the historical best value, thus obtaining the optimal radar interference classification model for class prediction of the input standardized radar interference image. Figure 5 As shown, the test samples are directly input into the trained network model to calculate the classification results.

[0033] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the methods described above, and the processor is configured to execute the program stored in the memory.

[0034] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A radar interference identification method based on WC3D_net network, characterized in that, Includes the following steps; Step 1: Obtain the preprocessed standardized radar jamming image I; let the true class label of I be denoted as . ,and 1,2,…, }, Indicates the total number of categories; Step 2: Construct a radar interference classification network, including: M stacked units, a convolutional feature extraction layer, an average pooling layer, a Dropout layer, a convolutional layer, a linear classification layer, and a softmax layer, and then... After processing, the predicted class probability set of I is obtained. ; Step 3: Construct the cross-entropy loss function L using equation (7): (7) In equation (6), for The one-hot encoded value of the i-th category; Step 4: Train the radar jamming classification network using the AdamW optimizer and minimize the cross-entropy loss function L to update the parameters until the cross-entropy loss function L converges, thereby obtaining the optimal radar jamming classification model, which is used to predict the category of the input standardized radar jamming image.

2. The radar interference identification method based on WC3D_net network according to claim 1, characterized in that, Step 2 includes the following steps; Step 2.1: Construct M stacked units, each stacked unit including: a convolution module, a wavelet decomposition module, and a wavelet convolution fusion module; and process I to obtain the Mth secondary fusion feature. ; When m=1, I is input into the m-th stacked unit, and the convolution module obtains the m-th convolutional feature map using equation (1). ; (1) In equation (1), Conv is the convolution operation, BN is batch normalization, and ReLU is the activation function; The wavelet decomposition module in the m-th stacked unit performs first-level Haar wavelet decomposition on I to obtain the m-th first-level low-frequency sub-image LL. m,1 The m-th first-order vertical high-frequency sub-image LH m,1 The m-th first-level high-frequency sub-image HL m,1 The m-th first-order diagonal high-frequency sub-image HH m,1 ; For LL m,1 Perform second-order Haar wavelet decomposition to obtain the m-th second-order low-frequency sub-image LL. m,2 The m-th second-order vertical high-frequency sub-image LH m,2 The m-th secondary level high-frequency sub-image HL m,2 The m-th second-order diagonal high-frequency sub-image HH m,2 ; The fusion module in the m-th stacking unit uses equation (2) to obtain the m-th first-level low-frequency enhancement feature. and the m-th second-order low-frequency enhancement feature : (2) The fusion module in the m-th stacking unit uses equation (3) to obtain the m-th first-level fusion feature. and the m-th secondary fusion feature : (3) In equation (2), This indicates that bilinear interpolation is used; This indicates element-wise addition; When m = 2, 3, ..., M, the (m-1)th secondary fusion feature will be... The m-th stacked unit is processed to obtain the m-th secondary fusion feature. Thus, the Mth stacking unit outputs the Mth secondary fusion feature. ; Step 2.2, the convolutional feature extraction layer pair The convolutional activation features are obtained by sequentially performing convolution, normalization, and activation function processing. ; Step 2.3, the average pooling layer pair Average pooling is performed on each channel to obtain the pooled feature vector. ; Step 2.4: The Dropout layer randomly discards... Any element in the array yields a sparsity feature. ; Step 2.5: The convolutional layer uses equation (4) to obtain the final features. ; (4) Step 2.6, the linear classification layer will After mapping to the class space, the Softmax activation function is used to transform the mapped feature vector into a probability distribution, which is then used to classify I, resulting in a set of predicted class probabilities. .

3. The radar interference identification method based on WC3D_net network according to claim 2, characterized in that, Step 2.6 includes the following steps; Step 2.6.1, using formula (5) to... Mapping to a C-dimensional category space, we obtain the score set y=( , … … ),in, This represents the score for the i-th category: y=W× +b (5) In equation (4), W is the weight matrix and b is the bias term; Step 2.6.2: Use equation (6) to obtain the predicted probability of I in the i-th category. Thus, the set of predicted category probabilities is obtained. ; = / (6)。 4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-3.

Images