A radar jamming signal recognition method based on an improved convolutional neural network
By introducing Inception and residual structures into convolutional neural networks and replacing fully connected layers with global average pooling layers, the problems of high computational cost and low recognition rate in radar interference identification are solved, achieving efficient and accurate radar interference signal identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2022-01-07
- Publication Date
- 2026-06-19
AI Technical Summary
Existing convolutional neural networks are computationally intensive, time-consuming, and have low accuracy in radar jamming identification, making it difficult to quickly and effectively identify enemy jamming signals in complex spatial environments.
By introducing Inception and residual structures and replacing fully connected layers with global average pooling layers, an improved convolutional neural network model is constructed, which reduces computational cost and enhances feature extraction capabilities.
It effectively reduces computational load, avoids feature loss, improves recognition efficiency and accuracy, enhances network convergence speed, and enables efficient identification of radar interference signals.
Smart Images
Figure CN116484171B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar signal recognition technology, specifically relating to a radar interference signal recognition method based on an improved convolutional neural network. Technical Background
[0002] The method of radar jamming identification using convolutional neural networks (CNNs) mainly extracts features through internal convolutional layers to achieve rapid identification and classification of received signals. Compared with traditional manual signal identification methods, the CNN-based method for radar jamming identification has advantages such as high accuracy and speed, which is of great significance for quickly identifying enemy jamming signals and taking corresponding anti-jamming measures in complex warfare environments. In addition, CNNs can effectively learn relevant features from a large number of samples, avoiding complex feature extraction processes. However, the current space environment is complex and variable, and the CNN-based radar jamming identification method generally suffers from drawbacks such as high computational cost, long processing time, and low accuracy, posing certain difficulties for radar jamming identification. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art and propose a radar interference signal identification method based on an improved convolutional neural network. This method improves the traditional convolutional neural network model by introducing an Inception structure and a residual structure, and replaces the original fully connected layer with a global average pooling layer (GAP). This solves the problems of large number of parameters, large amount of computation, overfitting, and loss of feature maps.
[0004] To achieve the above objectives, the present invention employs the following technical methods.
[0005] A radar jamming signal identification method based on an improved convolutional neural network includes the following steps:
[0006] Step 1: Establish a signal model containing interference and noise in the space, mainly including interference signals and additive white Gaussian noise signals.
[0007] Step 2: Perform wavelet transform and pulse compression on the signal model to obtain the corresponding time-frequency feature plane and pulse compression feature plane, and further perform binarization processing. Combine the processed binarized feature planes to form a new feature plane.
[0008] Step 3: Change the relevant parameters of different radar interference signals to construct a complete training sample and test sample dataset.
[0009] Step 4: Construct an improved convolutional neural network model and initialize the network parameters. Train the network using training samples, and then verify the network model's recognition performance by recognizing test samples.
[0010] Compared with the prior art, the present invention has the following beneficial technical effects:
[0011] (1) The radar interference signal identification method based on the improved convolutional neural network of the present invention introduces residual structure and Inception structure on the basis of traditional convolutional neural network model, which effectively reduces the amount of computation and avoids feature loss.
[0012] (2) The radar interference signal identification method based on the improved convolutional neural network of the present invention replaces the original fully connected layer with a global average pooling layer, which reduces network parameters, accelerates network convergence speed, and improves identification efficiency. Attached Figure Description
[0013] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0014] Figure 1 This is a flowchart of a radar jamming signal identification method based on an improved convolutional neural network.
[0015] Figure 2 a, b, c, d, and e are the joint characteristic planes after processing of amplitude-modulated noise interference, frequency-modulated noise interference, comb spectrum interference, spectral dispersion interference, and slice reconstruction interference.
[0016] Figure 3 This is a traditional convolutional neural network model.
[0017] Figure 4 For an improved convolutional neural network model.
[0018] Figure 5 This is a diagram of the Inception architecture.
[0019] Figure 6 This is a residual structure diagram.
[0020] Figure 7 A graph showing the relationship between recognition accuracy and the number of iterations.
[0021] Figure 8 The graph shows the relationship between the loss function and the number of iterations.
[0022] Figure 9 This is the confusion matrix for the specific identification results.
[0023] Figure 10This is a comparison chart showing the recognition accuracy of the improved convolutional neural network model and the traditional convolutional neural network model. Detailed Implementation
[0024] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0025] like Figure 1 As shown, the radar jamming signal identification method based on an improved convolutional neural network includes the following steps:
[0026] Step 1: Establish a signal model containing interference and noise in the space, mainly including interference signals and additive white Gaussian noise signals.
[0027] Specifically, the signal model x(t) is expressed as:
[0028] x(t)=j(t)+n(t) (1)
[0029] Where t is the time variable, j(t) represents the interference signal, and n(t) represents additive white Gaussian noise.
[0030] In embodiments of the present invention, the various radar interference signals include: frequency modulation noise interference, amplitude modulation noise interference, comb spectrum interference, spectrum dispersion interference, and slice reconstruction interference.
[0031] Step 2: Perform wavelet transform and pulse compression on the signal model to obtain the corresponding time-frequency feature plane and pulse compression feature plane, and further perform binarization processing. Combine the processed binarized feature planes to form a new feature plane.
[0032] Specifically, step 2 includes the following sub-steps:
[0033] Sub-step 2.1 involves performing wavelet transform and pulse compression on the signal model to obtain the corresponding time-frequency characteristic plane and pulse compression characteristic plane. Pulse compression utilizes the cross-correlation function between the received and transmitted signal delays to obtain the maximum peak value. The pulse compression result R(t) represents:
[0034] R(t) = s t (t)*s r (t) (2)
[0035] Where t is the time variable, R(t) is the cross-correlation function, and s t (t) represents the transmitted signal, s r (t) represents the received signal.
[0036] The basic idea of wavelet transform is to represent a signal using a set of wavelet basis functions, and to refine the signal at multiple scales by changing the scaling and translation parameters, thereby obtaining the corresponding frequency components within each time period. The wavelet transform WT(a,τ) is defined as follows:
[0037]
[0038] Where t is the time variable, f(t) is the original signal, a is the scale parameter, τ is the translation parameter, the wavelet basis is the CMO basis, and ψ(t) is the window function.
[0039] Sub-step 2.2: The feature plane obtained after the above processing is assumed to be an M×N dimensional matrix. For each row number i and column number j, there is an amplitude value f(i, j), where i = 1, 2, ..., M, j = 1, 2, ..., N. The original plane is binarized by setting a threshold value κ, as shown in the following formula:
[0040]
[0041] Where f′(i,j) represents the amplitude value of the (i,j)th position after the above processing, 0 represents black, and 255 represents white.
[0042] Sub-step 2.3: The time-frequency feature plane and pulse compression feature plane obtained after the above processing have a size of M×N, and the combined feature plane after column-wise combination has a size of 2M×N. In this example, the feature plane size after column-wise combination is 256×256. The specific feature map is shown below. Figure 2 As shown.
[0043] Step 3: Change the relevant parameters of different radar interference signals to construct training and test sample datasets.
[0044] Specifically, to ensure the robustness of the algorithm, a complete training and testing dataset was constructed by changing the relevant parameters of different radar interferences to ensure ergodicity. The training dataset contains 4000 sets of samples, and the testing dataset contains 1000 sets of samples.
[0045] Step 4: Construct an improved convolutional neural network model. Train the network using training samples to obtain a radar interference recognition model. Then input test samples to verify the network's recognition performance.
[0046] Specifically, step 4 includes the following sub-steps:
[0047] Sub-step 4.1 involves improving upon the traditional convolutional neural network model. The traditional and improved convolutional neural network models are as follows: Figure 3 and Figure 4As shown in the figure (using a 64×64 feature map as an example), both consist of three convolutional layers and three pooling layers. The convolutional layers use 5×5 kernels, and the pooling layers use 2×2 sampling windows. The difference lies in that, compared to the traditional convolutional network model, the improved model introduces an Inception structure and a residual structure, and replaces the original fully connected layer with a global average pooling layer in the last layer. The Inception structure and the residual structure are shown below. Figure 5 and Figure 6 As shown.
[0048] The Inception structure, by parallelizing convolutional kernels of different sizes, allows for different receptive fields in each channel, thereby extracting richer feature information. This structure contains multiple 1×1 convolutional kernels, which enhances the non-linear expressive power of the convolutional neural network, increases network depth, and effectively reduces the number of parameters. The residual structure propagates the feature information extracted by the pooling layer backward while simultaneously inputting it into the feature union layer, merging it with the feature map information from the convolutional layers in terms of channel count. This ensures that the feature union layer possesses both the low-level detail information extracted by the pooling layers and the global features extracted by the convolutional layers, effectively preventing feature loss.
[0049] Global average pooling layers average the entire feature map. Instead of a one-dimensional array, each class in the output is represented by a specific numerical value, which does not change the classification result of the network.
[0050] Assuming the feature maps input to both the fully connected layer and the global average pooling layer are of size m×N×N, then the computational cost required for the fully connected layer is m. 2 N 2 The computational cost of the global average pooling layer is N. 2 Global average pooling layers structurally regularize the entire network to prevent overfitting, while functionally being identical to fully connected layers. This improvement better integrates image features, avoiding feature loss, and effectively reduces the number of network model parameters and computational cost, while also preventing overfitting and gradient vanishing.
[0051] Sub-step 4.2 initializes the parameters of the improved convolutional neural network, including: selecting ReLU (Rectified Linear Unit) as the activation function, setting the amount of data used per iteration to 10, the maximum number of training epochs to 5, selecting the data shuffling strategy to shuffle once per epoch, setting the initial learning rate to 0.0001, and setting the validation frequency to 6.
[0052] Sub-step 4.3 involves performing relevant simulations using a computer. First, the network model is trained using all the training sample data. Then, test sample data is used to perform recognition and test its recognition performance. The relevant simulation parameter settings are shown in Table 1.
[0053] Table 1 Simulation parameter settings
[0054]
[0055] The recognition accuracy curve and the loss function curve are as follows: Figure 7 and Figure 8 As shown, the confusion matrix of the final identification result is as follows: Figure 9 As shown in the figure, the overall recognition accuracy based on the improved convolutional neural network model reaches 98.5%.
[0056] Figure 10 Line graphs show the accuracy of the improved convolutional neural network (CNN) model and the traditional CNN model under different signal-to-noise ratios (SNRs). Analysis of the graphs shows that the improved CNN model has a significantly higher recognition accuracy than the traditional CNN model, achieving nearly 100% accuracy at 10dB.
[0057] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A radar jamming signal identification method based on an improved convolutional neural network, characterized in that, Includes the following steps: Step 1: Establish a signal model containing interference and noise in the space, mainly including interference signals and additive white Gaussian noise signals; Step 2: Perform wavelet transform and pulse compression on the signal model to obtain the corresponding time-frequency feature plane and pulse compression feature plane, and further perform binarization processing. Combine the processed binarized feature planes to form a new feature plane. Step 3: Change the relevant parameters of different radar jamming to construct a complete training sample and test sample dataset; Step 4: Construct an improved convolutional neural network model by introducing the Inception structure and residual structure, replacing the last fully connected layer with a global average pooling layer, and initializing the network parameters; train the network based on training samples, and then verify the recognition performance of the network model by recognizing test samples. Step 2 includes the following sub-steps: Sub-step 2.1 involves performing wavelet transform and pulse compression on the signal model to obtain the corresponding time-frequency characteristic plane and pulse compression characteristic plane. The pulse compression result R(t) represents: R(t)=s t (t)*s r (t) Where t is the time variable, s t (t) represents the transmitted signal, s r (t) represents the received signal; The wavelet transform WT(a,τ) is defined as follows: Where t is the time variable, f(t) is the original signal, a is the scale parameter, τ is the translation parameter, the wavelet basis adopts the cmor basis, and ψ(t) is the window function; Sub-step 2.2: The feature plane obtained after the above processing is assumed to be an M×N dimensional matrix. For each row number i and column number j, there is an amplitude value f(i, j), where i = 1, 2, ..., M, j = 1, 2, ..., N. The original plane is binarized by setting a threshold value κ, as shown in the following formula: Where f′(i,j) represents the amplitude value of the (i,j)th position after the above processing, 0 represents black, and 255 represents white; Sub-step 2.3: The time-frequency feature plane and pulse compression feature plane obtained after the above processing have a size of M×N, and the joint feature plane after column combination has a size of 2M×N.
2. The radar jamming signal identification method based on the improved convolutional neural network according to claim 1, step 1 is specifically as follows: Specifically, a signal model containing interference and noise in space is established, mainly including interference signals and additive white Gaussian noise signals; x(t) = j(t) + n(t) in, j(t) represents the interference signal, and n(t) represents additive white Gaussian noise.
3. The radar jamming signal identification method based on the improved convolutional neural network according to claim 1, step 3 is specifically as follows: Specifically, to ensure the robustness of the algorithm, a complete training and testing dataset is constructed by changing relevant parameters of different radar jamming, such as jamming interval, delay, and period, to ensure ergodicity.
4. The radar jamming signal identification method based on the improved convolutional neural network according to claim 1, step 4 includes the following sub-steps: Sub-step 4.1 introduces the Inception structure and the residual structure, and replaces the last fully connected layer of the network model with a global average pooling layer; Sub-step 4.2: Initialize the parameters of the improved convolutional neural network; Sub-step 4.3 involves using a computer to perform relevant simulations. First, the network model is trained using all the training sample data, and then the recognition performance is tested using the test sample data.