Frequency hopping signal feature extraction method and system based on convolutional neural network

Abstract

The invention discloses a frequency hopping signal feature extraction method and system based on a convolutional neural network, relates to a signal image processing technology, and provides a scheme for solving the problem that different frequency hopping signals need to be redesigned and adjusted in the prior art. The method comprises the following steps: S1, generating a training set sample based on morphological processing; s2, constructing a convolutional neural network based on a single-channel classification network of ResNet18; s3, constructing a CrossEntroLoss loss function and an SGD (Service Gateway Diagnostic) optimizer; s4, training a convolutional neural network based on the samples; and S5, predicting a blind source frequency hopping signal. The invention further provides a system applying the method. The method has the advantages of being more intelligent and accurate; in order to solve the problem that a current frequency hopping signal time-frequency graph is too large, the single-channel convolutional neural network adopted by the invention abandons unimportant information, and the cost of network training is reduced in combination with image down-sampling.

Description

Technical Field

[0001] The present invention relates to signal image processing technology, and in particular to a frequency hopping signal feature extraction method and system based on convolutional neural network. Background Art

[0002] Traditional algorithms for frequency hopping signal feature extraction mainly include: methods based on autocorrelation analysis, methods based on time-frequency analysis, methods based on time-frequency sparsity, and methods based on spatial spectrum direction finding. The method based on autocorrelation analysis cannot solve the problem of feature extraction of multi-network station signals, and the method based on time-frequency sparsity cannot solve the problem of low signal-to-noise ratio. The method based on spatial spectrum direction finding has the disadvantage of strong dependence on hardware equipment. Therefore, most research has begun to focus on methods based on time-frequency analysis. The time-frequency analysis method mainly converts the frequency hopping signal from the time domain to the time-frequency domain for processing. However, the resolution of the time-frequency diagram is limited by Heisenberg's uncertainty law. The parameters of the blind source frequency hopping signal need to be constantly modified, which increases a lot of workload.

[0003] In traditional frequency hopping signal processing, it is usually necessary to manually design special feature extraction algorithms. These manually designed feature extraction algorithms require domain expertise and experience, and often need to be redesigned and adjusted for different frequency hopping signals. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for extracting frequency hopping signal features based on convolutional neural networks to solve the problems existing in the above-mentioned prior art.

[0005] The method for extracting frequency hopping signal features based on convolutional neural network in the present invention comprises the following steps:

[0006] S1. Generate training set samples based on morphological processing;

[0007] S2. Build a convolutional neural network based on the single-channel classification network of ResNet18;

[0008] S3. Construct CrossEntroLoss loss function and SGD optimizer;

[0009] S4. Training convolutional neural network based on samples;

[0010] S5. Predict blind source frequency hopping signals.

[0011] The frequency hopping signal feature extraction system based on convolutional neural network described in the present invention comprises a label generation module, a ResNet18 construction module, a CrossEntroLoss loss function construction module, an SGD optimizer construction module, a training module and a prediction module; and is used to apply the frequency hopping signal feature extraction method;

[0012] The label generation module is used to generate labels for training samples based on morphological processing;

[0013] The ResNet18 building module is used to build a single-channel classification network based on ResNet18;

[0014] The CrossEntroLoss loss function building module is used to build the CrossEntroLoss loss function

[0015] The SGD optimizer building module is used to build the SGD optimizer;

[0016] The training module is used to train the ResNet18 convolutional neural network based on simulated samples;

[0017] The prediction module is used to generate a prediction result based on a preprocessed blind source frequency hopping signal by using a trained convolutional neural network.

[0018] The frequency hopping signal feature extraction method and system based on convolutional neural network described in the present invention have the advantages of being more intelligent and accurate. In view of the problem that the current frequency hopping signal time-frequency diagram is too large, the single-channel convolutional neural network adopted in the present invention discards unimportant information and combines image downsampling to reduce the cost of network training.

[0019] Deep learning can learn more abstract and more representative feature representations from the original frequency hopping signal data through end-to-end training. The deep learning model can automatically learn the feature representation suitable for the task through a multi-layer convolutional neural network structure without manually designing the feature extraction algorithm. This end-to-end learning method can better adapt to different types of frequency hopping signals and tasks. For blind source frequency hopping signals, we need to consider their possible various forms. Therefore, the application of deep learning in frequency hopping signal processing requires sufficient data for training. We need to design appropriate models and training strategies for specific tasks and data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the process of the method described in the present invention.

[0021] Figure 2 It is a flow chart of generating prediction results based on preprocessed blind source frequency hopping signals using a trained convolutional neural network in the present invention.

[0022] Figure 3 It is a schematic diagram of a sample of a frequency-frequency diagram of a frequency-hopping signal before morphological processing in the present invention.

[0023] Figure 4 It is a schematic diagram of a sample of a frequency-frequency diagram of a frequency-hopping signal after morphological processing in the present invention.

[0024] Figure 5 It is a schematic diagram of a sample of a time-frequency diagram of a blind source frequency hopping signal after preprocessing in the present invention. DETAILED DESCRIPTION

[0025] The frequency hopping signal feature extraction method based on convolutional neural network described in the present invention is as follows Figures 1 to 5 As shown, the following steps are included:

[0026] S1. Generate training set samples based on morphological processing.

[0027] The characteristics of frequency hopping signals are better reflected in the time-frequency domain, which is manifested as multiple line segments of varying lengths at a certain frequency point. In general, frequency hopping signals contain many backgrounds such as fixed-frequency and swept-frequency signals and various types of noise such as white noise and colored noise. The presence of these signal noises interferes with the training of convolutional neural networks. Technically, it is more desirable for convolutional neural networks to learn the characteristics of frequency hopping signals. Therefore, such signals are removed based on morphology. At the same time, in order to ensure the training efficiency and effect of convolutional neural networks, the frequency hopping signal time-frequency diagram is downsampled and converted to.

[0028] The specific sub-steps are as follows:

[0029] S11. Remove white noise and colored noise in the time-frequency diagram based on the dynamic energy threshold algorithm.

[0030] S12. Based on the characteristics of fast hopping speed and short duration of frequency hopping signals, background signals with duration exceeding a preset value, such as fixed frequency and swept frequency signals, are removed according to the duration threshold.

[0031] S13. Downsample the frequency-frequency diagram of the frequency-hopping signal with the same number of network stations and convert it into a grayscale diagram.

[0032] After applying the algorithm in this step, the characteristics of the frequency hopping signal itself are retained, significantly reducing the interference of background and noise to the convolutional neural network.

[0033] S2. Build a convolutional neural network based on the single-channel classification network of ResNet18.

[0034] The specific sub-steps are as follows:

[0035] S21. Modify the input channel. The original ResNet18 network is designed for three-channel (RGB) images. In order to adapt it to single-channel grayscale image input, it is necessary to modify the number of input channels of the first convolutional layer from three channels to one channel to meet the grayscale image input.

[0036] S22. Build the network structure. Cascade the batch normalization layer, activation function, pooling layer, and each residual block of ResNet18 in sequence. The combination of these layers can enhance the network's feature extraction ability and nonlinear modeling ability.

[0037] S23. Modify the fully connected layer. Modify the last fully connected layer of ResNet18 to adapt to the classification task requirements. Specifically, set the output dimension of the fully connected layer to the number of target categories, that is, the number of categories of the number of stations, and add an adaptive activation function, such as softmax, to output a probability distribution.

[0038] S3. Construct the CrossEntroLoss loss function and SGD optimizer.

[0039] The specific sub-steps are as follows:

[0040] S31. Define the cross entropy loss function. The CrossEntroLoss loss function is used to measure the difference between the network output and the true label. Instantiate the loss function and prepare it for use in the training process.

[0041] S32. Instantiate the SGD optimizer and set initial parameters, including learning rate, momentum, etc. These parameters will affect the performance and training efficiency of the optimizer.

[0042] S33. Set the learning rate scheduler. In order to further improve the training effect, set the learning rate scheduler StepLR. The learning rate scheduler can adjust the learning rate according to the given step size and decay factor after each epoch, thereby helping the model to converge better.

[0043] S4. Training convolutional neural network based on samples.

[0044] The specific sub-steps are as follows:

[0045] S41. Divide the training set samples generated in step S1 into two parts, a training set and a validation set, in a certain proportion.

[0046] S42. Set the batch size, input each batch of data into the convolutional neural network to get the output, calculate the loss of the network output and the generated label based on the CrossEntroLoss loss function, and optimize the parameters of ResNet18 based on the SGD optimizer.

[0047] S43. At the end of each epoch training, the learning rate scheduler is used to update the learning rate of the SGD optimizer.

[0048] S44. Use the validation set data loader to iterate the validation set data, calculate the output through the model forward propagation, use the torch.max() function to find the maximum value and corresponding index in each sample output, and then compare it with the label to count the number of correct predictions.

[0049] S45. After all epochs are completed, the temporary weights are saved as the final weights.

[0050] S5. Predict blind source frequency hopping signals.

[0051] The specific sub-steps are as follows:

[0052] S51. Use step S1 to pre-process the time-frequency diagram of the blind source frequency hopping signal to obtain the time-frequency diagram of the frequency hopping signal that can be directly processed by the convolutional neural network.

[0053] S52. Input the processed frequency-hopping signal time-frequency diagram into the convolutional neural network to obtain the prediction result of the number of network stations.

[0054] After processing using the method described in the present invention, the accuracy is shown in Table 1 below. It is verified that the number of stations of most types of frequency hopping signals can be predicted with a high accuracy, and the performance of the model is reflected.

[0055] Table 1 Model training and testing results

[0056]

[0057] The method provided by the present invention can extract features of blind source frequency hopping signals and accurately predict the number of network stations. In order to effectively implement the method, the present invention also provides a corresponding frequency hopping signal feature extraction system, including a label generation module, a ResNet18 construction module, a CrossEntroLoss loss function construction module, an SGD optimizer construction module, a training module and a prediction module.

[0058] The label generation module is used to generate labels for training samples based on morphological processing;

[0059] The ResNet18 building module is used to build a single-channel classification network based on ResNet18;

[0060] The CrossEntroLoss loss function building module is used to build the CrossEntroLoss loss function

[0061] The SGD optimizer building module is used to build the SGD optimizer;

[0062] The training module is used to train the ResNet18 convolutional neural network based on simulated samples;

[0063] The prediction module is used to generate a prediction result based on a preprocessed blind source frequency hopping signal by using a trained convolutional neural network.

[0064] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. A frequency hopping signal feature extraction method based on a convolutional neural network, characterized in that: The following steps are involved: S1. Generate training set samples based on morphological processing; S2. Build a convolutional neural network based on the single-channel classification network of ResNet18; S3. Construct CrossEntroLoss loss function and SGD optimizer; S4. Training convolutional neural network based on samples; S5. Predict blind source frequency hopping signals.

2. The method for extracting frequency hopping signal features based on convolutional neural network according to claim 1, characterized in that: The step S1 includes the following specific sub-steps: S11. Remove white noise and colored noise in the time-frequency diagram based on the dynamic energy threshold algorithm; S12. For signals whose duration exceeds a preset value, they are removed according to the duration threshold; S13. Downsample the frequency-frequency diagram of the frequency-hopping signal with the same number of network stations and convert it into a grayscale diagram.

3. The method for extracting frequency hopping signal features based on convolutional neural network according to claim 2, characterized in that: The step S2 includes the following sub-steps: S21. Modify the input channel; modify the number of input channels of the first convolutional layer from three channels to one channel to meet the grayscale image input; S22. Build the network structure; cascade the batch normalization layer, activation function, pooling layer and each residual block of ResNet18 in sequence; S23. Modify the fully connected layer; modify the last fully connected layer of ResNet18, set the output dimension of the fully connected layer to the number of target categories, and add an adapted activation function to output a probability distribution.

4. The method for extracting frequency hopping signal features based on convolutional neural network according to claim 3, characterized in that: The step S3 includes the following sub-steps: S31. Define the cross entropy loss function; S32. Instantiate the SGD optimizer and set initial parameters; S33. Set the learning rate scheduler.

5. The method for extracting frequency hopping signal features based on convolutional neural network according to claim 4, characterized in that: The step S4 includes the following sub-steps: S41. Divide the training set samples generated in step S1 into a training set and a validation set according to a certain ratio; S42. Set the batch size, input each batch of data into the convolutional neural network to obtain the output, calculate the loss of the network output and the generated label based on the cross entropy loss function, and optimize the parameters of ResNet18 based on the SGD optimizer; S43. At the end of each epoch training, the learning rate scheduler is used to update the learning rate of the SGD optimizer; S44. Use the validation set data loader to iterate the validation set data, calculate the output through the model forward propagation, use the torch.max() function to find the maximum value and corresponding index in each sample output, and then compare it with the label to count the number of correct predictions; S45. After all epochs are completed, the temporary weights are saved as the final weights.

6. The method for extracting frequency hopping signal features based on convolutional neural network according to claim 5, characterized in that: The step S5 includes the following sub-steps: S51. Preprocess the time-frequency diagram of the blind source frequency hopping signal using step S1 to obtain a time-frequency diagram of the frequency hopping signal that can be directly processed by the convolutional neural network; S52. Input the processed frequency-hopping signal time-frequency diagram into the convolutional neural network to obtain the prediction result of the number of network stations.

7. A frequency hopping signal feature extraction system based on convolutional neural network, characterized in that: It includes a label generation module, a ResNet18 construction module, a CrossEntroLoss loss function construction module, an SGD optimizer construction module, a training module and a prediction module; and is used to apply the frequency hopping signal feature extraction method according to any one of claims 1 to 6; The label generation module is used to generate labels for training samples based on morphological processing; The ResNet18 building module is used to build a single-channel classification network based on ResNet18; The CrossEntroLoss loss function building module is used to build the CrossEntroLoss loss function The SGD optimizer building module is used to build the SGD optimizer; The training module is used to train the ResNet18 convolutional neural network based on simulated samples; The prediction module is used to generate a prediction result based on a preprocessed blind source frequency hopping signal by using a trained convolutional neural network.

Images