Interference signal identification method, system and device based on artificial feature map and convolutional neural network, and medium

Abstract

The invention discloses an interference signal identification method, system and device based on the combination of an artificial feature map and a convolutional neural network, and a medium, and the method comprises the steps: interference signal preprocessing, signal time domain segmentation, artificial feature extraction, construction and training of the convolutional neural network, obtaining of an interference signal identification model, and testing of the interference signal identification model. Obtaining a test result; the system, the equipment and the medium are used for realizing the method. According to the method, the artificial feature map is combined with the convolutional neural network, so that the calculation complexity of the convolutional neural network can be reduced while the efficient recognition performance is ensured, and the feasibility of hardware implementation is improved; the technical problems of information redundancy and overlarge network parameter quantity of input data of a neural network in an existing interference signal identification method based on deep learning are solved.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and in particular relates to an interference signal recognition method, system, device and medium based on artificial feature graph and convolutional neural network. Background Art

[0002] With the rapid development of modern communication, radar, navigation, electronic warfare and other technologies, interference signal identification plays an increasingly important role in the field of wireless communication and signal processing. Interference signals not only cause the signal quality of the communication system to deteriorate, but may even affect the stability and security of the system. Therefore, accurately identifying interference signals and taking timely countermeasures are the key to ensuring the normal operation of the communication system.

[0003] Typical interference signal recognition methods are mainly divided into three parts: interference signal preprocessing, interference signal feature extraction, and interference signal classification. According to whether the feature extraction is artificial, it can be divided into artificial feature extraction and automatic feature extraction. Artificial feature extraction is mainly composed of feature parameters with physical meaning or feature parameters defined by experts with rich knowledge and experience. Automatic feature extraction mainly uses neural networks to perform deep feature extraction on the input data. After feature extraction, common classification methods include decision trees, support vector machines, and fully connected neural networks. As deep learning technology continues to make breakthroughs in image processing, natural language processing, and other fields, its powerful capabilities in feature extraction and pattern recognition have been widely verified. This technical advantage has also gradually penetrated into the field of interference signal recognition, making the interference signal recognition method based on deep learning a hot research direction. The main implementation steps of the interference signal recognition method based on deep learning are to input the time domain or frequency domain or time-frequency diagram information of the interference signal into the designed feature extraction network for automatic feature extraction. These feature extraction networks can be one of the convolutional neural networks, recurrent neural networks, etc., and then send the extracted feature vectors to the fully connected network for classification and recognition.

[0004] However, interference signal recognition methods based on deep learning have the following two problems: (1) The original signal in the time domain or frequency domain is usually directly used as the input of the neural network, which makes the input data contain a large amount of redundant information, greatly increases the difficulty of training, and makes the convergence process of the neural network more complicated and time-consuming; (2) In order to explore the deep features of the original signal in the time domain or frequency domain to improve the recognition performance, deep learning-based methods usually need to stack more neural network layers, which leads to an excessively large number of network parameters, bringing huge challenges to engineering implementation and almost impossible to complete efficiently on hardware.

[0005] The Chinese invention patent with the publication number CN108509911B proposes a method for identifying interference signals based on a convolutional neural network. Its input layer uses a sample format of N×1×3 for input, where N is a positive integer, and 3 represents that the three dimensions of the input sample are the real part, imaginary part, and frequency-domain amplitude spectrum of the complex baseband signal. The design of the convolutional neural network model uses 2 convolutional layers, 4 Inception layers, and 1 fully connected layer. Although this method can achieve a 100% correct recognition rate for each signal when the signal-to-interference ratio is higher than -2dB, the number of network layers and network parameters used are too large, which is not conducive to engineering implementation; Shang Gaoyang (Shang Gaoyang. Research and Implementation of Wireless Communication Interference Signal Cognition Technology [D]. University of Electronic Science and Technology of China, 2023. DOI: 10.27005 / d.cnki.gdzku.2023.003219.) proposed a method for identifying interference signals based on a convolutional neural network combined with joint multi-domain features. This method inputs the frequency-domain Welch spectrum sequence and the time-frequency diagram into two sub-neural networks respectively to extract feature vectors, and then fuses the feature vectors in the two domains and inputs them to the fully connected layer for classification and recognition. The neural network model of this method uses 7 convolutional layers, 7 BN layers (Batch Normalization), and 2 fully connected layers, and can achieve a 100% correct recognition rate for 8 interference signals when the signal-to-interference ratio is greater than -6dB, but the number of model parameters of this neural network is also too large and is not conducive to being deployed on a hardware platform with limited resources. Summary of the Invention

[0006] In order to overcome the above-mentioned shortcomings in the prior art, the purpose of the present invention is to provide a method, system, device, and medium for identifying interference signals based on the combination of artificial feature maps and convolutional neural networks. By combining artificial feature maps with convolutional neural networks, it is possible to reduce the computational complexity of the convolutional neural network while ensuring high recognition performance, improve the feasibility of hardware implementation, and solve the technical problems of information redundancy in the input data of the neural network and excessive network parameters in the existing interference signal recognition methods based on deep learning.

[0007] In order to achieve the above-mentioned invention purpose, the technical solutions adopted by the present invention are as follows:

[0008] A method for identifying interference signals based on the combination of artificial feature maps and convolutional neural networks, comprising the following steps:

[0009] Step 1: Preprocessing of interference signals, performing power normalization processing on each sample in the generated original interference signal data;

[0010] Step 2: Signal time-domain segmentation. Each sample in the original data of the preprocessed interference signal is continuously segmented in the time domain to obtain time-domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32;

[0011] Step 3: Artificial feature extraction. Select N adjacent time-domain segments in each sample, perform artificial feature extraction on each time-domain segment respectively, and construct an artificial feature map with a size of 9×N;

[0012] Step 4: Construct a convolutional neural network and train it to obtain an interference signal recognition model;

[0013] Step 5: Input the artificial feature map into the interference signal recognition model for testing to obtain the test results.

[0014] The feature categories in step 3 include frequency-domain moment skewness, frequency-domain moment kurtosis, spectral kurtosis, frequency-domain envelope fluctuation degree, average spectral flatness coefficient, time-domain moment skewness, time-domain moment kurtosis, time-domain peak-to-average ratio, and time-domain envelope fluctuation degree.

[0015] In step 3, the feature data obtained from each sample is used to construct a feature map. That is, the dimension of the time-domain segment is used as the length of the artificial feature map, and the feature category dimension is used as the width of the artificial feature map, so as to construct an artificial feature map with a length of N and a width of 9.

[0016] In step 3, the feature data of the artificial feature map is preprocessed and normalized in sequence.

[0017] First, the frequency-domain feature data in the artificial feature map is separately taken the logarithm base 2, logarithm base 10, or square root for data preprocessing, while keeping the time-domain feature data unchanged;

[0018] Then, the maximum and minimum normalization processing is respectively performed on the time-domain and frequency-domain feature values, and each feature value in the entire artificial feature map is restricted between 0 and 1.

[0019] The input size of the convolutional neural network constructed in step 4 is 9×N. The first layer is a convolutional layer with a convolutional kernel size of 3×3, the number of channels is 6, and the activation function used for the output is leakyrelu; the second layer is a convolutional layer with a convolutional kernel size of 3×3, the number of channels is 16, and the activation function used for the output is leakyrelu; the third layer is a convolutional layer with a convolutional kernel size of 3×3, the number of channels is 120, and the activation function used for the output is leakyrelu; the fourth layer is a fully connected layer, the number of input and output neurons is 120 and 84 respectively, and the activation function used is leakyrelu; the fifth layer is a fully connected layer, the number of input and output neurons is 84 and M respectively; the sixth layer is a softmax layer, and the recognition probabilities of M types of interference signals are output.

[0020] The present invention also provides an interference signal recognition system based on artificial feature maps and convolutional neural networks, including:

[0021] An interference signal preprocessing module: used to perform power normalization processing on each sample in the generated raw interference signal data;

[0022] A signal time-domain segmentation module: continuously segment each sample in the preprocessed raw interference signal data in the time-domain dimension to obtain time-domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32;

[0023] An artificial feature extraction module: select N adjacent time-domain segments in each sample, perform artificial feature extraction on each time-domain segment respectively, and construct an artificial feature map with a size of 9×N;

[0024] A convolutional neural network construction and training module: used to construct a convolutional neural network and perform training to obtain an interference signal recognition model;

[0025] A testing module: input the artificial feature map into the interference signal recognition model for testing to obtain a test result.

[0026] The present invention also provides an interference signal recognition device based on artificial feature maps and convolutional neural networks, including:

[0027] A memory: stores a computer program for the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks, and is a computer-readable device;

[0028] A processor: used to implement the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks when executing the computer program.

[0029] The present invention also provides a computer-readable storage medium, which stores a computer program that can implement the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks when executed by a processor.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. Reduces the redundancy of input data: Existing methods usually directly use the raw signals in the time domain or frequency domain as input, which will cause a large amount of redundant information in the input data, increasing the training difficulty and computational burden. However, the present invention converts the raw signals into representative feature maps through artificial feature extraction, thus greatly reducing the redundant information. This enables the neural network to focus more on key features and improves the training efficiency.

[0032] 2. Reduced the number of network parameters and computational complexity: Existing methods usually require stacking a large number of network layers, resulting in an excessive number of neural network parameters and high computational complexity, making it difficult to achieve efficient hardware deployment. In contrast, the present invention constructs an artificial feature map with a size of 9×N, reducing the dimension of the neural network input data to a smaller size (9×N). At the same time, relatively few convolutional layers and fully connected layers are used, and the network structure is simpler compared to traditional deep learning methods, effectively reducing the number of network parameters and computational complexity, making it easier to be deployed and run on hardware platforms with limited resources.

[0033] 3. More efficient feature extraction: By selecting multiple key feature categories in the frequency domain and time domain, the present invention can effectively capture the useful information in the signal, making the identification of interference signals more accurate and robust. In contrast, most existing methods rely on neural networks to automatically extract features, which may ignore some physical features crucial for identification.

[0034] 4. Better recognition performance and training efficiency: The artificial feature map adopted by the present invention not only retains the signal features with actual physical significance, but also further optimizes the distribution of feature data in the artificial feature map by performing log2 processing on the frequency domain feature data and performing maximum-minimum normalization processing on the time domain and frequency domain feature data respectively, enabling the convolutional neural network to converge more quickly during the training process, thereby improving the recognition progress and reducing the training time.

[0035] In summary, compared with the prior art, by combining the artificial feature map with the convolutional neural network, the present invention ensures high recognition performance while reducing the computational complexity of the convolutional neural network, improving the feasibility of hardware implementation, and solving the technical problems of information redundancy in the input data of the neural network and excessive network parameters in the existing interference signal recognition methods based on deep learning. Description of the Drawings

[0036] Figure 1 It is a schematic flow chart of an interference signal recognition method based on an artificial feature map and a convolutional neural network.

[0037] Figure 2 It is a grayscale image of the artificial feature map provided by the embodiment.

[0038] Figure 3 It is a grayscale image of the artificial feature map provided by the comparative example.

[0039] Figure 4 It is a schematic structural diagram of a convolutional neural network.

[0040] Figure 5 It is a loss value curve during the training process of the interference signal recognition model provided by the embodiment.

[0041] Figure 6 The accuracy curve of the interference signal recognition model provided for the embodiment during the training process.

[0042] Figure 7 The confusion matrix of the interference signal recognition model provided for the embodiment for the recognition output of the test set.

[0043] Figure 8 The interference recognition accuracy curve of the interference signal recognition model provided for the embodiment under different signal-to-interference-plus-noise ratio (SINR) environments. Detailed implementation manners

[0044] The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments.

[0045] As Figure 1 shown, an interference signal recognition method based on the combination of artificial feature maps and convolutional neural networks includes the following steps:

[0046] Step 1: Generate interference signals. Within a certain SINR range, several different jamming signals are simulated and generated as the original interference signal data;

[0047] Considering that the channel type is an additive white Gaussian noise (AWGN) channel, the SINR range is from -20 to 20 dB with an interval of 2 dB. Under each SINR condition, 1000 samples of each type of interference signal are randomly generated in Matlab software as the original interference signal data. The types of the different jamming signals include 6 types: single-tone interference signal, multi-tone interference signal, linear frequency sweep interference signal, noise amplitude modulation interference signal, noise frequency modulation interference signal, and narrowband noise interference signal;

[0048] The mathematical model of the single-tone interference signal is expressed as:

[0049]

[0050] where P J represents the interference power, f J represents the interference center frequency, θ J represents the initial phase that follows a uniform distribution on [0, 2π), and j represents the imaginary unit;

[0051] The mathematical model of the multi-tone interference signal is expressed as:

[0052]

[0053] where N represents the number of tones, f i represents the center frequency of the i-th tone, and θ i represents the initial phase of the i-th tone that also follows a uniform distribution on [0, 2π);

[0054] The mathematical model of the linear frequency sweep interference signal is expressed as:

[0055]

[0056] In the formula, L(τ) represents a periodic linear signal, and K FM represents the frequency modulation slope, also known as the frequency modulation sensitivity;

[0057] The mathematical model of the noise amplitude modulation interference signal is expressed as:

[0058] J 4 (t) = [U 0 + U n (t)] exp[j(2πf J t + θ J )]

[0059] In the formula, U 0 represents the amplitude of the carrier signal, and U n (t) represents a band-limited Gaussian white noise signal, and the interference bandwidth is basically the same as the bandwidth of the band-limited Gaussian white noise;

[0060] The mathematical model of the noise frequency modulation interference signal is expressed as:

[0061]

[0062] In the formula, U n (t) represents a band-limited Gaussian white noise signal;

[0063] The mathematical model of the narrowband noise interference signal is expressed as:

[0064] J 6 (t) = U n (t) exp[j(2πf J t + θ J )]

[0065] In the formula, U n (t) represents a band-limited Gaussian white noise signal;

[0066] Step 2: Preprocess the interference signal, and perform power normalization processing on each sample in the original interference signal data generated;

[0067] The formula for power normalization processing is as follows:

[0068]

[0069] In the formula, j p (n) represents the interference signal after power normalization processing, j(n) represents the interference signal before power normalization processing, and N represents the number of time-domain samples of the interference signal;

[0070] Step 3: Segment the signal in the time domain. Continuously segment each sample in the original data of the preprocessed interference signal in the time domain dimension to obtain time domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32;

[0071] In this embodiment, each sample is continuously segmented into 32 time domain segments of equal length in the time domain dimension;

[0072] Step 4: Manually extract features. Select N adjacent time domain segments in each sample, and manually extract features from each time domain segment respectively. The feature categories include frequency domain moment skewness, frequency domain moment kurtosis, spectral kurtosis, frequency domain envelope fluctuation degree, average spectral flatness coefficient, time domain moment skewness, time domain moment kurtosis, time domain peak-to-average ratio, and time domain envelope fluctuation degree, a total of nine feature categories;

[0073] In this embodiment, features are extracted from 32 time domain segments of each sample respectively, and the following nine feature categories are selected for the features:

[0074] The calculation formula for the frequency domain moment skewness is as follows:

[0075]

[0076] In the formula, P(k) represents the amplitude spectrum of the interference signal, μ represents the mean value of P(k), σ represents the standard deviation of P(k), and E represents the expected value, that is, the statistical average of (P(k) - μ) 3 is performed;

[0077] The calculation formula for the frequency domain moment kurtosis is as follows:

[0078]

[0079] The calculation formula for the spectral kurtosis is as follows:

[0080]

[0081] In the formula, max(P(k)) represents the maximum value of the interference signal amplitude spectrum;

[0082] The calculation formula for the frequency domain envelope fluctuation degree is as follows:

[0083]

[0084] The calculation formula for the average spectral flatness coefficient is as follows:

[0085]

[0086] In the formula, S p (k) = S u (k) - Sa (k), S u (k) represents the power spectrum of the interference signal after mean normalization. For S u (k), perform moving average filtering to obtain S a (k), and set the moving window length to 7;

[0087] The calculation formula for the skewness of the time-domain moment is as follows:

[0088]

[0089] In the formula, j(n) represents the time-domain sample points of the interference signal, and μ j represents the mean value of j(n), and σ j represents the standard deviation of j(n);

[0090] The calculation formula for the kurtosis of the time-domain moment is as follows;

[0091]

[0092] The calculation formula for the peak-to-average ratio of the time domain is as follows:

[0093]

[0094] In the formula, max(|j(n)|) represents the maximum value of the time-domain sample points of the interference signal;

[0095] The calculation formula for the envelope fluctuation degree of the time domain is as follows:

[0096]

[0097] After the artificial feature extraction process in step 4, 9×32 feature data can be obtained;

[0098] Step 5: Construct an artificial feature map. For the feature data obtained from each sample, construct a feature map, that is, use the dimension of the time-domain segment as the length of the artificial feature map and the dimension of the feature category as the width of the artificial feature map, so as to construct an artificial feature map with a length of N and a width of 9;

[0099] In this embodiment, an artificial feature map with a length of 32 and a width of 9 is constructed.

[0100] Step 6: Perform preprocessing and normalization on the feature data of the artificial feature map in sequence;

[0101] First, perform data preprocessing on the frequency-domain feature data in the artificial feature map by taking log2, log10, or square root separately, while keeping the time-domain feature data unchanged. Because the differences in various frequency-domain feature data are too large, after normalization, the data distribution may be concentrated near 0 and 1, which is not conducive to the convergence of the neural network. Taking log2 processing can solve this problem;

[0102] Then, perform min-max normalization on the time-domain and frequency-domain eigenvalues respectively, restricting each eigenvalue in the entire artificial feature map to be between 0 and 1.

[0103] Comparative example: Directly perform global normalization on the feature data of the artificial feature map to obtain an artificial feature map, which is a grayscale map generated using Matlab software, as Figure 3 shown; in the grayscale map, the closer the pixel point is to white, the closer the eigenvalue of that point is to 1, and vice versa, the closer it is to black, the closer the eigenvalue of that point is to 0. Figure 3 In the grayscale map in , only the pixel points of the eigenvalues of one category (frequency-domain moment kurtosis) approach white, and the pixel points of the eigenvalues of other categories are almost all black. The data distribution is concentrated near 0 and 1. This indicates that direct global normalization will make the data too concentrated, resulting in the loss of detailed information, reduced feature discrimination, and being unfavorable for the convergence of the neural network; while in this embodiment, the artificial feature map obtained by performing preprocessing and normalization in sequence, as Figure 2 shown, Figure 2 shows that the data distribution is not overly concentrated near 0 and 1, which is beneficial to the convergence of the neural network and improves the generalization ability of the model.

[0104] Step 7: Construct and divide the total dataset. Use the normalized artificial feature maps generated from each sample in the original interference signal data as the total dataset, and divide the total dataset into a training set, a validation set, and a test set according to a certain ratio.

[0105] In this embodiment, 1000 samples are generated for each of the 6 types of interference signals under a total of 21 signal-to-interference-plus-noise ratio (SINR) conditions in Step 1. Therefore, the total number of samples in the dataset is 1000×6×21 = 126000; divide the 1000 samples of each type of interference signal under each SINR condition into a training set, a validation set, and a test set according to a quantity ratio of 6:2:2. Therefore, the total number of samples in the training set is 75600, and the total number of samples in the validation set and the test set are both 25200.

[0106] Step 8: Construct a convolutional neural network. Design a reasonable convolutional neural network according to the size of the artificial feature map constructed in Step 5.

[0107] The size of the artificial feature map constructed in Step 5 is 9×32. The structure of the convolutional neural network constructed in this embodiment is as Figure 4As shown, the input size of the neural network is the same as that of the artificial feature map; the first layer is a convolutional layer with 6 channels and a convolutional kernel size of 3×3. The activation function used for the output of this layer is leakyrelu; the second layer is a convolutional layer with 16 channels and a convolutional kernel size of 3×3. The activation function used for the output is leakyrelu; the third layer is a convolutional layer with 120 channels and a convolutional kernel size of 3×3. The activation function used for the output is leakyrelu; the fourth layer is a fully connected layer with the number of input and output neurons being 120 and 84 respectively. The activation function used is leakyrelu; the fifth layer is a fully connected layer with the number of input and output neurons being 84 and 6 respectively; the sixth layer is a softmax layer that outputs the recognition probabilities of 6 types of interference signals;

[0108] Step 9: Train the convolutional neural network. Use the training set and validation set in Step 7 to train the constructed convolutional neural network to obtain an interference signal recognition model;

[0109] Train the convolutional neural network constructed in Step 8 in the pycharm software. The hyperparameters during the training process are set as follows: the batch size is 256, the learning rate is 0.0001, the number of training epochs is 600, the loss function is the cross-entropy loss function, and the optimizer used in the backpropagation process is the Stochastic Gradient Descent optimizer (SGD);

[0110] Figure 5 and Figure 6 respectively show the loss value curve and the model accuracy curve during the training process of the convolutional neural network model in this embodiment; it can be seen from the figure that the training curve (train) and the validation curve (val) almost completely overlap, and the curve convergence speed is relatively fast, the change is stable, and the fluctuation is small. Finally, the loss value curve converges to about 0.03, while the model accuracy curve stabilizes at about 0.99; this indicates that the neural network training result is relatively ideal, the model has strong learning ability and generalization ability, and can perform well on both the training set and the validation set;

[0111] Step 10: Test the interference signal recognition model. Use the test set in Step 7 to test the interference signal recognition model in Step 9 to obtain the test results to verify the performance of the constructed interference signal recognition model;

[0112] Adopt a test set independent of the training set and the validation set to test the trained convolutional neural network. The confusion matrix of the test results is as Figure 7As shown, under the condition that the channel is an additive white Gaussian noise channel and the signal-to-interference-plus-noise ratio (SINR) ranges from -20 dB to 20 dB, the interference signal recognition model constructed by the present invention has an average correct recognition rate of up to 99.34% for 6 types of interference signals. Among them, the correct recognition rate of multi-tone interference signals is 97.95%, which is the lowest among all categories; the correct recognition rate of narrowband noise interference signals is 99.95%, which is the highest. Figure 8 This is the interference recognition accuracy curve of the interference signal recognition model constructed in this embodiment under different SINR environments. When the SINR is higher than -14 dB, the recognition accuracy of each type of interference signal is almost 100%. Compared with the convolutional neural network model in the Chinese invention patent with the publication number CN108509911B, the interference signal recognition model constructed in this embodiment not only improves the interference recognition performance by nearly 12 dB, but also has a simpler network structure and lower computational complexity.

[0113] The present invention also provides an interference signal recognition system based on artificial feature maps and convolutional neural networks, including:

[0114] Interference signal preprocessing module: used to perform power normalization processing on each sample in the generated original interference signal data.

[0115] Signal time-domain segmentation module: continuously segment each sample in the preprocessed original interference signal data in the time-domain dimension to obtain time-domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32.

[0116] Artificial feature extraction module: select N adjacent time-domain segments in each sample, perform artificial feature extraction on each time-domain segment respectively, and construct an artificial feature map with a size of 9×N.

[0117] Convolutional neural network construction and training module: used to construct a convolutional neural network and perform training to obtain an interference signal recognition model.

[0118] Testing module: input the artificial feature map into the interference signal recognition model for testing to obtain the test results.

[0119] The present invention also provides an interference signal recognition device based on artificial feature maps and convolutional neural networks, including:

[0120] Memory: stores the computer program of the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks, and is a computer-readable device.

[0121] Processor: used to implement the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks when executing the computer program.

[0122] The present invention also provides a computer-readable storage medium storing a computer program, which when executed by a processor, can implement the above-mentioned interference signal recognition method based on artificial feature maps and convolutional neural networks.

Claims

1. A method for identifying interference signals based on artificial feature maps and convolutional neural networks, characterized in that: The steps include: Step 1: Preprocess the interference signal, and perform power normalization processing on each sample in the generated interference signal raw data; Step 2: Signal time domain segmentation: each sample in the preprocessed original interference signal data is continuously segmented in the time domain dimension to obtain time domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32; Step 3: Manual feature extraction: select N adjacent time domain segments in each sample, perform manual feature extraction on each time domain segment, and construct an artificial feature map of size 9×N; Step 4: Construct a convolutional neural network and train it to obtain an interference signal recognition model; Step 5: Input the artificial feature map into the interference signal recognition model for testing to obtain the test results.

2. The interference signal recognition method based on artificial feature graph and convolutional neural network according to claim 1 is characterized in that: The feature categories in step 3 include frequency domain moment skewness, frequency domain moment kurtosis, spectrum kurtosis, frequency domain envelope fluctuation, average spectrum flatness coefficient, time domain moment skewness, time domain moment kurtosis, time domain peak-to-average ratio and time domain envelope fluctuation.

3. The interference signal recognition method based on artificial feature graph and convolutional neural network according to claim 1 is characterized in that: In step 3, the feature data obtained from each sample is used to construct a feature map, that is, the dimension of the time domain segment is used as the length of the artificial feature map, and the feature category dimension is used as the width of the artificial feature map, so as to construct an artificial feature map with a length of N and a width of 9.

4. The interference signal recognition method based on the combination of artificial feature graph and convolutional neural network according to claim 1 is characterized in that: In step 3, the feature data of the artificial feature map are preprocessed and normalized in sequence.

5. The interference signal recognition method based on artificial feature graph and convolutional neural network according to claim 4 is characterized in that: First, the frequency domain feature data in the artificial feature map are individually taken to perform data preprocessing by taking log2, log10 or square root, while keeping the time domain feature data unchanged; Then, the maximum and minimum values ​​of the time domain and frequency domain eigenvalues ​​are normalized respectively, and each eigenvalue in the entire artificial feature map is limited to between 0 and 1.

6. The interference signal recognition method based on artificial feature graph and convolutional neural network according to claim 1 is characterized in that: The input size of the convolutional neural network constructed in step 4 is 9×N, the first layer is a convolutional layer with a convolution kernel size of 3×3, the number of channels is 6, and the activation function used for the output is leakyrelu; the second layer is a convolutional layer with a convolution kernel size of 3×3, the number of channels is 16, and the activation function used for the output is leakyrelu; the third layer is a convolutional layer with a convolution kernel size of 3×3, the number of channels is 120, and the activation function used for the output is leakyrelu; the fourth layer is a fully connected layer, the number of input and output neurons is 120 and 84 respectively, and the activation function used is leakyrelu; the fifth layer is a fully connected layer, the number of input and output neurons is 84 and M respectively; the sixth layer is a softmax layer, and the recognition probability of M types of interference signals is output.

7. An interference signal recognition system based on artificial feature graph and convolutional neural network based on the method of claims 1-6, characterized in that: include: Interference signal preprocessing module: used to perform power normalization processing on each sample in the generated interference signal raw data; Signal time domain segmentation module: Continuously segment each sample in the preprocessed interference signal raw data in the time domain dimension to obtain time domain segments of equal length, and the number of segments is not less than N segments, where N is a positive integer not less than 32; Artificial feature extraction module: select N adjacent time domain segments in each sample, extract artificial features from each time domain segment, and construct an artificial feature map of size 9×N; Convolutional neural network construction and training module: used to construct a convolutional neural network and perform training to obtain an interference signal recognition model; Test module: Input the artificial feature map into the interference signal recognition model for testing and obtain the test results.

8. An interference signal recognition device based on artificial feature graph and convolutional neural network, characterized in that: include: Memory: a computer program storing the interference signal recognition method based on artificial feature graph and convolutional neural network as described in any one of claims 1 to 6, which is a computer-readable device; Processor: used to implement the interference signal recognition method based on artificial feature map and convolutional neural network as described in any one of claims 1-6 when executing the computer program.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the interference signal recognition method based on artificial feature map and convolutional neural network described in any one of claims 1-6.

Images