Interference identification method, system and device based on deep learning, and medium

Abstract

The invention relates to the technical field of electronic communication, in particular to an interference identification method, system and device based on deep learning and a medium. According to the interference identification method, an input signal is subjected to CV-R-BiLSTM-A network structure processing and classification processing in sequence to obtain an interference identification result, and the CV-R-BiLSTM-A network structure processing comprises the step of processing the input signal through a complex convolutional network (CV), a residual network (R), a bidirectional long-short-term memory network (BiLSTM) and an attention layer (A) in sequence to obtain weighted feature data. The classification processing comprises the step of processing the weighted feature data through a full connection layer and a Softmax layer to obtain an interference identification result; the interference identification system applies the interference identification method and specifically comprises a deep learning module and a classification processing module. According to the method, complex interference signals can be processed, high recognition accuracy can be kept in a low-interference-to-noise-ratio environment, the calculation complexity is reduced, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic communication technology, and in particular to an interference identification method, system, device and medium based on deep learning. Background Art

[0002] With the advancement of communication technology, data link systems may face various interferences. In communication countermeasures, it is necessary to ensure the stability of the data link while also considering how to effectively deal with possible communication interference. Interference can be divided into natural and man-made interference. In recent years, non-spread spectrum interference technology has received increasing attention, especially with the development of artificial intelligence, and intelligent interference technology is gaining increasing attention. Intelligent interference technology includes interference identification, interference strategy formulation, and interference suppression. Deep learning has been successfully applied in fields such as image processing, so it is being applied to communication interference, leveraging its powerful feature extraction capabilities to solve the problem.

[0003] Interference identification methods in the prior art generally have the following problems:

[0004] Inadequate processing capabilities for low-INR signals: In low-INR environments, the intensity of interference and noise in the signal is significantly higher than that of the target signal, significantly obscuring the characteristics of the target signal. Traditional interference identification methods rely on the statistical or spectral characteristics of the signal, but in low-INR conditions, these characteristics can become blurred or even indistinguishable, significantly reducing identification accuracy.

[0005] Low recognition accuracy: The existing technology has a significantly reduced recognition accuracy in low interference-to-noise ratio environments, especially in processing complex interference signals such as noise frequency modulation and noise amplitude modulation.

[0006] The invention, patent publication number CN115296758A, entitled "A Method, System, Computer Device, and Storage Medium for Identifying Interference Signals," discloses a method for identifying interference signals based on a convolutional recurrent joint network. By combining a long short-term memory network (LSTM) with a deep residual network (ResNet), it achieves multi-dimensional feature extraction and classification of interference signals. This invention has a high recognition accuracy, but it has the following problems: Complex network structure: This invention uses a joint network of LSTM and ResNet. Although this improves recognition accuracy, the network structure is complex, and the training and inference time are long, making it difficult to meet real-time requirements; Limited processing capabilities for low interference-to-noise ratio signals: In low interference-to-noise ratio environments, the invention's recognition accuracy drops significantly, especially when processing complex interference signals such as noise frequency modulation and noise amplitude modulation; High computing resource consumption: Due to the use of a deep residual network and LSTM, this invention has a high demand for computing resources, especially when processing large-scale data, resulting in high computational costs. Summary of the Invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to propose an interference identification method, system, device and medium based on deep learning. The method uses a CV-R-BiLSTM-A (complex residual bidirectional long short-term memory attention) network structure to achieve efficient identification and classification of interference signals in received signals; the present invention achieves efficient and accurate interference signal identification, and solves the problem of low recognition accuracy and insufficient processing capability of low interference-to-noise ratio signals in the prior art.

[0008] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0009] In a first aspect, a method for interference identification based on deep learning comprises the following steps:

[0010] The input signal is sequentially processed by a CV-R-BiLSTM-A network structure and classified to obtain an interference recognition result. The CV-R-BiLSTM-A network structure processing includes sequentially processing the input signal by a complex convolutional network (CV), a residual network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A) to obtain weighted feature data. The classification processing includes processing the weighted feature data by a fully connected layer and a Softmax layer to obtain an interference recognition result.

[0011] Furthermore, the complex convolutional network (CV) includes one convolution layer and one maximum pooling layer. The convolution layer uses a 7*1 convolution kernel with a step size of 2; the maximum pooling layer uses a 2*1 convolution kernel with a step size of 2. The maximum pooling layer formula is as follows:

[0012]

[0013] Where x i is the input feature map, y i，j is the output feature map; k h and k w is the height and width of the pooling kernel, which are 2 and 1 respectively; S h and S w The step size of the pooling operation is 2, m is the height direction index (row direction) range: 0 to 1 (k h -1), n ​​is the index in the width direction (column direction) range: 0 to (k w -1).

[0014] Furthermore, the residual network (R) includes 4 residual convolution layers and 1 average pooling layer. The convolution kernel sizes of the 4 residual convolution layers are 5*1, 3*1, 3*1, and 3*1, respectively. The stride of the 4 residual convolution layers is 1. The output formula of the residual convolution layer is as follows:

[0015] y=F(x,{w i})+x

[0016] Where x is the input data, F(x,{w i}) is the output of the convolutional layer, w i is the weight parameter of the convolution layer, and y is the output of the residual convolution layer.

[0017] Furthermore, the bidirectional long short-term memory network (BiLSTM) consists of two layers of LSTM, the number of neurons in each layer of LSTM is 256, and the activation function of each layer of LSTM is the sigmoid formula:

[0018]

[0019] Where x t is the tth element of the input sequence, is the hidden state of the forward LSTM, is the hidden state of the backward LSTM, h t is the output of BiLSTM.

[0020] Furthermore, the attention layer (A) includes an attention layer, and the attention layer formula is as follows:

[0021]

[0022] Where X∈R T×d , the input feature matrix includes the features of T time steps, each feature dimension is d, For scaling, xx T It is to calculate the similarity matrix between features.

[0023] Furthermore, the interference identification method further includes interference signal preprocessing before CV-R-BiLSTM-A network structure processing, and the interference signal preprocessing includes a power normalization operation, and the normalization is as follows:

[0024]

[0025] Where x(n) is the input signal, N is the total sampling points, and n is the sampling point.

[0026] Furthermore, each convolutional layer in the complex convolutional network (CV), residual network (R), and attention layer (A) uses the CReLU (Concatenated Rectified Linear Unit) activation function, and a BN layer is added before each convolutional layer uses the CReLU activation function. The CReLU activation function is as follows:

[0027]

[0028] Where x is the input data, ReLU(x)=max(0,x) is the ReLU activation function, and ReLU(-x)=max(0,-x) is the negative part of ReLU. Represents a vector concatenation operation.

[0029] In a second aspect, a deep learning-based interference identification system is provided, which applies the interference identification method. The interference identification system includes a deep learning module and a classification processing module:

[0030] Deep learning module: The input signal is processed through the CV-R-BiLSTM-A network structure to obtain weighted feature data. The CV-R-BiLSTM-A network structure includes a complex convolutional network (CV), a residual network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A);

[0031] Classification processing module: The weighted feature data is processed through a fully connected layer and a Softmax layer to obtain an interference recognition result.

[0032] In a third aspect, an electronic device includes a memory and a processor:

[0033] Memory: used for storing a computer program for implementing the interference identification method;

[0034] Processor: configured to implement the interference identification method when executing the computer program.

[0035] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the interference identification method is implemented.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] (1) The present invention combines a complex convolutional network (CV) with a residual network (R), providing efficient feature extraction and improving the accuracy of interference signal recognition. Existing interference recognition methods typically use a single convolutional network or residual network for feature extraction, which cannot simultaneously take into account frequency and time domain features. The present invention extracts the frequency and time domain features of the signal through a complex convolutional network, and combines it with a residual network to solve the gradient vanishing problem in deep networks, thereby improving the network's training efficiency and generalization ability.

[0038] (2) The present invention incorporates a bidirectional long short-term memory (BiLSTM) network, which is capable of capturing temporal dependencies and improving the ability to identify complex interference signals. Existing interference identification methods typically use unidirectional LSTMs or simple convolutional networks to process temporal data, which cannot fully capture the signal's forward and backward correlation information. The present invention, through the BiLSTM module, can simultaneously consider both the forward and backward information of the signal, further extracting deep feature information, improving the comprehensiveness of feature extraction and significantly enhancing the ability to identify complex interference signals.

[0039] (3) The present invention includes an attention layer (A) module that automatically focuses on key features, thereby improving the accuracy of interference signal recognition in low interference-to-noise ratio environments. Existing interference recognition methods generally lack an automatic focus mechanism on key features, resulting in a significant decrease in recognition accuracy in low interference-to-noise ratio environments. The present invention, through the attention layer module, can automatically focus on key information in the input features and perform a weighted summation of this information to extract a more effective feature representation, significantly improving the accuracy of interference signal recognition in low interference-to-noise ratio environments.

[0040] In summary, the present invention constructs an efficient interference recognition network by organically combining a complex convolutional network, a residual network, a BiLSTM and an attention mechanism. The network can not only process complex interference signals, but also maintain a high recognition accuracy in a low interference-to-noise ratio environment: improve the interference signal recognition accuracy, through multi-dimensional feature extraction and automatic attention to key information, significantly improve the interference signal recognition accuracy, especially in a low interference-to-noise ratio environment; reduce computational complexity, through the combination of a complex convolutional network and a residual network, reduce the complexity of manual feature extraction, and improve computational efficiency; enhance the generalization ability of the model, through the BiLSTM module and the attention layer module, enhance the model's generalization ability for complex interference signals, and reduce overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a structural diagram of the interference identification network based on CV-R-BiLSTM-A of the present invention.

[0042] Figure 2 This is a flow chart of interference identification based on CV-R-BiLSTM-A of the present invention.

[0043] Figure 3 It is a workflow diagram of the present invention in two application modes.

[0044] Figure 4 η is the confusion matrix of the CV-R-BiLSTM-A network recognition when the interference-to-noise ratio is -5 dB in this embodiment.

[0045] Figure 5∈ R₁ is the confusion matrix of the CV-R-BiLSTM-A network under the interference-to-noise ratio of 10 dB in this embodiment.

[0046] Figure 6 This is the interference identification performance based on the CV-R-BiLSTM-A network in this embodiment.

[0047] Figure 7 This is a performance comparison between this embodiment and other different network models. DETAILED DESCRIPTION

[0048] The following is combined with Figure 1 To the attached Figure 7 The present invention is described in further detail:

[0049] First, an interference identification method based on deep learning is proposed, and the main steps are: the external input signal is first processed by the CV-R-BiLSTM-A network structure to obtain weighted feature data; the classification processing processes the weighted feature data through the fully connected layer and the Softmax layer to obtain the interference identification result.

[0050] like Figure 1 As shown, the CV-R-BiLSTM-A network structure includes a complex convolutional network (CV), a residual convolutional network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A); wherein the complex convolutional network (CV) is used to perform preliminary feature extraction on the input signal and capture the frequency domain and time domain features of the signal. Through the complex convolutional network, the frequency domain and time domain features of the signal can be efficiently extracted, laying the foundation for subsequent feature extraction and classification; the residual convolutional network (R) can extract deeper feature information through the residual convolutional network, alleviate the gradient disappearance problem in the deep network, and improve the training efficiency and generalization ability of the model; the bidirectional long short-term memory network (BiLSTM) is used to capture the temporal dependency in the input data and further extract deep feature information. Through the BiLSTM module, the temporal dependency in the input data can be captured and further extract deep feature information, significantly improving the recognition ability of complex interference signals; the attention layer (A) is used to calculate the attention distribution in the feature information and perform weighted summation on the input feature information to extract a more effective feature representation. Through the attention layer module, it can automatically focus on the key information in the input features, significantly improving the accuracy of interference signal recognition in low interference-to-noise ratio environments.

[0051] The fully connected layer and the Softmax classifier are used to map the extracted features to the classification space and perform final classification. Through the fully connected layer and the Softmax classifier, the extracted features can be mapped to the classification space and accurate classification of the interference signal can be achieved.

[0052] 1. Interference signal preprocessing:

[0053] First, the input signal is preprocessed. The strength of the interference signal may be different. In order to avoid the influence of different amplitudes and powers on the recognition performance, the input signal x(n) is power normalized:

[0054]

[0055] Where N is the total number of sampling points and n is the sampling point:

[0056] When using time domain truncation to acquire the original signal, it will lead to spectrum loss. In order to reduce the spectrum leakage caused by time domain truncation, the signal is processed by Hanning window. The time domain expression of Hanning window is:

[0057]

[0058] The windowed signal is x windowed (n) = x(n)·w(n), the Hanning window effectively suppresses the spectrum sidelobes by smoothing the edges of the signal, thereby improving the accuracy of frequency domain analysis.

[0059] The input signal sequence contains interference signals and Gaussian noise. Its mathematical model

[0060] y(t)=j(t)+η(t) (3)

[0061] Where j(t) is the interference signal, η(t) is Gaussian noise, and y(t) is the input signal. The input signal data format is the I and Q paths of the interference signal. The specific content of the one-hot encoding of the network label selection signal category is shown in (Table 1)

[0062] 2. Preliminary feature extraction (complex convolutional network (CV)):

[0063] The preprocessed input signal data is fed into the first convolutional layer for preliminary feature extraction. This convolutional layer uses a 7x1 convolution kernel with a stride of 2. Unless otherwise specified, each convolutional layer uses the CReLU (Concatenated Rectified Linear Unit) activation function. A batch normalization layer is added before each convolutional layer to improve training speed and mitigate gradient loss. This allows for efficient extraction of key features from the input data. The kernel size and stride are designed to minimize computational complexity while ensuring effective feature extraction.

[0064] The CReLU (Concatenated Rectified Linear Unit) activation function can enhance the ability to express features. By retaining both the positive and negative information of the input, CReLU can capture more feature information, especially when processing complex signals, and can better extract the frequency domain and time domain features of the signal. Alleviate the gradient vanishing problem. Compared with the traditional ReLU, CReLU can alleviate the gradient vanishing problem to a certain extent by splicing positive and negative information, especially in deep networks. Improve the generalization ability of the model. CReLU can extract more feature information, thereby improving the generalization ability of the model and reducing overfitting. The formula of the CReLU activation function can be expressed as:

[0065]

[0066] Where x is the input data, ReLU(x)=max(0,x) is the ReLU activation function, and ReLU(-x)=max(0,-x) is the negative part of ReLU. Represents a vector concatenation operation.

[0067] The function of the BN layer is to normalize the output of each convolutional layer to accelerate the network training process and alleviate the gradient disappearance problem. The specific formula is as follows:

[0068]

[0069] Where x i is the input data, μ B is the mean of the current batch of data, is the variance of the current batch of data, and ζ is a small constant used to prevent the denominator from being zero.

[0070] The output of the BN layer is:

[0071]

[0072] Here, γ and β are learnable parameters representing scaling and translation, respectively. Initialized to 1 and 0, they are optimized via gradient descent, with no strict range constraints. γ typically tends to be positive and is used to adjust the variance of the feature distribution, typically stabilizing in the range [0.1, 10]. β is typically distributed near the mean of the input data.

[0073] 3. Max pooling layer:

[0074] The feature data extracted by the convolution layer is then passed through a maximum pooling layer with a stride of 2 and a convolution kernel size of 2*1. The formula is as follows:

[0075]

[0076] Where x i is the input feature map, y i，j is the output feature map; k h and k w is the height and width of the pooling kernel, which are 2 and 1 respectively; S h and S w The step size of the pooling operation is 2, m is the height direction index (row direction) range: 0 to 1 (k h -1), n ​​is the index in the width direction (column direction) range: 0 is (k w -1).

[0077] The role of the maximum pooling layer is to further simplify data features, reduce computational complexity, and retain the most important feature information.

[0078] 4. Residual Convolutional Layer (R):

[0079] The data then passes through four residual convolutional layers (2 to 5). Residual connections are built between these convolutional layers, allowing information to be passed directly from front to back between layers. The parameters of each convolutional layer are set differently to adapt to the feature extraction requirements of different levels. The specific parameters are as follows:

[0080] Convolutional layer 2: The convolution kernel size is 5*1 and the stride is 1.

[0081] Convolutional layer 3: The convolution kernel size is 3*1 and the stride is 1.

[0082] Convolutional layer 4: The convolution kernel size is 3*1 and the stride is 1.

[0083] Convolutional layer 5: The convolution kernel size is 3*1 and the stride is 1.

[0084] The formula is as follows:

[0085] y=F(x,{w i})+x (8)

[0086] Where x is the input data, F(x,{w i}) is the output of the convolutional layer, w i is the weight parameter of the convolutional layer, and y is the output of the residual convolutional layer. To improve the network training speed and effectively alleviate the gradient loss problem, the present invention adds a Batch Normalization (BN) layer before using the activation function (CReLU) in each convolutional layer to improve the network training speed and effectively alleviate the gradient loss problem.

[0087] The Average Pooling Layer is a commonly used downsampling operation used to reduce the spatial size (height and width) of a feature map while retaining important information. The core idea is to calculate the average value of a local area of ​​the input feature map and use this value as the value of the corresponding position in the output feature map. The formula is as follows:

[0088]

[0089] Where y i,j is the value of the i-th row and j-th column of the output feature map, x(i·s+m),(j·s+n) are the elements in the corresponding window in the input feature map, k×k is the pooling window size, and s is the step size of each movement of the pooling window.

[0090] 5. Bidirectional Long Short-Term Memory Network (BiLSTM):

[0091] The data processed by the convolutional layer is then fed into a Bidirectional Long Short-Term Memory (BiLSTM) network. The BiLSTM consists of two LSTM layers, each with 256 neurons and a sigmoid activation function. The formula is as follows:

[0092]

[0093] Where x t is the tth element of the input sequence, is the hidden state of the forward LSTM, is the hidden state of the backward LSTM, h t is the output of BiLSTM.

[0094] BiLSTM can capture temporal dependencies in input data and further extract deep feature information. Compared with traditional LSTM, BiLSTM can simultaneously consider both forward and backward information of the signal, improving the comprehensiveness of feature extraction.

[0095] 6. Attention layer (A):

[0096] After BiLSTM, the data enters the attention layer. The attention layer formula is as follows:

[0097]

[0098] X∈R T×d , the input feature matrix (that is, the feature matrix obtained after the output of BiLSTM is processed by the activation function) includes the features of T time steps, and each feature dimension is d. T It is to calculate the similarity matrix between features. Scaling factor to prevent the dot product from being too large and causing the gradient to vanish. Softmax is normalized by row to generate the attention weight matrix. The output feature matrix shape remains T×d.

[0099] The attention layer calculates the attention distribution in the feature information and performs weighted summation on the input feature information to extract a more effective feature representation. The introduction of the attention layer enables the network to automatically focus on the key information in the input features, thereby improving the network's ability to recognize complex interference signals.

[0100] 7. Fully connected layer and Softmax classification:

[0101] The formula of the fully connected layer is: z = W x + b, y = f (z) (12)

[0102] W·x represents the matrix multiplication of the weight matrix and the input vector. b is the offset term, and an activation function f(x) is usually applied to the output, resulting in y = f(z).

[0103] The formula for the Softmax layer is:

[0104]

[0105] Among them, the molecule e xi For input x i The exponential operation of is a normalization term that ensures that the sum of all probabilities is 1.

[0106] Finally, the processed data is fed into the fully connected layer (FC) and the Softmax layer for final classification to obtain the interference identification results. The fully connected layer is responsible for mapping the extracted features to the classification space, while the Softmax layer is used to calculate the probability distribution of each category, thereby achieving accurate classification of the interference signal.

[0107] like Figure 2 and Figure 3 As shown, this method can be applied in two modes: training mode and working mode:

[0108] Training mode: Training data is performed by simulating the transmission of the original signal. After receiving the original signal and the interference signal from the channel, interference identification is performed on the received input signal (including the original signal and the interference signal), and the interference is classified, identified, and eliminated. After repeated operations, the final signal without the interference signal is obtained. Finally, the final signal is compared with the original signal for training. After multiple training cycles, the final signal is compared with the original signal, and the optimal number of training cycles is used as the termination condition (set to 50 times at the beginning).

[0109] Working mode: Receive the original signal and interference signal from the channel, classify, identify and eliminate the interference, obtain the final signal after multiple identification and elimination, record the data loss rate and bit error rate, and finally output the signal after removing the interference.

[0110] In a second aspect, a deep learning-based interference identification system is provided, which applies the interference identification method. The interference identification system includes a deep learning module and a classification processing module:

[0111] Deep learning module: The input signal is processed through the CV-R-BiLSTM-A network structure to obtain weighted feature data. The CV-R-BiLSTM-A network structure includes a complex convolutional network (CV), a residual network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A);

[0112] Classification processing module: The weighted feature data is processed through a fully connected layer and a Softmax layer to obtain an interference recognition result.

[0113] In a third aspect, an electronic device includes a memory and a processor:

[0114] Memory: used for storing a computer program for implementing the interference identification method;

[0115] Processor: configured to implement the interference identification method when executing the computer program.

[0116] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the interference identification method is implemented; the computer-readable storage medium includes: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store program codes.

[0117] The present method is further described in detail below by Example 1:

[0118] (1) Parameter settings

[0119] The simulation hardware environment of Example 1 uses a GPU (NVIDIA RTX 3060), and the software uses Matlab and the deep learning framework TensorFlow. The network initialization method uses MSRA initialization. By adjusting the initial distribution of weights, MSRA initialization can effectively solve the gradient vanishing problem and accelerate the convergence speed of the network. Its formula is as follows:

[0120]

[0121] W is the weight matrix, N represents the normal distribution, n in is the number of input neurons.

[0122] The loss function is cross entropy, the batch size (the number of samples used in each training iteration) is 16, and the learning rate is 0.001, which controls the step size of the model parameter updates. The Adam optimizer is used, which combines the advantages of the momentum method and adaptive learning rate to accelerate model convergence. β1 = 0.9 (first-order momentum decay rate), β2 = 0.999 (second-order momentum decay rate), and ∈ = 1e-8 (a small constant to prevent division by zero).

[0123] Example 1 Classifies and identifies eight types of interference signals through a CV-R-BiLSTM-A interference identification network.

[0124] Eight interference signals were generated using Matlab: single-tone interference, multi-tone interference, linear frequency sweep interference, noise FM interference, noise AM interference, partial noise interference, chirp FM interference, and sawtooth FM interference. The sampling rate was set to 10 MHz, and Gaussian white noise was added to the channel. The basic information for the eight interference types is as follows:

[0125] 1) Single tone interference

[0126] The expression of single-tone interference is:

[0127]

[0128] Where A is the amplitude, f c is the interference frequency, is the initial phase.

[0129] 2) Multi-tone interference

[0130] Multi-tone interference is a complex baseband expression of the superposition of multiple single-tone interferences:

[0131]

[0132] Among them A m is the amplitude of the Mth single-tone interference of multi-tone interference, and similarly f m ; are the frequency and phase of the Mth single-tone interference respectively.

[0133] 3) Partial frequency band noise interference

[0134] Partial band interference concentrates the noise energy within a specified frequency band, and its expression is:

[0135]

[0136] Where U(t) has a mean of 0 and a variance of Gaussian noise, f c is the center frequency of the interference signal, is the initial phase.

[0137] 4) Linear frequency sweep interference

[0138] The frequency and time of linear swept frequency interference are linearly related, and its time domain expression is:

[0139]

[0140] Where A is the amplitude, f0 is the initial frequency, is the initial phase, k is the frequency modulation coefficient, and T is the signal duration.

[0141] 5) Chirp FM interference

[0142] The single-component Chirp signal expression is:

[0143]

[0144] Where A is the amplitude, f0 is the initial frequency, k0 is the frequency modulation slope of the signal, and T is the signal duration.

[0145] 6) Noise FM interference

[0146] This interference has a linear relationship between frequency and modulation noise voltage. Its expression is:

[0147]

[0148] Where A is the amplitude, f c is the center frequency of the interference signal, k fm is the frequency modulation coefficient, ξ(t) has a mean of 0 and a variance of The bandwidth of the interference signal is determined by the frequency modulation coefficient and variance.

[0149] 7) Noise AM interference

[0150] The noise amplitude modulation interference expression is:

[0151]

[0152] Where u0 is the carrier amplitude, w j is the carrier angular frequency, U n (t) is the baseband noise, Uniformly distributed in [0,2π].

[0153] 8) Sawtooth FM interference

[0154] The sawtooth FM signal uses a sawtooth wave signal for frequency modulation, which is similar to the noise FM signal. Its expression is:

[0155] J(t)=Acos(2πf c t+2πK∫m(t)dt) (22)

[0156] Where A represents the amplitude, f c is the carrier frequency of the signal, K is the modulation coefficient, and m(t) is the periodic sawtooth wave signal.

[0157] The following formula gives the frequency component f in each frequency band of the sawtooth FM signal: c ±nF power.

[0158]

[0159] Where F is the periodic frequency of the sawtooth wave, 2Δf i is the FM deviation of the signal, P o is the total interference power, C and S are the Fresnel integrals.

[0160] When the FM bandwidth is large, that is, 2Δf i >>F, F at this time can be ignored and can be simplified to the following formula:

[0161]

[0162] From the above formula, it can be concluded that the sawtooth frequency modulation signal has F / 2Δf i components, and the interference power of each component is basically the same, and is evenly distributed in the available frequency band in a comb shape.

[0163] The interference-to-noise ratio (INR) is defined as the ratio of the interference signal strength to the noise intensity within the communication band. The INR range was set to [-10:1:15] dB. 2500 samples were randomly generated for each signal type at each INR. The sample sizes in the training, test, and validation sets were set in a ratio of 6:2:2. All subsequent experimental parameters remained consistent.

[0164] (2) Network model validity analysis

[0165] Evaluation indicators include:

[0166] Accuracy: The ratio of the total number of correctly detected positive and negative samples to the total number of samples

[0167]

[0168] True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN). In this design, we study the recognition of eight types of interference signals and use accuracy to evaluate the network model's ability to identify interference signals.

[0169] Precision: The probability of an actual positive sample being predicted as a correct sample.

[0170]

[0171] Recall: The probability that the prediction is positive among the actual positive samples.

[0172]

[0173] To verify the importance of each module in the designed network, three group ablation experiments were conducted. Under identical training and testing conditions, the overall recognition rates of different networks for various interference signals within the test set, ranging from -10dB to 15dB, are shown in Table 1. Table 1 shows that the CV-R-BiLSTM-A network achieved the highest accuracy for distinguishing the eight interference signals. Using only a complex residual network performed the worst, achieving the lowest recognition rates for all eight interference signals. Adding a BiLSTM and an attention layer to the network yielded comparable recognition performance, with varying degrees of accuracy for each type of interference. However, compared to the complex residual network, both networks achieved improved accuracy for each type of interference, demonstrating the effectiveness of the designed BiLSTM and attention layer modules in improving network recognition. Overall, the CV-R-BiLSTM-A network improved overall recognition rates by approximately 10% compared to the CV-ResNet network.

[0174] Table 1 One-hot encoding used in eight interference types in this embodiment

[0175]

[0176]

[0177] (3) Analysis and comparison of simulation effects

[0178] After the CV-R-BiLSTM-A interference recognition network is trained, it is tested on the test set. Figures 4 and 5As shown in the figure, the horizontal axis represents the predicted value labels, and the vertical axis represents the true value labels. The two confusion matrices show that at low interference-to-noise ratios, the CV-R-BiLSTM-A network performs poorly in identifying three interference signals: noise frequency modulation, noise amplitude modulation, and partial noise. The recognition rate is approximately 70%. Compared to other interference signals, the recognition accuracy is as low as approximately 84%. Because noise characteristics are less pronounced at low interference-to-noise ratios, the confusion matrices show that misclassifications of the three interference signals are evenly distributed, with no concentration of misclassifications on any one type. However, when comparing the interference ratio at 10 dB, we find that as the interference increases, the network's recognition accuracy improves, with recognition rates exceeding 95% for noise frequency modulation, noise amplitude modulation, and partial noise.

[0179] like Figure 6 The figure shows the interference recognition performance curve of the CV-R-BiLSTM-A network for various interference signals at different interference-to-noise ratios. For each interference signal in the figure, the recognition accuracy increases with the increase of the interference-to-noise ratio. When the interference-to-noise ratio is greater than 8dB, the network recognition accuracy tends to stabilize. Figure 6 It can be seen intuitively that the network's recognition accuracy for noise frequency modulation, noise amplitude modulation and some noise is relatively low when the interference-to-noise ratio is less than 0dB; the network has the best recognition effect for linear sweep frequency interference, with an accuracy of more than 90% at -5dB.

[0180] (4) CV-R-BiLSTM-A network structure compared with other networks

[0181] In addition, the network structure of the present invention is compared with the existing network in terms of network performance. The network model selected has been tested for the validity of these networks through a large number of experiments, and six networks (LeNet-5, VGG16, ResNet34, InceptionV4, AlexNet, DenseNet) are built under the same simulation parameter design for testing. Table 2 shows the accuracy of several networks in identifying interference signals. As shown in Table 3, several existing networks are only better for the recognition effect of several individual interference types, such as better for the recognition effect of single-tone and multi-tone interference AlexNet network, better for the recognition effect of linear frequency sweep interference ResNet34 and DenseNet network, better for the recognition effect of sawtooth frequency modulation interference VGG 16 network, but not all outstanding for the recognition effect of noise frequency modulation, noise amplitude modulation, and partial noise. Overall, the CV-R-BiLSTM-A network is the best for the recognition effect of eight types of interference signals. Figure 7 The curves showing the overall recognition accuracy of different networks as the interference-noise ratio changes are shown. Figure 7As can be seen, the CV-R-BiLSTM-A recognition network has a significant advantage over other networks in recognition performance. When the interference-to-noise ratio is -5dB, the network designed in this paper still has a recognition accuracy of approximately 50%, while other networks have almost no recognition capability at this level. When the interference-to-noise ratio exceeds 5dB, the CV-R-BiLSTM-A network's recognition capability becomes even more advantageous, reaching a recognition rate of over 95%. At this point, the recognition rates of other networks range from 80% to 90%. The above figures and tables demonstrate that the designed CV-R-BiLSTM-A network model is more suitable than other networks for identifying and classifying eight types of interference signals, demonstrating the effectiveness of this network design.

[0182] Table 2 Ablation experiment based on CV-R-BiLSTM-A in this embodiment

[0183]

[0184]

[0185] In Table 2, the Chinese names corresponding to different deep learning networks are as follows:

[0186] CV-ResNet: Convolutional Residual Network, CV-R-BiLSTM: Convolutional Residual-Bidirectional Long Short-Term Memory Network, CV-RA: Convolutional Residual-Attention Network, CV-R-BiLSTM-A: Convolutional Residual-Bidirectional Long Short-Term Memory-Attention Network.

[0187] Table 3 Comparison of interference identification accuracy in different networks in this embodiment

[0188]

[0189] In Table 3, the Chinese names corresponding to different deep learning networks are as follows:

[0190] LeNet-5: Proposed by Yann LeCun et al., it is an early classic model of convolutional neural network (CNN); VGG16: Proposed by the Visual Geometry Group (VGG) of the University of Oxford, it is a classic deep convolutional neural network known for its simple and regular structure;

[0191] ResNet34: Proposed by Microsoft Research, it introduced the concept of "residual connection", which solved the gradient vanishing problem in deep networks and made it possible to train ultra-deep networks.

[0192] InceptionV4: One of the Inception series models proposed by Google, it further optimizes the ability to extract multi-scale features and incorporates the concept of residual networks.

[0193] AlexNet: Proposed by Alex Krizhevsky and others, it is considered an important milestone in the revival of deep learning;

[0194] DenseNet (Dense Connection Network): Proposed by Gao Huang et al., it uses "dense connection" to connect each layer directly to all other layers, thereby improving feature propagation efficiency and reducing the number of parameters.

[0195] The working principle of the present invention is:

[0196] In the present invention, the interference identification method based on deep learning realizes the effective identification and classification of interference signals through the following steps: signal preprocessing, power normalization and Hanning window processing of the input interference signal to reduce spectrum leakage and ensure the uniformity of the signal amplitude; complex convolutional network (CV) feature extraction, using a 7*1 convolution kernel with a step size of 2 to perform preliminary feature extraction on the preprocessed signal to capture the frequency domain and time domain characteristics of the signal; residual convolutional network (R) feature extraction, through 4 residual convolution layers (2 to 5), the convolution kernel sizes are 5*1, 3*1, 3*1, 3*1, and the step size is 1, to extract deeper feature information, and the residual connection allows information to be directly transferred between different layers to alleviate the gradient disappearance problem, among which the maximum pooling The layer uses a 2*1 pooling kernel with a step size of 2 to further simplify data features and reduce computational complexity; the bidirectional long short-term memory network (BiLSTM) temporal feature extraction consists of two layers of LSTM, with 256 neurons in each layer and a sigmoid activation function to capture the temporal dependencies in the input data and further extract deep feature information; the attention layer mechanism (A) extracts key features, calculates the attention distribution in the feature information, and performs weighted summation on the input feature information to extract a more effective feature representation; the fully connected layer and the Softmax classifier, the fully connected layer maps the extracted features to the classification space, and the Softmax classifier calculates the probability distribution of each category to achieve accurate classification of interference signals.

Claims

1. A method for interference identification based on deep learning, characterized in that: The following steps are involved: The input signal is sequentially processed by a CV-R-BiLSTM-A network structure and classified to obtain an interference recognition result. The CV-R-BiLSTM-A network structure processing includes sequentially processing the input signal by a complex convolutional network (CV), a residual network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A) to obtain weighted feature data. The classification processing includes processing the weighted feature data by a fully connected layer and a Softmax layer to obtain an interference recognition result.

2. The interference identification method according to claim 1, wherein: The complex convolutional network (CV) includes one convolution layer and one maximum pooling layer. The convolution layer uses a 7*1 convolution kernel with a step size of 2; the maximum pooling layer uses a 2*1 convolution kernel with a step size of 2. The maximum pooling layer formula is as follows: Where x i is the input feature map, y i，j is the output feature map; k h and k w is the height and width of the pooling kernel, which are 2 and 1 respectively; S h and S w The step size of the pooling operation is 2, m is the height direction index (row direction) range: 0 to 1 (k h -1), n ​​is the index in the width direction (column direction) range: 0 to (k w -1).

3. The interference identification method according to claim 1, wherein: The residual network (R) contains 4 residual convolution layers and 1 average pooling layer. The convolution kernel sizes of the 4 residual convolution layers are 5*1, 3*1, 3*1, and 3*1 respectively. The stride of the 4 residual convolution layers is 1. The output formula of the residual convolution layer is as follows: y=F(x,{w i })+x Where x is the input data, F(x,{w i }) is the output of the convolutional layer, w i is the weight parameter of the convolution layer, and y is the output of the residual convolution layer.

4. The interference identification method according to claim 1, wherein: The bidirectional long short-term memory network (BiLSTM) consists of two layers of LSTM. The number of neurons in each LSTM layer is 256, and the activation function of each LSTM layer is the sigmoid formula: Where x t is the tth element of the input sequence, is the hidden state of the forward LSTM, is the hidden state of the backward LSTM, h t is the output of BiLSTM.

5. The interference identification method according to claim 1, wherein: The attention layer (A) includes the attention layer, and the attention layer formula is as follows: Where X∈R T×d , the input feature matrix includes the features of T time steps, each feature dimension is d, For scaling, xx T It is to calculate the similarity matrix between features.

6. The interference identification method according to claim 1, wherein: The interference identification method further includes interference signal preprocessing before CV-R-BiLSTM-A network structure processing, and the interference signal preprocessing includes a power normalization operation, and the normalization is as follows: Where x(n) is the input signal, N is the total sampling points, and n is the sampling point.

7. The interference identification method according to claim 1, wherein: Each convolutional layer in the complex convolutional network (CV), residual network (R), and attention layer (A) uses the CReLU (Concatenated Rectified Linear Unit) activation function, and a batch normalization layer is added before each convolutional layer uses the CReLU activation function. The CReLU activation function is as follows: Where x is the input data, ReLU(x)=max(0,x) is the ReLU activation function, and ReLU(-x)=max(0,-x) is the negative part of ReLU. Represents a vector concatenation operation.

8. An interference identification system based on deep learning, applying the interference identification method, characterized in that: The interference identification system includes a deep learning module and a classification processing module: Deep learning module: The input signal is processed through the CV-R-BiLSTM-A network structure to obtain weighted feature data. The CV-R-BiLSTM-A network structure includes a complex convolutional network (CV), a residual network (R), a bidirectional long short-term memory network (BiLSTM), and an attention layer (A); Classification processing module: The weighted feature data is processed through a fully connected layer and a Softmax layer to obtain an interference recognition result.

9. An electronic device comprising a memory and a processor, characterized in that: Memory: used for storing a computer program for implementing the interference identification method; Processor: configured to implement the interference identification method when executing the computer program.

10. 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, the interference identification method is implemented.

Images