Wireless interference source automatic identification and classification system and method based on deep learning

Abstract

The invention relates to the technical field of wireless communication, and provides a wireless interference source automatic identification and classification system and method based on deep learning, and the method comprises the steps: 1, receiving an original wireless interference signal, and carrying out the denoising and filtering processing; signal energy is converted into a power spectrum and a time-frequency diagram through fast Fourier transform (FFT) and continuous wavelet transform (CWT); step 2, constructing a time-domain graph branch and a frequency-domain graph branch through a convolutional neural network CNN and a self-attention mechanism Transform architecture, and capturing frequency-domain features and time-frequency features of the interference signal at the same time; 3, fusing the frequency domain and time-frequency domain features through a multi-scale feature fusion network MT in combination with a convolutional neural network and a self-attention mechanism; and step 4, adopting a Softmax activation function to carry out classification identification on the fused features, and outputting the category of the wireless interference source. According to the invention, the wireless interference source can be efficiently, accurately and automatically identified and classified.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and more specifically, to an automatic identification and classification system and method for wireless interference sources based on deep learning. Background Art

[0002] Wireless communication systems are widely used in the civil aviation field. However, these systems are vulnerable to wireless interference, which can lead to a decline in communication quality and even cause system failures. Therefore, the accurate identification and classification of wireless interference sources have become a core technical issue for improving the anti-interference ability of communication systems. Existing wireless interference source identification methods can be mainly divided into classification methods based on traditional feature extraction and classification methods based on deep learning. Traditional methods often rely on manually designed features, which are computationally complex and difficult to adapt to changing interference environments. On the other hand, deep learning-based methods, especially convolutional neural networks (CNNs), although successful in many scenarios, have limitations in dealing with long-range dependencies due to their local computational characteristics, which affects the accuracy and efficiency of identification. In recent years, deep learning models incorporating self-attention mechanisms, such as Transformer, have been proven to have significant advantages in capturing long-range dependencies and enhancing feature representation capabilities.

[0003] Traditional wireless interference source identification methods usually rely on manually designed features, such as higher-order cumulants (HOC), time-frequency features, etc. These methods require manual selection of appropriate features and extraction, resulting in a large amount of manual intervention. The quality of feature design directly affects the identification accuracy, and different interference types may require different feature extraction methods, making it difficult for the system to adapt and optimize in a changing environment. Although the application of traditional convolutional neural networks (CNNs) in interference source identification can extract local features of signals through local convolutional operations, their local computational characteristics make CNNs perform poorly in capturing long-range dependencies. For interference signals with complex spatio-temporal features, CNNs cannot fully model global dependencies, thus affecting the accurate identification of interference signals, especially in complex and dynamically changing wireless environments. Although deep learning methods have achieved some success in interference source classification, most existing models have high computational complexity when faced with high-dimensional features and large-scale data. The training and inference speeds of the models are slow, especially in real-time applications, which may lead to the identification efficiency not meeting the actual requirements. In addition, the generalization ability of existing models is relatively weak, and overfitting is likely to occur in different wireless environments, resulting in the models being unable to effectively handle new or unseen interference types. The wireless interference environment is often complex and variable, and different types of interference signals may exhibit different features in the time domain, frequency domain, or time-frequency domain. Existing interference source identification methods perform poorly in scenarios where multiple interference signals coexist, and it is difficult to distinguish the subtle differences between interference sources, especially in cases where the noise is strong or the signal is weak. Traditional methods are prone to misclassification or missed classification in complex interference environments, limiting their effectiveness in practical applications. When traditional CNN architectures perform feature extraction, they mainly rely on local convolutional operations and are difficult to effectively capture long-range dependencies. Models based on self-attention mechanisms, such as Transformer, can model long-range dependencies through global interactions. However, existing CNN-based wireless interference source identification methods have not effectively introduced this global modeling ability, resulting in the inability to fully utilize the global information in interference signals during the extraction and classification of complex signal features. Multi-scale feature fusion is an important means to improve the accuracy of interference source identification, but existing methods usually only consider information at a single scale during feature fusion and fail to fully explore the feature information of signals at different scales. This approach leads to the ineffective combination of multi-scale information, thus affecting the identification accuracy of the model under different interference types and different signal-to-noise ratios.

[0004] Therefore, how to combine the local feature extraction ability of CNNs and the global dependency modeling ability of Transformer to propose an efficient and accurate automatic identification and classification system and method for wireless interference sources has become the key to solving the problems of the existing technology. Summary of the Invention

[0005] The content of the present invention is to provide a deep learning-based automatic identification and classification system and method for wireless interference sources, which can efficiently and accurately automatically identify and classify wireless interference sources.

[0006] A deep learning-based automatic identification and classification method for wireless interference sources according to the present invention includes the following steps:

[0007] Step 1: Receive the original wireless interference signal and perform denoising and filtering processing; convert the signal energy into a power spectrum diagram and a time-frequency diagram distributed in the frequency domain and the time-frequency domain through the fast Fourier transform (FFT) and the continuous wavelet transform (CWT);

[0008] Step 2: Construct a time-domain graph branch and a frequency-domain graph branch through a convolutional neural network (CNN) and a self-attention mechanism Transformer architecture respectively, and capture the frequency-domain features and time-frequency features of the interference signal at the same time; the time-domain graph branch takes the time-frequency graph of the interference signal as the input, and uses a 2D-CNN and a Transformer encoder to obtain embedded features; the input of the frequency-domain graph branch is the frequency-domain graph of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder;

[0009] Step 3: Combine the convolutional neural network and the self-attention mechanism to fuse the frequency-domain and time-frequency domain features through a deep multi-scale feature fusion network (MT), and capture the local and global dependencies in the signal;

[0010] Step 4: Use the Softmax activation function to classify and identify the fused features, and output the category of the wireless interference source.

[0011] Preferably, the types of interference signals include amplitude modulation (AM), binary phase shift keying (BPSK), single tone (CW), linear frequency modulation (LFM), multi-tone (MTJ), partial band noise (PBNJ), periodic pulse (PPNJ), sinusoidal frequency modulation (SFM).

[0012] Preferably, the denoising and filtering processing is specifically: limit the interference signal within a specific frequency band range through a band-pass filter to enhance the signal features while reducing the noise components in the signal.

[0013] Preferably, in the time-domain graph MDA-TFI branch, given an image with a size of as the input of the MDA-TFI branch, where W0 and H0 represent the width and height of the input image, and C0 represents the number of channels, construct a convolutional layer that maps O c to the embedded features, and this stage is expressed as:

[0014] O c = f c(m) (I, θ c )

[0015] In the formula, f c(m) represents the CNN stage with m convolutional layers, and θ c represents the trainable parameters of the convolutional layer; is the output of the convolutional stage, where the resolution is W×H and the number of channels is denoted as C;

[0016] Next, the feature O extracted by the CNN stage c is used to extract global features through the Transform stage; the Transform stage includes an encoder, and the input of the encoder of the Transform stage is a one-dimensional embedding sequence. O c is converted into a one-dimensional sequence, that is where has a dimension of C, and K = W×H represents the number of embedded tokens; all the embedded patch tokens are concatenated together to form

[0017] The time-frequency map branch is expressed as:

[0018]

[0019] α0 = concat[C cls , I'] + PE

[0020] α′ l = MSA(NL(α l-1 )) + α l-1

[0021] α″ l = FFN(NL(α′ l )) + α′ l

[0022] In the formula, θ are all the trainable parameters of the network, α0 is the initial variable, PE is the positional encoding, and α' l , α″ l are the L outputs of MSA and FFN respectively; NL is the normalization, and α l-1 is the output of the previous layer;

[0023] Finally, the FC layer using Softmax converts the L-th class label C cls into the classification probability.

[0024] Preferably, in the frequency-domain map branch, the convolutional layer and the Transformer layer extract local context and model global interaction respectively; when extracting the features of the frequency domain, it is expressed as The length of the original sequence is L0, and C f is the number of channels; a one-dimensional convolutional layer is adopted, which is expressed as:

[0025] X′F =f 1d-conv (X F ,φ)

[0026] where f 1d-conv (.,φ) is a series of one-dimensional convolution operations, is of length L and output channel C f Output; X F represents the original sequence, φ represents the parameters of the one-dimensional convolutional layer;

[0027] Next, for the embedded feature X' F Connect a classification tag Trainable 1D position embeddings Add directly to X' F ;after, It is called the input of the Transformer; it is then processed by stacked transformer layers as follows:

[0028] X′ F =f 1d-conv (X F ,φ)

[0029]

[0030] in, is the initial input feature of the Transformer encoder, X fp is the output of each layer in the Transformer encoder, and They are the output of the L-layer multi-head self-attention MSA and the frequency domain output of the L-layer feed-forward network FFN; the number of Transformer encoder layers of the multi-head self-attention mechanism MDN-FS is represented by N f ;The final output of MDN-FS is where x fclk is the output classification label of MDN-FS, and X fpt is the output block embedding of MDN-FS.

[0031] The present invention provides a system for automatic identification and classification of wireless interference sources based on deep learning, which adopts the method for automatic identification and classification of wireless interference sources based on deep learning as described in any one of claims 1 to 5, and comprises a signal preprocessing module, a feature extraction module MDA, a feature fusion module and a classification module connected in sequence;

[0032] The signal preprocessing module is used to receive the original wireless interference signal and perform denoising and filtering processing; the signal energy is converted into a power spectrum and time-frequency diagram distributed in the frequency domain and time-frequency domain through fast Fourier transform FFT and continuous wavelet transform CWT;

[0033] The feature extraction module includes a time-domain graph branch and a frequency-domain graph branch constructed by a convolutional neural network (CNN) and a self-attention mechanism (Transformer) architecture respectively, which are used to capture the frequency-domain features and time-frequency features of interference signals simultaneously. The time-domain graph branch takes the time-frequency graph of the interference signal as input and uses a 2D-CNN and a Transformer encoder to obtain embedded features. The input of the frequency-domain graph branch is the frequency-domain graph of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder.

[0034] The feature fusion module is used to fuse the frequency-domain and time-frequency-domain features through a deep feature fusion network by combining a convolutional neural network and a self-attention mechanism, so as to capture the local and global dependencies in the signal.

[0035] The classification and recognition module uses a Softmax activation function to classify and recognize the fused features and output the categories of wireless interference sources.

[0036] The beneficial effects of the present invention are as follows:

[0037] The present invention aims to improve the classification accuracy of wireless interference sources. Especially in a complex interference environment, it can accurately identify various types of interference signals, covering different spectral characteristics and interference patterns. At the same time, to further enhance the generalization ability of the model and optimize the adaptability of the model to different channel conditions and unknown interferences, the present invention adopts an improved multi-scale feature fusion mechanism (MT), combining the global dependency modeling ability of Transformer and the local feature extraction advantage of CNN, so that the system can still maintain stable recognition performance when facing different signal distributions and environmental changes. In addition, to improve the computing efficiency and optimize the computing resource consumption of the system, the present invention introduces a lightweight Transformer structure, optimizes the computational complexity of the attention mechanism, and combines one-dimensional convolution (Conv1D) to reduce the computational overhead, enabling the system to operate efficiently on low-computing-resource devices and being applicable to edge computing environments. Meanwhile, the present invention fully considers the system robustness during design, optimizes for different interference-to-noise ratios (INRs) and different sampling rate conditions, and ensures that the model can still maintain a high recognition accuracy in a harsh wireless environment. In addition, considering the actual application requirements, the present invention supports multiple hardware acceleration schemes, including heterogeneous computing platforms such as GPUs, FPGAs, and ASICs, so as to be flexibly deployed in different application scenarios, thereby further enhancing the scalability and practicality of the system. Finally, through the present invention, various types of wireless interference sources can be efficiently and accurately identified and classified in a complex wireless environment, providing a more intelligent, reliable, and efficient anti-interference solution for wireless communication systems. Description of the Drawings

[0038] Figure 1 It is a flowchart of an automatic identification and classification method for wireless interference sources based on deep learning;

[0039] Figure 2 It is a schematic diagram of the time-frequency diagram branch;

[0040] Figure 3 It is a schematic diagram of the frequency-domain diagram branch. Specific implementation manner

[0041] To further understand the content of the present invention, the present invention will be described in detail in combination with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and not for limiting it.

[0042] Embodiment

[0043] As Figure 1 shown, this embodiment provides an automatic identification and classification method for wireless interference sources based on deep learning, which includes the following steps:

[0044] Step 1: Receive the original wireless interference signal, and perform denoising and filtering processing; convert the signal energy into a power spectrum diagram and a time-frequency diagram distributed in the frequency domain and the time-frequency domain through the fast Fourier transform (FFT) and the continuous wavelet transform (CWT).

[0045] The types of interference signals include amplitude modulation (AM), binary phase shift keying (BPSK), single tone (CW), linear frequency modulation (LFM), multi-tone (MTJ), partial band noise (PBNJ), periodic pulse (PPNJ), sinusoidal frequency modulation (SFM).

[0046] Denoising and filtering processing: Limit the interference signal within a specific frequency band range through a band-pass filter to enhance the signal features while reducing the noise components in the signal for subsequent processing.

[0047] Convert the time-domain signal into a frequency-domain signal through the fast Fourier transform (FFT) in MATLAB to obtain the power spectrum diagram of the signal. Then convert the time-domain signal into a time-frequency domain signal through the continuous wavelet transform (CWT) to obtain the time-frequency diagram.

[0048] Step 2: Construct a time-domain diagram branch and a frequency-domain diagram branch through a convolutional neural network (CNN) and a self-attention mechanism Transformer architecture respectively, and capture the frequency-domain features and time-frequency features of the interference signal at the same time;

[0049] As Figure 2 shown, the time-domain diagram branch takes the time-frequency diagram of the interference signal as the input, and uses a 2D-CNN and a Transformer encoder to obtain embedded features; the input of the frequency-domain diagram branch is the frequency-domain diagram of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder;

[0050] In the time-domain graph MDA-TFI branch, given an image of size , as the input of the MDA-TFI branch, where W0 and H0 represent the width and height of the input image, and C0 represents the number of channels, a convolutional layer that maps O c to the embedded features is constructed. This stage is expressed as:

[0051] O c = f c(m) (I, θ c )

[0052] where f c(m) represents the CNN stage with m convolutional layers, and θ c represents the trainable parameters of the convolutional layer; is the output of the convolutional stage, where the resolution is W×H, and the number of channels is denoted as C;

[0053] Next, the features O c extracted in the CNN stage are used to extract global features through the Transform stage; the Transform includes an encoder. The input of the encoder of the Transform is a one-dimensional embedded sequence, and O c is converted into a one-dimensional sequence, that is, where has a dimension of C, and K = W×H represents the number of embedded tokens; all the embedded patch tokens are concatenated together to form

[0054] The time-frequency graph branch is expressed as:

[0055]

[0056] α0 = concat[C cls , I'] + PE

[0057] α′ l = MSA(NL(α l-1 )) + α l-1

[0058] α″ l = FFN(NL(α′ l )) + α′ l

[0059] where θ are all the trainable parameters of the network, α0 is the initial variable, PE is the positional encoding, and α' l , α″ l are the L outputs of MSA and FFN respectively; NL is the normalization, and α l-1 is the output of the previous layer;

[0060] Finally, the Softmax FC layer is used to label the Lth category C cls Convert to classification probability.

[0061] like Figure 3 As shown in the frequency domain graph branch, the convolution layer and the Transformer layer extract local context and model global interaction respectively; when extracting frequency domain features, it is represented as The original sequence length is L0, C f is the number of channels; a one-dimensional convolutional layer is used, expressed as:

[0062] X′ F =f 1d-conv (X F ,φ)

[0063] where f 1d-conv (.,φ) is a series of one-dimensional convolution operations, is of length L and output channel C f Output; X F represents the original sequence, φ represents the parameters of the one-dimensional convolutional layer;

[0064] Next, for the embedded feature X' F Connect a classification tag Trainable 1D position embeddings Add directly to X' F ;after, It is called the input of the Transformer; it is then processed by stacked transformer layers as follows:

[0065] X′ F =f 1d-conv (X F ,φ)

[0066]

[0067] in, is the initial input feature of the Transformer encoder, X fp is the output of each layer in the Transformer encoder, and They are the output of the L-layer multi-head self-attention MSA and the frequency domain output of the L-layer feed-forward network FFN; the number of Transformer encoder layers of the multi-head self-attention mechanism MDN-FS is represented by N f ;The final output of MDN-FS is where x fclk is the output classification label of MDN-FS, and X fpt is the output block embedding of MDN-FS.

[0068] Step 3: Combine the convolutional neural network and the self-attention mechanism to fuse the frequency-domain and time-frequency-domain features through the deep multi-scale feature fusion network MT, and capture the local and global dependencies in the signal.

[0069] Multi-scale fusion: Fuse the features extracted from the frequency domain and the time-frequency domain. The fusion process is processed by a deep neural network. In particular, the self-attention mechanism of the Transformer is combined with the local feature extraction ability of the CNN. The Transformer can model long-range dependencies through global interactions, while the CNN is good at extracting local features.

[0070] Feature extraction: Further process and optimize the fused features through a multi-layer fully connected network to obtain more effective interference signal features, and provide high-quality inputs for the subsequent classification module.

[0071] Step 4: Use the Softmax activation function to classify and identify the fused features, and output the category of the wireless interference source.

[0072] This embodiment provides a deep learning-based automatic identification and classification system for wireless interference sources, which adopts the above-mentioned deep learning-based automatic identification and classification method for wireless interference sources, and includes a signal preprocessing module, a feature extraction module MDA, a feature fusion module, and a classification module connected in sequence;

[0073] The signal preprocessing module is used to receive the original wireless interference signal and perform denoising and filtering processing; convert the signal energy into a power spectrum diagram and a time-frequency diagram distributed in the frequency domain and the time-frequency domain through the fast Fourier transform FFT and the continuous wavelet transform CWT;

[0074] The feature extraction module includes a time-domain graph branch and a frequency-domain graph branch constructed by the convolutional neural network CNN and the self-attention mechanism Transformer architecture respectively, and is used to simultaneously capture the frequency-domain features and time-frequency features of the interference signal; the time-domain graph branch takes the time-frequency graph of the interference signal as input, and uses a 2D-CNN and a Transformer encoder to obtain embedded features; the input of the frequency-domain graph branch is the frequency-domain graph of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder;

[0075] The feature fusion module is used to combine the convolutional neural network and the self-attention mechanism to fuse the frequency-domain and time-frequency-domain features through a deep feature fusion network, and capture the local and global dependencies in the signal;

[0076] The classification and recognition module uses the Softmax activation function to classify and identify the fused features, and outputs the category of the wireless interference source.

[0077] In this embodiment, by combining the advantages of CNN and Transformer, the present invention can effectively extract the local and global features of interference signals, thus significantly improving the recognition accuracy of wireless interference sources. Especially in a complex wireless environment, the system can identify various types of interference sources, and the classification accuracy is better than that of traditional methods.

[0078] In this embodiment, through the multi-scale feature fusion mechanism, the present invention can better adapt to different scales of signals and enhance the robustness of the system under different interference types and signal intensities. This mechanism enables the system to maintain high classification performance even in the face of complex interference scenarios.

[0079] This embodiment can operate efficiently in a dynamically changing and complex wireless communication environment. It can not only identify known types of interference sources but also has good adaptability and scalability, capable of adapting to the recognition tasks of new types of interference sources.

[0080] The above has schematically described the present invention and its implementation manners. This description is not restrictive, and what is shown in the drawings is only one of the implementation manners of the present invention. The actual structure is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural manners and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. An automatic identification and classification method for wireless interference sources based on deep learning, characterized in that: It includes the following steps: Step 1: Receive the original wireless interference signal and perform denoising and filtering processing; convert the signal energy into power spectrum diagrams and time-frequency diagrams distributed in the frequency domain and time-frequency domain through fast Fourier transform (FFT) and continuous wavelet transform (CWT); Step 2: Construct a time-domain diagram branch and a frequency-domain diagram branch respectively through a convolutional neural network (CNN) and a self-attention mechanism Transformer architecture to capture the frequency-domain features and time-frequency features of the interference signal simultaneously; The time-domain diagram branch takes the time-frequency diagram of the interference signal as input and uses 2D-CNN and a Transformer encoder to obtain embedded features; the input of the frequency-domain diagram branch is the frequency-domain diagram of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder; Step 3: Combine the convolutional neural network and the self-attention mechanism to fuse the frequency-domain and time-frequency domain features through a deep multi-scale feature fusion network (MT) to capture the local and global dependencies in the signal; Step 4: Use the Softmax activation function to classify and identify the fused features and output the category of the wireless interference source.

2. The automatic identification and classification method of wireless interference sources based on deep learning according to claim 1, characterized in that: The types of interference signals include amplitude modulation (AM), binary phase shift keying (BPSK), single tone (CW), linear frequency modulation (LFM), multi-tone (MTJ), partial band noise (PBNJ), periodic pulse (PPNJ), and sinusoidal frequency modulation (SFM).

3. The automatic identification and classification method of wireless interference sources based on deep learning according to claim 2, characterized in that: The denoising and filtering processing specifically is: limit the interference signal within a specific frequency band range through a band-pass filter to enhance the signal features while reducing the noise components in the signal.

4. The automatic identification and classification method of wireless interference sources based on deep learning according to claim 3, characterized in that: In the time-domain graph MDA-TFI branch, given an image of size as the input to the MDA-TFI branch, where W0 and H0 represent the width and height of the input image, and C0 represents the number of channels, a convolutional layer that maps O c to the embedded feature is constructed. This stage is expressed as: O c = f c(m) (I, θ c ) where f c(m) represents the CNN stage with m convolutional layers, and θ c represents the trainable parameters of the convolutional layer; is the output of the convolutional stage, where the resolution is W×H and the number of channels is denoted as C; Next, the features O extracted in the CNN stage c are used to extract global features through the Transform stage; The Transform includes an encoder. The encoder input of the Transform is a one-dimensional embedding sequence, which transforms O c into a one-dimensional sequence, that is where has a dimension of C, and K = W × H represents the number of embedded tokens; all the embedded patch tokens are concatenated together to form The time-frequency diagram branch is represented as: α0 = concat[C cls , I'] + PE α' l = MSA(NL(α l-1 )) + α l-1 α” l = FFN(NL(α' l )) + α' l where θ are all the trainable parameters of the network, α0 is the initial variable, PE is the positional encoding, α' l , α” l are the L outputs of MSA and FFN respectively; NL is the normalization, and α l-1 is the output of the previous layer; Finally, the FC layer using Softmax converts the $L$-th class label $C$ cls into classification probabilities.

5. The automatic identification and classification method of wireless interference sources based on deep learning according to claim 4, characterized in that: In the frequency domain graph branch, the convolutional layer and the Transformer layer extract local context and model global interactions respectively; when extracting features in the frequency domain, it is expressed as The original sequence length is L0, and C f is the number of channels; a one-dimensional convolutional layer is used, which is expressed as: X' F = f 1d-conv (X F , φ) where f 1d-conv (., φ) is a series of one-dimensional convolution operations, which is the output of length L and output channels C f ; X F represents the original sequence, and φ represents the parameters of the one-dimensional convolutional layer; Next, for the embedded feature X' F Connect a classification token Trainable one-dimensional positional embedding And directly add it to X' F ; After that, It is called the input of the Transformer; and then it is processed through a stacked transformer layer as follows: X' F = f 1d-conv (X F , φ) Among them, is the initial input feature of the Transformer encoder, X fp is the output of each layer in the Transformer encoder, and are the output of the multi-head self-attention MSA of layer L and the frequency-domain output of the feed-forward network FFN of layer L, respectively; the number of layers of the Transformer encoder of the multi-head self-attention mechanism MDN-FS is denoted as N f ; the final output of MDN-FS is where x fclk is the output classification token of MDN-FS, while X fpt is the output block embedding of MDN-FS.

6. An automatic identification and classification system for wireless interference sources based on deep learning, characterized in that: It adopts the automatic identification and classification method of wireless interference sources based on deep learning as described in any one of claims 1-5, and includes a signal preprocessing module, a feature extraction module (MDA), a feature fusion module, and a classification module connected in sequence; The signal preprocessing module is used to receive the original wireless interference signal and perform denoising and filtering processing; convert the signal energy into power spectrum diagrams and time-frequency diagrams distributed in the frequency domain and time-frequency domain through fast Fourier transform (FFT) and continuous wavelet transform (CWT); The feature extraction module includes a time-domain diagram branch and a frequency-domain diagram branch constructed respectively through a convolutional neural network (CNN) and a self-attention mechanism Transformer architecture, and is used to capture the frequency-domain features and time-frequency features of the interference signal simultaneously; The time-domain diagram branch takes the time-frequency diagram of the interference signal as input and uses 2D-CNN and a Transformer encoder to obtain embedded features; the input of the frequency-domain diagram branch is the frequency-domain diagram of the interference signal, and the processing layer includes a 1D-CNN layer and a Transformer encoder; The feature fusion module is used to combine the convolutional neural network and the self-attention mechanism to fuse the frequency-domain and time-frequency domain features through a deep feature fusion network to capture the local and global dependencies in the signal; The classification and recognition module uses the Softmax activation function to classify and identify the fused features and output the category of the wireless interference source.

Images