A deep learning-based wireless signal device detection method and system

Abstract

The application discloses a wireless signal device detection method and system based on deep learning, which comprises the following steps: step one, collecting and preprocessing original signal data; step two, using fast Fourier transform to extract time-frequency features and perform multi-spectrum analysis; step three, performing Chebyshev polynomial approximation convolution operation through a ChebNet network to construct a signal correlation graph; step four, performing spectrum analysis through an improved PinSAGE model to obtain a spectrum feature set; step five, extracting device features of each wireless signal device and matching the device features with the spectrum feature set; step six, performing time-space clustering on the matched feature vector set; and step seven, analyzing the device type and working state of the wireless signal device. Through the ChebNet network and the improved PinSAGE model, the application optimizes signal analysis and interference suppression, and realizes high-precision detection of wireless devices.

Description

Technical Field

[0001] This invention relates to the field of wireless device detection technology, and in particular to a method and system for detecting wireless signal devices based on deep learning. Background Technology

[0002] With the continuous development of wireless communication technology, wireless signal devices have been widely used in modern society, especially in the fields of the Internet of Things, smart devices, industrial automation, and smart cities. Wireless signal devices transmit data and communicate via wireless signals; therefore, accurate detection and management of these devices are crucial. Traditional wireless signal device detection methods typically rely on manual methods or simple algorithm-based detection tools, which suffer from low accuracy, low efficiency, and inability to handle complex signal environments. With the development of deep learning technology, deep learning-based signal analysis methods have gradually become a research hotspot in the field of wireless signal device detection, providing efficient and accurate device identification and status monitoring in more complex signal environments. Existing wireless signal device detection methods mainly rely on time-domain and frequency-domain signal analysis, identifying device type and status by extracting time-frequency features and power spectra. However, these methods typically face several problems. First, traditional spectrum analysis methods have limitations in extracting signal frequency band features, failing to effectively address complex noise interference and signal overlap between devices, resulting in low device identification accuracy. Second, during the detection process of wireless signal devices, the spatiotemporal characteristics of the signal and the correlation between devices are often ignored, leading to an inability to comprehensively capture device behavior patterns and signal changes. Therefore, existing technologies perform poorly in dynamic interference environments and are unable to meet the requirements for high precision and high real-time performance.

[0003] To address the aforementioned issues, deep learning technology has been introduced into wireless signal device detection, particularly the combination of Graph Neural Networks (GNNs) and Convolutional Neural Networks (CNNs), which is widely used for signal feature learning and device identification. However, existing deep learning-based wireless signal device detection methods still face several technical challenges. Firstly, existing methods typically employ traditional signal processing and feature extraction algorithms, failing to fully exploit the complex correlations between signals. For instance, traditional convolutional neural networks often focus only on local signal features, neglecting the spatiotemporal correlations between signals from different devices and the mutual influence of their spectra, thus affecting the accuracy of signal analysis results. Secondly, while GNNs can model the correlations between signal sources to some extent, existing technologies still have shortcomings in how to better combine spectrograms and signal correlation diagrams, and how to effectively process multi-spectral information. Therefore, existing methods exhibit poor robustness in multi-device and complex interference environments, making it difficult to achieve efficient matching and accurate classification between signals and devices.

[0004] Therefore, how to provide a method and system for detecting wireless signal devices based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method and system for detecting wireless signal devices. This invention optimizes signal analysis and interference suppression by combining the ChebNet network and an improved PinSAGE model, thereby improving the accuracy and robustness of wireless signal device detection. High-quality time-frequency features are extracted using Fast Fourier Transform and multi-spectral analysis, enhancing the accuracy of signal-device matching. K-Means clustering and spatiotemporal relationship identification effectively remove interference signals, ensuring signal set optimization and improving the accuracy of device classification, operational status assessment, and interference source identification, providing reliable support for intelligent monitoring of wireless devices.

[0006] A method for detecting wireless signal devices based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Collect raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure; Step 2: Use Fast Fourier Transform to extract the time-frequency features of the wireless signal data, and perform multi-spectral analysis to generate time-frequency feature vectors for each frequency band; Step 3: Input the time-frequency feature vector of each frequency band into the ChebNet network, and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; Step 4: Input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module, and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. Step 5: Extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; Step 6: Perform spatiotemporal clustering on the matching feature vector set to identify the spatiotemporal relationships between signals, and perform interference suppression to generate an interference suppression signal set; Step 7: Based on the interference suppression signal set, analyze the device type and operating status of the wireless signal device to generate the final detection result.

[0007] Optionally, the raw signal data includes time-domain data, frequency-domain data, and ambient noise; the preprocessing steps specifically include: A bandpass filter is used to filter out low-frequency noise below a preset frequency in the time domain data of the original signal data, and the corresponding bandwidth is set according to the frequency band of the received signal. For the frequency domain data in the original signal data, a high-pass filter is used to filter out high-frequency interference signals with frequencies greater than a preset frequency.

[0008] Optionally, step two specifically involves: The wireless signal data is segmented into multiple wireless signal data segments using windowing processing. A fast Fourier transform is applied to each wireless signal data segment to obtain the frequency domain representation of the current wireless signal data segment. By calculating the amplitude spectrum and phase spectrum in the frequency domain, the amplitude and phase information of each frequency band are obtained, and a spectrum diagram is generated, which represents the energy distribution of the signal in different frequency bands; The spectrum in the spectrogram is divided into multiple frequency bands. The frequency range and energy distribution of each frequency band are analyzed to identify the frequency range in the signal and calculate the signal-to-noise ratio of each frequency band. Select the spectral data corresponding to the frequency bands whose frequency range is within the preset threshold and whose signal-to-noise ratio is greater than the preset threshold to obtain the filtered time-frequency features; By combining the frequency characteristics, amplitude, and phase information of each frequency band in the filtered time-frequency features, a time-frequency feature vector for each frequency band is generated.

[0009] Optionally, step three specifically includes: The time-frequency feature vector of each frequency band is input into the input layer of the ChebNet network, and the input layer uses the time-frequency feature vector of each frequency band as a node to obtain a node set; Calculate the Manhattan distance between every two nodes in the node set. If the Manhattan distance is less than a preset distance threshold, establish an edge connection between the corresponding nodes to obtain an edge connection set. Construct a time-frequency feature map based on the node set and edge connection set; The time-frequency feature map is input into the convolutional layer of the ChebNet network. Through Chebyshev polynomial approximate convolution operation, the aggregated features of the nodes are generated, and the aggregated feature map is obtained. The aggregated feature map is input into a nonlinear activation layer, and the ReLU nonlinear activation function is used to activate the features of each node, reducing negative features to zero while keeping positive features unchanged, thus generating a signal correlation map.

[0010] Optionally, the improved PinSAGE model is specifically as follows: In the spectrum band node construction module, each frequency band in the spectrum graph is mapped to a spectrum band node. Each spectrum band node represents the feature data of a frequency band, generating a spectrum band node set. The spectrum band node includes the amplitude frequency domain features and phase frequency domain features of each spectrum band. The spectral band nodes are input into the signal feature encoding module, and a convolutional autoencoder is used to encode the input set of spectral band nodes. The convolutional autoencoder extracts local spectral features of the spectral band through multiple convolutional layers, and maps the local spectral features of each node into a low-dimensional embedding vector through a spectral reconstruction mechanism to generate a set of spectral feature vectors. The spectral reconstruction mechanism is as follows: The local spectral features of each node are selected from the spectrum diagram. By selecting multiple consecutive spectral bands, the amplitude and phase information are combined into a matrix to form the local spectral matrix of the corresponding node. Principal component analysis is used to transform the spectral matrix into a low-dimensional vector representation, thus obtaining the spectral eigenvectors. The spectral feature vector set is input to the multimodal node association and transmission module. During node association and transmission, information is transmitted by combining the spectral feature vector and the connection relationship in the signal association graph to generate a fused association feature set. The information transmission is based on the edge connection relationship in the signal association graph, and the spectral feature vector is propagated in the signal association graph. Each node receives the information of the neighboring nodes and updates it by weighted average in combination with its own feature vector. The fused correlation feature set is input into the spectrum selection module. Different fused correlation features in the fused correlation feature set are used as nodes, and the Euclidean distance between different nodes is calculated as the edge weight of the connection edge to obtain the fused correlation feature map. The spectral features corresponding to nodes whose edge weights in the fusion correlation feature graph are greater than a preset threshold are selected to obtain the final spectral feature set.

[0011] Optionally, step five specifically includes: Extract the device features of each wireless signal device to generate a device feature vector set. The device features include the device model, working status, device location, and transmitted signal strength. Calculate the cosine similarity between each device feature vector and spectrum feature vector in the device feature set and spectrum feature set, respectively. If the cosine similarity is greater than the preset similarity threshold, it is considered that there is a strong similarity between the signal and the device, and a corresponding matching feature vector set is generated.

[0012] Optionally, step six specifically includes: K vectors are randomly selected from the set of matching feature vectors as initial cluster centers, where K is a preset number of clusters; Calculate the Euclidean distance between each matching feature vector and the initial cluster center, and assign each vector to the initial cluster center with the closest Euclidean distance; Based on the mean of all matching feature vectors in each cluster, the position of the cluster center is recalculated, and the cluster center is iteratively updated until the change of the cluster center is less than the preset threshold or the maximum number of iterations is reached, thus obtaining the final K cluster centers. Based on the clustering results, the matching feature vector set is divided into K signal clusters, each signal cluster corresponding to a spatiotemporal signal relationship, thus generating a spatiotemporal signal cluster set; By analyzing the signals in each spatiotemporal signal cluster, interference signals are identified. The interference signals are those whose difference from the mean characteristic value of the spatiotemporal signal cluster is greater than a preset threshold, and whose signal strength is lower than a preset threshold or whose spectral characteristics are inconsistent. The identified interference signals are filtered out to generate an interference suppression signal set.

[0013] Optionally, step seven specifically includes: Based on the interference suppression signal set, the transmission mode, spectral bandwidth and signal strength information of the signal are analyzed and matched with the predefined device type to determine the device type to which the signal source belongs, and the final device type identification result is obtained. The device type identification result indicates the device type corresponding to each signal source. Based on the frequency changes and amplitude fluctuations of the interference suppression signal, the working status of the equipment is assessed to determine whether the equipment is in normal working condition or whether there is an abnormal working condition, and the equipment working status assessment result is generated. Based on the equipment type identification results and equipment working status assessment results, interference sources are identified using signal interference information, and the location and type of interference sources are determined, generating interference source identification results. The final detection result is generated by combining the device type identification result, the device working status assessment result, and the interference source identification result.

[0014] A wireless signal device detection system based on deep learning according to an embodiment of the present invention includes the following modules: The data acquisition and preprocessing module is used to acquire raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure. The time-frequency feature analysis module is used to extract the time-frequency features of the wireless signal data using fast Fourier transform and to perform multi-spectral analysis to generate a time-frequency feature vector for each frequency band. The signal correlation graph construction module is used to input the time-frequency feature vector of each frequency band into the ChebNet network and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; The spectrum analysis module is used to input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. The matching analysis module is used to extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; The spatiotemporal clustering module is used to perform spatiotemporal clustering on the matching feature vector set, identify the spatiotemporal relationship between signals, and perform interference suppression to generate an interference suppression signal set. The final detection module is used to analyze the device type and operating status of the wireless signal device based on the interference suppression signal set, and generate the final detection result.

[0015] The beneficial effects of this invention are: The method of this invention effectively addresses several challenges faced by existing wireless signal device detection technologies. Existing wireless signal device detection methods are unstable and have low accuracy in complex environments, especially with multiple devices and signal interference. However, by combining deep learning techniques, particularly the improved PinSAGE model and ChebNet network, the spatiotemporal correlations between signal sources can be better captured, improving the accuracy of device identification and signal analysis. Utilizing Fast Fourier Transform (FFT) and multi-spectral analysis, the time-frequency features of the wireless signal are extracted and optimized, eliminating the limitations of traditional methods in signal frequency band feature extraction, resulting in more accurate and clear spectral information. This process effectively improves signal identification, especially in environments with strong spectral interference, significantly enhancing the accuracy of device identification. The improved PinSAGE model, through deep integration of the spectrogram and signal correlation graph, enhances the model's ability to learn complex correlations between signal sources, effectively avoiding the problem of insufficient processing of signal correlations between devices in traditional methods. Precise matching of device features and spectral features ensures accurate correspondence between each signal source and device, further improving the accuracy of device type and operating status identification. Simultaneously, by performing spatiotemporal clustering on the signals, the spatiotemporal relationships between signals can be identified, signal clusters can be accurately divided, the impact of interference signals on effective signals can be reduced, and signal quality can be improved. The introduction of this interference suppression mechanism makes signal analysis and device identification more robust in multi-device and complex environments. By combining K-Means clustering and interference suppression techniques, the method of this invention can automatically identify and remove irrelevant interference signals, effectively optimize the signal set, and ensure the efficiency and accuracy of device detection. This method not only improves the accuracy of device classification and operational status assessment but also enables real-time signal monitoring and fault diagnosis in practical applications, providing strong support for the intelligent management of wireless devices. Therefore, using the method of this invention can significantly improve the accuracy, efficiency, and adaptability of wireless signal device detection, overcoming the problems of traditional technologies being unable to cope with complex interference and multi-device environments, and has broad application prospects and practical value. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based wireless signal device detection method proposed in this invention. Figure 2 This is a schematic diagram of the structure of a deep learning-based wireless signal device detection system proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 A method for detecting wireless signal devices based on deep learning, comprising: Step 1: Collect raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure; Step 2: Use Fast Fourier Transform to extract the time-frequency features of the wireless signal data, and perform multi-spectral analysis to generate time-frequency feature vectors for each frequency band; Step 3: Input the time-frequency feature vector of each frequency band into the ChebNet network, and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; Step 4: Input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module, and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. Step 5: Extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; Step 6: Perform spatiotemporal clustering on the matching feature vector set to identify the spatiotemporal relationships between signals, and perform interference suppression to generate an interference suppression signal set; Step 7: Based on the interference suppression signal set, analyze the device type and operating status of the wireless signal device to generate the final detection result.

[0019] In this embodiment, the raw signal data includes time-domain data, frequency-domain data, and environmental noise. Raw signal data is typically generated by wireless signal devices and transmitted through a transmission path, and may be affected by environmental noise, signal attenuation, or other signal interference. Therefore, preprocessing the raw signal data is a crucial step in improving the accuracy of subsequent signal analysis and device detection.

[0020] The preprocessing steps specifically include: Time-domain data preprocessing: The time-domain data in the original signal is filtered using a bandpass filter. The bandpass filter removes low-frequency noise below a preset frequency while preserving the effective frequency bands of the signal. Specifically, the low-frequency noise includes DC offset noise below 1 Hz and low-frequency interference signals caused by environmental changes. For example, if the target signal's frequency range is 10 Hz to 5 kHz, the preset frequency thresholds can be set to 1 Hz and 5 kHz. The bandpass filter removes noise below 1 Hz and filters out high-frequency noise above 5 kHz to preserve the frequency components of the target signal. In this way, the impact of low-frequency noise on signal quality during signal acquisition can be reduced, while ensuring the integrity of important frequency components.

[0021] Frequency domain data preprocessing: For the frequency domain data in the original signal data, a high-pass filter is used for filtering. The function of the high-pass filter is to filter out high-frequency interference signals above a preset frequency in the frequency domain, ensuring that the spectral portion of the signal is within the required effective frequency band. For example, for a target signal with a frequency band of 100Hz to 3kHz, the preset threshold for high-frequency interference signals can be set to 3kHz. The high-pass filter will filter out all high-frequency interference signals above 3kHz, thereby reducing the impact of high-frequency noise and unwanted frequency components on signal analysis.

[0022] Through the above filtering steps, noise components in the signal are effectively suppressed while preserving important information, providing a clearer and more accurate data foundation for subsequent feature extraction and signal analysis. This filtering method effectively improves signal quality, making subsequent signal analysis and equipment detection more precise, and reducing errors and uncertainties caused by signal noise.

[0023] In this embodiment, step two specifically includes: Windowing is employed to segment wireless signal data into multiple segments. The purpose of windowing is to divide continuous signal data into several time-domain segments, each of which can be analyzed independently. Specifically, common window types include rectangular windows, Hamming windows, and Hanning windows, with the appropriate window function selected based on the signal characteristics. The length of each wireless signal data segment can be determined according to the signal frequency and analysis requirements, typically set to a duration between 10 ms and 1 second. This segmentation effectively reduces the impact of signal non-stationarity on frequency domain analysis, ensuring a certain degree of temporal stability for each signal segment.

[0024] A Fast Fourier Transform (FFT) is applied to each wireless signal data segment to convert the time-domain signal into a frequency-domain signal. FFT is an efficient algorithm for calculating the Fourier transform; by decomposing the time-domain signal into different frequency components, it can accurately obtain the frequency-domain representation of the current wireless signal data segment. The frequency-domain representation of each data segment includes amplitude and phase information in the spectrum, and the FFT processing result typically contains two parts: amplitude spectrum and phase spectrum.

[0025] By calculating the amplitude and phase spectra in the frequency domain, the amplitude and phase information of each frequency band are obtained. The amplitude spectrum represents the intensity of different frequency components, reflecting the energy distribution of the signal in various frequency bands; the phase spectrum provides the phase information of each frequency component of the signal. By calculating these spectral values, a spectrogram is generated, which represents the energy distribution of the signal in different frequency bands. The spectrogram can intuitively display the intensity and distribution of the signal in various frequency bands and is the basis for subsequent analysis.

[0026] The spectrum in a spectrogram is divided into multiple frequency bands. Each band corresponds to a specific frequency range of the signal, and the choice of band usually depends on the signal's operating frequency range. For example, if a wireless signal device operates in the 2.4 GHz Wi-Fi band, the spectrogram can be divided into several smaller bands, and the frequency range and energy distribution of each band can be analyzed. By analyzing the frequency range, signal strength, and energy distribution of each band, the characteristics of each frequency band in the signal can be effectively identified.

[0027] Based on the analysis of the frequency range and energy distribution of each frequency band, the signal-to-noise ratio (SNR) of each band is calculated. SNR is a crucial indicator of signal quality. For each frequency band in the spectrum, its SNR is calculated. If the SNR is greater than a preset threshold, the band is considered a valid signal band. The preset SNR threshold can be set between 10dB and 30dB, and can be adjusted according to different application scenarios and signal requirements.

[0028] Spectral data corresponding to frequency bands with a frequency range within a preset threshold and a signal-to-noise ratio greater than a preset threshold are selected and filtered to obtain filtered time-frequency features. The filtered time-frequency features typically include information such as the amplitude, phase, and frequency range of the effective frequency bands. Then, by combining the frequency characteristics, amplitude, and phase information of each frequency band in the filtered time-frequency features, a time-frequency feature vector is generated for each frequency band. Each time-frequency feature vector represents the characteristics of that frequency band and can be used for subsequent device detection, signal analysis, and classification tasks.

[0029] Through these steps, the final time-frequency feature vector set not only retains the frequency domain information of the signal but also removes noise components, ensuring the accuracy and effectiveness of subsequent analysis. These time-frequency feature vectors will serve as input data for subsequent device identification and signal classification, greatly improving the accuracy and robustness of the detection method.

[0030] In this embodiment, step three specifically includes: The time-frequency feature vector of each frequency band is input into the input layer of the ChebNet network, and the input layer uses the time-frequency feature vector of each frequency band as a node to obtain a node set; Calculate the Manhattan distance between every two nodes in the node set. If the Manhattan distance is less than a preset distance threshold, establish an edge connection between the corresponding nodes to obtain an edge connection set. Construct a time-frequency feature map based on the node set and edge connection set; The time-frequency feature map is input into the convolutional layer of the ChebNet network. Through Chebyshev polynomial approximate convolution operation, the aggregated features of the nodes are generated, and the aggregated feature map is obtained. The aggregated feature map is input into a nonlinear activation layer, and the ReLU nonlinear activation function is used to activate the features of each node, reducing negative features to zero while keeping positive features unchanged, thus generating a signal correlation map.

[0031] This step, by utilizing the ChebNet network and graph convolution methods, effectively improves the accuracy of signal correlation modeling and feature extraction in wireless signal device detection. Traditional signal analysis methods typically only process local signal features and cannot comprehensively capture the complex correlations between signals between devices. By inputting the time-frequency feature vectors of each frequency band into the ChebNet network and constructing a signal correlation graph by calculating Manhattan distance, the system can more accurately reveal the relationships between signal sources, thereby improving device type identification and signal matching. Furthermore, through the combination of Chebyshev polynomial approximate convolution operations and the nonlinear activation function (ReLU), the system can not only efficiently aggregate node features but also effectively suppress noise and interference, preserving valid signal features. This structure makes signal feature extraction more accurate and robust, enabling stable operation in complex wireless signal environments. The generated signal correlation graph provides reliable data support for subsequent device detection and classification, solving the detection accuracy problems caused by mutual interference or signal instability in traditional methods. This method, through accurate signal correlation analysis and optimization, effectively improves the accuracy, efficiency, and robustness of wireless signal device detection.

[0032] In this embodiment, the improved PinSAGE model is specifically as follows: In the spectrum band node construction module, each frequency band in the spectrum graph is mapped to a spectrum band node. Each spectrum band node represents the feature data of a frequency band, generating a spectrum band node set. The spectrum band node includes the amplitude frequency domain features and phase frequency domain features of each spectrum band. The spectral band nodes are input into the signal feature encoding module, and a convolutional autoencoder is used to encode the input set of spectral band nodes. The convolutional autoencoder extracts local spectral features of the spectral band through multiple convolutional layers, and maps the local spectral features of each node into a low-dimensional embedding vector through a spectral reconstruction mechanism to generate a set of spectral feature vectors. The spectral reconstruction mechanism is as follows: The local spectral features of each node are selected from the spectrum diagram. By selecting multiple consecutive spectral bands, the amplitude and phase information are combined into a matrix to form the local spectral matrix of the corresponding node. Principal component analysis is used to transform the spectral matrix into a low-dimensional vector representation, thus obtaining the spectral eigenvectors. The spectral feature vector set is input to the multimodal node association and transmission module. During node association and transmission, information is transmitted by combining the spectral feature vector and the connection relationship in the signal association graph to generate a fused association feature set. The information transmission is based on the edge connection relationship in the signal association graph, and the spectral feature vector is propagated in the signal association graph. Each node receives the information of the neighboring nodes and updates it by weighted average in combination with its own feature vector. The fused correlation feature set is input into the spectrum selection module. Different fused correlation features in the fused correlation feature set are used as nodes, and the Euclidean distance between different nodes is calculated as the edge weight of the connection edge to obtain the fused correlation feature map. The spectral features corresponding to nodes whose edge weights in the fusion correlation feature graph are greater than a preset threshold are selected to obtain the final spectral feature set.

[0033] This step significantly improves the accuracy and robustness of wireless signal device detection through an improved PinSAGE model. Compared with traditional methods, this invention can extract the spectral features of signals more accurately and enhances the ability to model the correlation between signal sources through deep fusion of spectrograms and signal correlation graphs. Encoding spectral features using a convolutional autoencoder effectively compresses the dimensionality of spectral information while retaining important frequency, amplitude, and phase features, making subsequent analysis more efficient. The combination of a spectrum reconstruction mechanism and principal component analysis optimizes feature representation and improves the quality of signal analysis.

[0034] Furthermore, the information transmission mechanism ensures accurate transmission of correlation information between nodes by combining the characteristics of neighboring nodes for weighted average updates, thus enhancing the learning ability of spatiotemporal correlations between signal sources. This approach significantly improves the model's stability and accuracy in complex interference environments.

[0035] By filtering highly correlated features through a spectrum selection module, redundant information is reduced, and an optimized spectrum feature set is ultimately generated. This makes the matching of signal sources and devices more accurate, improving the effectiveness of device detection and classification. Overall, this invention provides an efficient and accurate method for wireless signal device detection through precise spectrum feature extraction, interference suppression, and signal correlation learning. It overcomes many limitations of traditional technologies and has high practicality and potential for widespread application.

[0036] In this embodiment, step five specifically includes: Device features are extracted from each wireless signal device. These features include the device model, operating status, device location, and transmitted signal strength. The device model represents the device category or type; different models typically have different operating characteristics and signal transmission features. The device operating status indicates the device's current operational status, such as whether it is in normal operation or standby mode. The device location refers to the device's position in physical space, which is usually closely related to the signal propagation path and signal strength. The transmitted signal strength reflects the device's transmission power, a crucial feature in wireless communication. These features comprehensively describe the operating status of each device and its correlation with the signal. After all these features are extracted, a device feature vector set is generated, where each device feature vector represents complete feature information for a single device.

[0037] After obtaining the device feature vector set, the next step is to calculate the similarity between each device feature vector and the spectral feature vector. For this purpose, cosine similarity is used as the similarity metric. Cosine similarity measures the similarity between two vectors by calculating the angle between them.

[0038] If the calculated cosine similarity is greater than a preset similarity threshold, a strong similarity is considered to exist between the signal and the device, meaning that the device may be the source of the signal or has a strong correlation with it. In this case, a corresponding matching feature vector set is generated. This matching feature vector set contains matching relationship information between the signal source and the device, with each matching feature vector representing the degree of correlation and characteristics between a signal and a device. In this way, a precise matching relationship can be established between the device and the signal, providing a reliable basis for subsequent device classification, status assessment, etc.

[0039] In practical applications, the similarity threshold is dynamically adjusted based on factors such as the device's operating status, signal type, and signal strength. Typically, the preset similarity threshold can be adjusted between 0.8 and 0.95 to ensure high matching accuracy. Higher similarity thresholds (e.g., above 0.9) are generally used for precise matching when device and signal characteristics are highly consistent, while lower similarity thresholds are suitable for preliminary matching in situations with weak signal strength or high environmental noise.

[0040] Through the above steps, the system can accurately match each signal source with its corresponding device in a multi-device, multi-signal environment, providing high-precision data support for device type identification, status monitoring, and subsequent signal analysis.

[0041] In this embodiment, step six specifically includes: K vectors are randomly selected from the set of matching feature vectors as initial cluster centers, where K is a preset number of clusters; Calculate the Euclidean distance between each matching feature vector and the initial cluster center, and assign each vector to the initial cluster center with the closest Euclidean distance; Based on the mean of all matching feature vectors in each cluster, the position of the cluster center is recalculated, and the cluster center is iteratively updated until the change of the cluster center is less than the preset threshold or the maximum number of iterations is reached, thus obtaining the final K cluster centers. Based on the clustering results, the matching feature vector set is divided into K signal clusters, each signal cluster corresponding to a spatiotemporal signal relationship, thus generating a spatiotemporal signal cluster set; By analyzing the signals in each spatiotemporal signal cluster, interference signals are identified. The interference signals are those whose difference from the mean characteristic value of the spatiotemporal signal cluster is greater than a preset threshold, and whose signal strength is lower than a preset threshold or whose spectral characteristics are inconsistent. The identified interference signals are filtered out to generate an interference suppression signal set.

[0042] This step significantly improves the accuracy and stability of wireless signal device detection by introducing K-Means clustering and interference suppression techniques. First, K-Means clustering divides signal matching feature vectors into multiple signal clusters, making the spatiotemporal relationships between signals clearer. This method effectively identifies the similarity between signals and optimizes signal analysis through clustering results. Dynamic updates during the clustering process and precise calculation of cluster centers improve the accuracy of signal classification and ensure the reliability of signal clusters. The interference signal identification and suppression strategy significantly enhances the system's robustness in complex environments. By identifying signals that do not match the spatiotemporal signal cluster features and effectively distinguishing interference signals through threshold settings, the system can remove noise components and retain valid signals, thereby improving signal quality and accuracy. The final generated interference-suppressed signal set provides clearer signal data for subsequent device type identification and operational status assessment.

[0043] Through the above optimizations, the method of this invention can stably and efficiently perform device detection and signal classification in multi-device and complex interference environments. Compared with traditional methods, this method has stronger adaptability and accuracy, and can provide high-quality device detection results in dynamically changing wireless signal environments, thus having broad application prospects.

[0044] In this embodiment, step seven specifically includes: Based on the aforementioned interference suppression signal set, information such as the signal transmission mode, spectral bandwidth, and signal strength is first analyzed. These signal characteristics reflect the signal's operating state, transmission method, and communication requirements. For example, the signal transmission mode can characterize whether it is unidirectional transmission, bidirectional communication, or multi-point broadcasting; the spectral bandwidth directly affects the signal's transmission speed and coverage; and the signal strength can be used to estimate the device's transmit power and operating range. This information can then be matched against a predefined device type library to determine the device type to which the signal source belongs. The predefined device type library can include the communication characteristics of various common devices, such as wireless base stations, sensor nodes, Wi-Fi devices, and Bluetooth devices. The matching of signal sources and device types is based on a comprehensive analysis of the signal's spectrum, transmission characteristics, and other communication parameters, ultimately yielding a device type identification result. This result indicates the device type corresponding to each signal source, such as a base station, sensor, or router.

[0045] Based on the frequency and amplitude fluctuation information of the interference suppression signal, the operating status of the equipment is further evaluated. The operating status of the equipment can usually be determined by the changes in the signal in the time and frequency domains. Frequency changes and amplitude fluctuations reflect whether the equipment is within its normal operating range and whether there is signal distortion or other abnormalities. For example, if the signal frequency changes significantly, it may indicate a equipment malfunction or a switch to an abnormal operating mode; if the amplitude fluctuation is too large, it may indicate that the equipment is in an unstable state or is experiencing external interference. By analyzing the frequency and amplitude fluctuations of the signal, the operating status of the equipment can be evaluated to determine whether the equipment is in a normal operating state or whether there is an abnormal operating state. This analysis generates an equipment operating status evaluation result, indicating whether the equipment is currently operating normally, in standby mode, or experiencing a malfunction.

[0046] Based on the equipment type identification and operational status assessment results, interference sources are further identified using interference information within the signal. Interference source identification primarily involves analyzing the signal's spectral characteristics, particularly identifying portions of the signal that deviate from the expected frequency range and intensity distribution. Interference signals may originate from signal leakage from other equipment, electromagnetic interference, or external environmental influences. By detecting these anomalous signal characteristics, the location and type of interference source can be identified; for example, whether a specific frequency band is occupied by other unrelated equipment or whether there is spectral interference between devices. This method generates interference source identification results, which pinpoint the interference source within the signal and provide its location and type information, facilitating further optimization of signal transmission and equipment scheduling.

[0047] The device type identification results, device operating status assessment results, and interference source identification results are combined to generate the final detection result. This final detection result includes the device type, operating status, signal characteristics, and possible interference sources. Through this comprehensive analysis, the operating status, type, and performance of wireless signal devices during signal transmission can be fully evaluated, providing accurate basis for subsequent fault diagnosis, performance optimization, and network management. The final detection result not only supports the monitoring of wireless devices but can also be used to dynamically adjust device operating parameters, optimize device operating efficiency and signal transmission quality, and improve the overall performance and reliability of the system.

[0048] refer to Figure 2 A deep learning-based wireless signal device detection system includes the following modules: The data acquisition and preprocessing module is used to acquire raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure. The time-frequency feature analysis module is used to extract the time-frequency features of the wireless signal data using fast Fourier transform and to perform multi-spectral analysis to generate a time-frequency feature vector for each frequency band. The signal correlation graph construction module is used to input the time-frequency feature vector of each frequency band into the ChebNet network and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; The spectrum analysis module is used to input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. The matching analysis module is used to extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; The spatiotemporal clustering module is used to perform spatiotemporal clustering on the matching feature vector set, identify the spatiotemporal relationship between signals, and perform interference suppression to generate an interference suppression signal set. The final detection module is used to analyze the device type and operating status of the wireless signal device based on the interference suppression signal set, and generate the final detection result.

[0049] Example 1: To verify the feasibility of this invention in practice, it was applied to an industrial wireless communication system. In this system, multiple wireless signal devices are used for data transmission, and the devices are interconnected and exchange data via wireless signals. However, due to factors such as electromagnetic interference in the environment, high device density, and signal overlap, the device signals are subject to varying degrees of interference. This makes it impossible for traditional methods to effectively identify and classify device types and operating states, thus affecting the stability and reliability of the system. Multiple devices communicate with each other using wireless communication protocols such as Wi-Fi and Zigbee. Due to the large number of devices and the complex signal frequency bands and interference environment, traditional methods cannot effectively distinguish the signals of each device, especially in cases of severe signal overlap or strong interference, where the device type and operating state are often inaccurately identified.

[0050] In this scenario, we apply the method of this invention to solve the aforementioned problems. First, raw signal data from wireless signal devices, including time-domain data, frequency-domain data, and environmental noise, is collected and preprocessed to generate wireless signal data with a unified structure. Next, the time-frequency features of the signal are extracted using Fast Fourier Transform (FFT), and multi-spectral analysis is performed to generate time-frequency feature vectors for each frequency band. Then, these time-frequency feature vectors are input into a ChebNet network, where a signal correlation graph is constructed using Chebyshev polynomial approximate convolution operations. The spectrum graph and signal correlation graph are then input into an improved PinSAGE model for spectrum band node construction, signal feature encoding, and multi-modal node association propagation, ultimately generating an optimized spectrum feature set.

[0051] By extracting the device features of each wireless signal device, combining them with a spectral feature set, calculating the similarity between devices and signals, and performing K-Means clustering analysis to identify the spatiotemporal relationships between signals and suppress interference. Finally, an interference-suppressed signal set is generated. Further analysis of these data reveals device types and operating states, accurately identifying devices and effectively filtering out interfering signals. In our experiments, we compared the performance of three traditional signal device detection methods with the method of this invention. Experimental data came from signal acquisition from 10 wireless devices within the park, covering multiple frequency bands including Wi-Fi and Zigbee. Under the same interference environment, the traditional methods achieved a device type identification accuracy of 65% and a device operating state identification accuracy of only 60%. In contrast, the method of this invention achieved a device type identification accuracy of 95%, a device operating state identification accuracy of 92%, and a significant improvement in the accuracy of interference source identification.

[0052] The following table compares the effectiveness of different methods in device identification and interference suppression: Table 1 Comparison of Device Identification and Interference Suppression Effects

[0053] As clearly shown in Table 1, the method of this invention significantly outperforms traditional methods in three aspects: device type identification, device operating status identification, and interference source identification. Traditional method 1 relies on simple signal strength and spectrum analysis, achieving an accuracy rate of 70% for device type identification, 65% for device operating status identification, and 60% for interference source identification. This method cannot effectively handle complex interference and signal overlap between multiple devices, resulting in low identification accuracy. Traditional method 2 uses support vector machines to classify signal features. Although it has made some progress in identifying device type and operating status, achieving an accuracy rate of 75% for device type identification and 68% for device operating status identification, the accuracy rate for interference source identification remains low at only 63% because the model fails to fully capture the spatiotemporal correlation of signals. Traditional method 3 is based on signal comparison and similarity analysis, achieving an accuracy rate of 72% for device type identification, 70% for device operating status identification, and 62% for interference source identification. This method mainly matches signals by calculating the similarity between them and device features, but its inability to handle complex relationships and spatiotemporal dynamic changes between signal sources leads to low accuracy in interference source identification. In contrast, the method of this invention, by introducing an improved PinSAGE model and deep learning technology, achieves significant improvements in signal feature extraction, signal correlation learning, and interference suppression. The accuracy rate for device type identification reaches 95%, the accuracy rate for device operating status identification is 92%, and the accuracy rate for interference source identification reaches 98%. By combining multiple advanced technologies, this invention effectively overcomes the limitations of traditional methods in complex environments, improves the accuracy and robustness of signal analysis, and can operate stably in complex interference environments, accurately identifying signal and interference sources. It has stronger practical application value and wider application prospects.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting wireless signal devices based on deep learning, characterized in that, include: Step 1: Collect raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure; Step 2: Use Fast Fourier Transform to extract the time-frequency features of the wireless signal data, and perform multi-spectral analysis to generate time-frequency feature vectors for each frequency band; Step 3: Input the time-frequency feature vector of each frequency band into the ChebNet network, and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; Step 4: Input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module, and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. Step 5: Extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; Step 6: Perform spatiotemporal clustering on the matching feature vector set to identify the spatiotemporal relationships between signals, and perform interference suppression to generate an interference suppression signal set; Step 7: Based on the interference suppression signal set, analyze the device type and operating status of the wireless signal device to generate the final detection result.

2. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, The raw signal data includes time-domain data, frequency-domain data, and environmental noise; the preprocessing steps specifically include: A bandpass filter is used to filter out low-frequency noise below a preset frequency in the time domain data of the original signal data, and the corresponding bandwidth is set according to the frequency band of the received signal. For the frequency domain data in the original signal data, a high-pass filter is used to filter out high-frequency interference signals with frequencies greater than a preset frequency.

3. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, Step two specifically involves: The wireless signal data is segmented into multiple wireless signal data segments using windowing processing. A fast Fourier transform is applied to each wireless signal data segment to obtain the frequency domain representation of the current wireless signal data segment. By calculating the amplitude spectrum and phase spectrum in the frequency domain, the amplitude and phase information of each frequency band are obtained, and a spectrum diagram is generated, which represents the energy distribution of the signal in different frequency bands; The spectrum in the spectrogram is divided into multiple frequency bands. The frequency range and energy distribution of each frequency band are analyzed to identify the frequency range in the signal and calculate the signal-to-noise ratio of each frequency band. Select the spectral data corresponding to the frequency bands whose frequency range is within the preset threshold and whose signal-to-noise ratio is greater than the preset threshold to obtain the filtered time-frequency features; By combining the frequency characteristics, amplitude, and phase information of each frequency band in the filtered time-frequency features, a time-frequency feature vector for each frequency band is generated.

4. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, Step three specifically involves: The time-frequency feature vector of each frequency band is input into the input layer of the ChebNet network, and the input layer uses the time-frequency feature vector of each frequency band as a node to obtain a node set; Calculate the Manhattan distance between every two nodes in the node set. If the Manhattan distance is less than a preset distance threshold, establish an edge connection between the corresponding nodes to obtain an edge connection set. Construct a time-frequency feature map based on the node set and edge connection set; The time-frequency feature map is input into the convolutional layer of the ChebNet network. Through Chebyshev polynomial approximate convolution operation, the aggregated features of the nodes are generated, and the aggregated feature map is obtained. The aggregated feature map is input into a nonlinear activation layer, and the ReLU nonlinear activation function is used to activate the features of each node, reducing negative features to zero while keeping positive features unchanged, thus generating a signal correlation map.

5. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, The improved PinSAGE model is specifically as follows: In the spectrum band node construction module, each frequency band in the spectrum graph is mapped to a spectrum band node. Each spectrum band node represents the feature data of a frequency band, generating a spectrum band node set. The spectrum band node includes the amplitude frequency domain features and phase frequency domain features of each spectrum band. The spectral band nodes are input into the signal feature encoding module, and a convolutional autoencoder is used to encode the input set of spectral band nodes. The convolutional autoencoder extracts local spectral features of the spectral band through multiple convolutional layers, and maps the local spectral features of each node into a low-dimensional embedding vector through a spectral reconstruction mechanism to generate a set of spectral feature vectors. The spectral reconstruction mechanism is as follows: The local spectral features of each node are selected from the spectrum diagram. By selecting multiple consecutive spectral bands, the amplitude and phase information are combined into a matrix to form the local spectral matrix of the corresponding node. Principal component analysis is used to transform the spectral matrix into a low-dimensional vector representation, thus obtaining the spectral eigenvectors. The spectral feature vector set is input to the multimodal node association and transmission module. During node association and transmission, information is transmitted by combining the spectral feature vector and the connection relationship in the signal association graph to generate a fused association feature set. The information transmission is based on the edge connection relationship in the signal association graph, and the spectral feature vector is propagated in the signal association graph. Each node receives the information of the neighboring nodes and updates it by weighted average in combination with its own feature vector. The fused correlation feature set is input into the spectrum selection module. Different fused correlation features in the fused correlation feature set are used as nodes, and the Euclidean distance between different nodes is calculated as the edge weight of the connection edge to obtain the fused correlation feature map. The spectral features corresponding to nodes whose edge weights in the fusion correlation feature graph are greater than a preset threshold are selected to obtain the final spectral feature set.

6. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, Step five specifically involves: Extract the device features of each wireless signal device to generate a device feature vector set. The device features include the device model, working status, device location, and transmitted signal strength. Calculate the cosine similarity between each device feature vector and spectrum feature vector in the device feature set and spectrum feature set, respectively. If the cosine similarity is greater than the preset similarity threshold, it is considered that there is a strong similarity between the signal and the device, and a corresponding matching feature vector set is generated.

7. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, Step six specifically involves: K vectors are randomly selected from the set of matching feature vectors as initial cluster centers, where K is a preset number of clusters; Calculate the Euclidean distance between each matching feature vector and the initial cluster center, and assign each vector to the initial cluster center with the closest Euclidean distance; Based on the mean of all matching feature vectors in each cluster, the position of the cluster center is recalculated, and the cluster center is iteratively updated until the change of the cluster center is less than the preset threshold or the maximum number of iterations is reached, thus obtaining the final K cluster centers. Based on the clustering results, the matching feature vector set is divided into K signal clusters, each signal cluster corresponding to a spatiotemporal signal relationship, thus generating a spatiotemporal signal cluster set; By analyzing the signals in each spatiotemporal signal cluster, interference signals are identified. The interference signals are those whose difference from the mean characteristic value of the spatiotemporal signal cluster is greater than a preset threshold, and whose signal strength is lower than a preset threshold or whose spectral characteristics are inconsistent. The identified interference signals are filtered out to generate an interference suppression signal set.

8. The method for detecting wireless signal devices based on deep learning according to claim 1, characterized in that, Step seven specifically involves: Based on the interference suppression signal set, the transmission mode, spectral bandwidth and signal strength information of the signal are analyzed and matched with the predefined device type to determine the device type to which the signal source belongs, and the final device type identification result is obtained. The device type identification result indicates the device type corresponding to each signal source. Based on the frequency changes and amplitude fluctuations of the interference suppression signal, the working status of the equipment is assessed to determine whether the equipment is in normal working condition or whether there is an abnormal working condition, and the equipment working status assessment result is generated. Based on the equipment type identification results and equipment working status assessment results, interference sources are identified using signal interference information, and the location and type of interference sources are determined, generating interference source identification results. The final detection result is generated by combining the device type identification result, the device working status assessment result, and the interference source identification result.

9. A deep learning-based wireless signal device detection system, comprising executing the deep learning-based wireless signal device detection method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire raw signal data from wireless signal devices and preprocess the raw signal data to generate wireless signal data with a unified structure. The time-frequency feature analysis module is used to extract the time-frequency features of the wireless signal data using fast Fourier transform and to perform multi-spectral analysis to generate a time-frequency feature vector for each frequency band. The signal correlation graph construction module is used to input the time-frequency feature vector of each frequency band into the ChebNet network and construct the signal correlation graph through Chebyshev polynomial approximate convolution operation; The spectrum analysis module is used to input the spectrum diagram and the signal correlation diagram into the improved PinSAGE model. The spectrum band node construction module, signal feature encoding module, multimodal node correlation and transfer module and spectrum selection module perform spectrum analysis to obtain the spectrum feature set. The matching analysis module is used to extract the device features of each wireless signal device and match them with the spectrum feature set to generate a matching feature vector set; The spatiotemporal clustering module is used to perform spatiotemporal clustering on the matching feature vector set, identify the spatiotemporal relationship between signals, and perform interference suppression to generate an interference suppression signal set. The final detection module is used to analyze the device type and operating status of the wireless signal device based on the interference suppression signal set, and generate the final detection result.

Images