Power equipment state monitoring method based on non-contact leakage current sensor

Abstract

The invention is suitable for the technical field of electrical equipment state monitoring, and provides an electrical equipment state monitoring method based on a non-contact leakage current sensor, and the method comprises the steps: collecting a leakage current signal of the surface of an insulating part of electrical equipment, and obtaining infrared thermal image data and an ultrasonic signal; variational mode decomposition is carried out on the leakage current signal to obtain a plurality of intrinsic mode functions; separating a leakage current effective component and an independent noise source based on a blind source separation algorithm; extracting a time-frequency characteristic of the effective component of the separated leakage current, extracting a local temperature gradient characteristic of the infrared thermal image data and a frequency spectrum energy characteristic of the ultrasonic signal, and generating a multi-modal characteristic vector; analyzing the space-time relevance of the multi-modal feature vector, dynamically distributing each modal weight coefficient for fusion, and generating a comprehensive fault feature; according to the method, the signal-to-noise ratio of the weak leakage current signal is improved, and the misjudgment rate is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring, and in particular to a power equipment status monitoring method based on a non-contact leakage current sensor. Background Art

[0002] The safe and stable operation of power equipment is central to ensuring grid reliability, and condition monitoring of its insulation components is particularly critical. Currently, non-contact monitoring technology has become a research hotspot, thanks to its advantages of requiring no contact with equipment and long-term deployment. Among these, monitoring solutions based on leakage current sensors can effectively identify potential faults such as partial discharge and insulation aging by collecting leakage current signals from the insulation surface. Current technologies have achieved multimodal data fusion, combining auxiliary information such as infrared thermal imaging and ultrasonic signals to improve fault identification accuracy. Signal processing techniques such as variational mode decomposition and blind source separation are also used to suppress power frequency noise and extract effective features.

[0003] Conventional signal feature extraction methods struggle to adaptively capture the multi-scale time-frequency characteristics of partial discharge signals in leakage currents, especially in the presence of strong power-frequency noise and environmental interference, which can easily lead to loss of valid features or misjudgment of noise. Furthermore, multimodal data fusion employs a static weighting strategy that ignores the dynamic coupling of leakage current, infrared thermal imaging, and ultrasonic signals in time and space. This results in insufficient characterization of comprehensive fault features, resulting in incomplete and potentially inaccurate monitoring of equipment operating status. Summary of the Invention

[0004] The present application provides a method for monitoring the status of electric power equipment based on a non-contact leakage current sensor, which is used to solve the problem of incomplete detection and large errors.

[0005] The present application provides a method for monitoring the state of electric power equipment based on a non-contact leakage current sensor, comprising:

[0006] The non-contact leakage current sensor collects leakage current signals from the surface of insulating components of power equipment, and acquires infrared thermal imaging data and ultrasonic signals to form a multimodal monitoring data set.

[0007] Performing variational modal decomposition on the leakage current signal, and obtaining a plurality of intrinsic modal functions by optimizing the bandwidth and center frequency of each modal to separate the power frequency noise and the preset frequency component of partial discharge; and separating the effective component of the leakage current and the independent noise source from the intrinsic modal functions based on a blind source separation algorithm;

[0008] Extract the time-frequency characteristics of the effective component of the leakage current after separation, and extract the local temperature gradient characteristics of the infrared thermal imaging data and the spectral energy characteristics of the ultrasonic signal to generate a multimodal feature vector;

[0009] Utilizing the attention mechanism to analyze the spatiotemporal correlation of the multimodal feature vectors, dynamically assigning weight coefficients to each modality for fusion, and generating comprehensive fault features;

[0010] Based on the comprehensive fault feature vector and the preset edge computing device, the insulating component is analyzed to obtain the target defect location, and the defect location information and early warning signal are output to the monitoring terminal in real time.

[0011] Furthermore, the leakage current signal is subjected to variational modal decomposition, and multiple natural mode functions are obtained by separating the power frequency noise and the preset frequency components of partial discharge by optimizing the bandwidth and center frequency of each modal. The functions include:

[0012] A variational model is constructed to decompose the leakage current signal into multiple narrowband intrinsic mode functions by constraining the bandwidth and center frequency of each intrinsic mode function.

[0013] Determine the number of decomposed modes based on the spectrum characteristics of the leakage current signal and initialize the center frequency of each mode;

[0014] The bandwidth constraint parameters and center frequency of each mode are adjusted through an iterative optimization algorithm, so that the modal center frequency corresponding to the power frequency noise converges to the power frequency range, and the modal center frequency corresponding to the high-frequency component of partial discharge converges to the preset high-frequency range.

[0015] Furthermore, the separation of the effective component of the leakage current and the independent noise source from the intrinsic mode function based on a blind source separation algorithm includes:

[0016] Environmental noise reference signals are collected in real time by deploying environmental noise sensors around power equipment;

[0017] Constructing a hybrid measurement matrix using the ambient noise reference signal and the intrinsic mode function, and solving the hybrid measurement matrix based on an independent component analysis algorithm to separate statistically independent source signals;

[0018] The effective components of the leakage current are screened according to the energy distribution characteristics and frequency characteristics of the source signal, and the noise source components that meet the correlation threshold with the environmental noise reference signal are eliminated.

[0019] Furthermore, the extraction of the time-frequency characteristics of the effective component of the leakage current after separation includes:

[0020] The effective component of the leakage current is input into a preset multi-branch convolutional neural network; wherein the first branch uses a one-dimensional convolution kernel to extract time series features to capture the periodic changes of the leakage current; the second branch uses a two-dimensional convolution kernel to process the time-frequency spectrum, and uses an adaptive long strip convolution kernel to extract frequency band change features vertically and time continuity features horizontally;

[0021] A learnable frequency domain constraint layer is embedded in the multi-branch convolutional neural network, and convolution kernel parameters are initialized to wavelet basis functions of a preset partial discharge frequency band to enhance feature extraction of high-frequency discharge components;

[0022] The output weights of different convolution channels are dynamically activated through the gating mechanism, and the features that meet the preset correlation threshold with the time-frequency characteristics of the current signal are screened to generate the time-frequency feature vector that characterizes the partial discharge behavior.

[0023] Furthermore, the extraction of the time-frequency characteristics of the effective component of the leakage current after separation, and the extraction of the local temperature gradient characteristics of the infrared thermal imaging data and the spectrum energy characteristics of the ultrasonic signal to generate a multimodal feature vector includes:

[0024] Performing channel attention weighting on the time-frequency characteristics of the effective component of the leakage current to suppress power frequency noise interference and enhance the significant characteristics of the high-frequency discharge pulse;

[0025] Performing spatial attention modeling on the local temperature gradient characteristics of the infrared thermal imaging data to extract the statistical distribution characteristics of the temperature gradient direction and amplitude of the hotspot area and eliminate background thermal interference;

[0026] Performing multi-scale decomposition on the spectrum energy characteristics of the ultrasonic signal, extracting the energy fluctuation characteristics within a preset frequency band, and quantifying the coupling relationship between mechanical vibration and discharge;

[0027] The weighted time-frequency features, temperature gradient features, and spectrum energy features are concatenated according to a preset dimension to generate a preliminary multimodal feature vector.

[0028] Furthermore, the attention mechanism is used to analyze the spatiotemporal correlation of the multimodal feature vectors, dynamically assign weight coefficients to each modality for fusion, and generate comprehensive fault features, including:

[0029] The cross-modal correlation weights between electrical, thermal, and mechanical modalities are calculated through a preset cross-modal spatiotemporal attention network.

[0030] The weight coefficient of each modal is adjusted according to the real-time operating conditions and the probability distribution of fault types, and comprehensive fault characteristics are generated through weighted fusion.

[0031] Furthermore, the preset cross-modal spatiotemporal attention network includes:

[0032] For electrical modal features, the temporal attention module is used to capture the temporal clustering of discharge pulses, and the frequency domain attention module is used to focus on the energy distribution in the high-frequency band.

[0033] For thermal modal features, the spatial attention module is used to locate temperature anomaly areas, and the temporal attention module is used to analyze temperature evolution trends.

[0034] For mechanical modal features, the frequency points with energy surges are identified through the spectrum energy attention module.

[0035] Furthermore, the cross-modal correlation weights among the electrical modality, the thermal modality, and the mechanical modality are calculated by using a preset cross-modal spatiotemporal attention network, including:

[0036] The electrical modal features, thermal modal features, and mechanical modal features are grouped by sensor type, and the attention weights within the groups and between key groups are calculated;

[0037] Introducing causal convolutional layers into temporal attention calculations to constrain weight distribution to rely solely on historical temporal information;

[0038] Through lightweight matrix compression technology, the weight matrix of the fully connected layer is decomposed into a low-rank matrix product, and the attention weight is binarized to retain only the correlation features of the key modalities.

[0039] Furthermore, the method of using the attention mechanism to analyze the spatiotemporal correlation of the multimodal feature vectors, dynamically assigning weight coefficients to each modality for fusion, and generating a comprehensive fault feature also includes:

[0040] Performing spatiotemporal alignment on the high-frequency pulse amplitude of the electrical mode, the local temperature gradient amplitude of the thermal mode, and the frequency points of the spectral energy surge of the mechanical mode, and assigning a weight coefficient to each mode based on the correlation characteristics;

[0041] The aligned multimodal features are fused to generate a comprehensive fault feature vector that is adapted to the edge computing model input.

[0042] Furthermore, the analyzing the insulating component based on the comprehensive fault feature vector and a preset edge computing device to obtain the target defect location includes:

[0043] Deploying a lightweight diagnostic model in the edge computing device, inputting the comprehensive fault characteristics into the lightweight diagnostic model to predict defect probability, and outputting a defect probability distribution map of each area of ​​the insulation component;

[0044] Based on the defect probability distribution map and a preset spatial mapping relationship between the insulating components, determining an area whose probability exceeds a dynamic threshold as a candidate defect area;

[0045] The candidate defect areas are matched using the comprehensive feature vector to obtain a target defect location.

[0046] It can be seen from the above technical solutions that this application has the following advantages:

[0047] This application uses non-contact leakage current sensors, infrared thermal imagers, and ultrasonic sensors to synchronously collect multimodal data. It uses variational mode decomposition and blind source separation techniques to separate power frequency noise and high-frequency discharge components. It combines the attention mechanism to dynamically fuse the spatiotemporal correlation characteristics of multimodal data to generate highly discriminative comprehensive fault features. It deploys lightweight diagnostic models on edge computing devices to accurately locate insulation defects at the component level and output defect locations and warning signals in real time. By reducing the reliance on contact for power outage detection, it improves the signal-to-noise ratio of weak leakage current signals, reduces the misjudgment rate of monitoring, and supports online real-time monitoring of high-voltage equipment, providing efficient and reliable data support for predictive maintenance of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The figure is a flow chart of an embodiment of a method for monitoring the state of electric equipment based on a non-contact leakage current sensor in the present invention. DETAILED DESCRIPTION

[0049] The terms "first," "second," "third," "fourth," etc. (if any) in the specification and claims of the present application and in the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to," and any variations thereof, are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or apparatus.

[0050] Example 1

[0051] The implementation method in this embodiment can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, without specific limitation. The following describes the power equipment status monitoring method based on the non-contact leakage current sensor in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application includes the following steps:

[0052] S11. Collect leakage current signals from the surface of insulating components of power equipment using a non-contact leakage current sensor, and obtain infrared thermal imaging data and ultrasonic signals to form a multimodal monitoring data set;

[0053] In this embodiment, the non-contact leakage current sensor may include a capacitive coupling or electromagnetic induction sensor, and the insulating component may include an insulator or bushing. The non-contact leakage current sensor non-invasively collects leakage current signals from the surface of the insulating component, avoiding equipment power outages or circuit modifications. An infrared thermal imager captures temperature distribution data on the device surface, and thermal radiation imaging technology is used to identify areas of local temperature anomalies. An ultrasonic sensor array is also deployed to collect high-frequency acoustic signals generated by partial discharge or mechanical vibration during device operation.

[0054] All sensor data is time-stamped and aligned using a unified timing synchronization module. Spatial matching is performed based on device spatial coordinates, such as insulator segment number and bushing installation location, to form a structured dataset containing multimodal electrical, thermal, and mechanical information. This dataset is stored in a standardized format, providing multi-dimensional collaborative input for subsequent signal processing and fault diagnosis.

[0055] S12. Perform variational modal decomposition on the leakage current signal, separating the power frequency noise and the preset frequency components of partial discharge by optimizing the bandwidth and center frequency of each modal, to obtain multiple intrinsic modal functions; based on a blind source separation algorithm combined with an ambient noise reference signal, separate the effective leakage current component and independent noise sources from the intrinsic modal functions;

[0056] In this embodiment, variational modal decomposition is performed on the leakage current signal, and multiple natural mode functions are obtained by separating the power frequency noise and the preset frequency components of partial discharge by optimizing the bandwidth and center frequency of each mode. The steps include:

[0057] This step constructs a variational model to adaptively decompose the leakage current signal into multiple narrowband intrinsic mode functions (IMFs), accurately separate the power frequency noise and the high-frequency components of partial discharge, and solve the modal aliasing problem of traditional empirical mode decomposition (EMD).

[0058] 1. Construct a variational model to decompose the leakage current signal into multiple narrowband intrinsic mode functions by constraining the bandwidth and center frequency of each intrinsic mode function;

[0059] The objective function of the variational model is defined as minimizing the sum of the bandwidths of all IMFs while constraining the synthetic signal of each IMF to be consistent with the original signal. The mathematical model is:

[0060]

[0061] Where: u k is the kth IMF, ω k is its center frequency, K is the number of modes; constraints: s(t) is the original leakage current signal.

[0062] 2. Determine the number of decomposed modes based on the spectral characteristics of the leakage current signal and initialize the center frequency of each mode;

[0063] Based on the spectrum characteristics of the leakage current signal (such as the number of main peaks and frequency band distribution), the number of decomposition layers K is adaptively determined by the energy accumulation method; the center frequency w of each IMF is initialized. k , usually sorted from low to high by spectrum amplitude, such as the power frequency 50Hz corresponds to the low-frequency mode, and the high-frequency component of partial discharge corresponds to the high-frequency mode.

[0064] 3. The bandwidth constraint parameters and center frequency of each mode are adjusted through an iterative optimization algorithm so that the modal center frequency corresponding to the power frequency noise converges to the power frequency range, and the modal center frequency corresponding to the high-frequency component of partial discharge converges to the preset high-frequency range.

[0065] Iteratively update u using the alternating direction method of multipliers (ADMM) k With w k , dynamically adjusting the bandwidth constraint parameters to control the modal bandwidth. Convergence condition: Optimization is terminated when the change in the IMF center frequency between adjacent iterations is less than a preset threshold, which is flexibly set based on the convergence scenario. Output: Power frequency noise (IM) is output when the center frequency F converges to around 50Hz, and partial discharge (IMF) is output when the center frequency converges to a preset high-frequency range (10kHz-1MHz).

[0066] In this embodiment, separating the effective component of the leakage current and the independent noise source from the intrinsic mode function based on a blind source separation algorithm includes the following steps:

[0067] This step uses the independent component analysis (ICA) algorithm, combined with the environmental noise reference signal, to separate the effective leakage current component and independent noise sources from the IMF after VMD decomposition, significantly improving the signal-to-noise ratio.

[0068] 1. Use environmental noise sensors deployed around power equipment to collect environmental noise reference signals in real time;

[0069] Noise sensors, such as electromagnetic compatibility antennas or acoustic sensors, are deployed around power equipment to collect real-time environmental noise signals, such as corona discharge and radio frequency interference. The noise signals are timestamped with the IMF decomposed by VMD to ensure data synchronization.

[0070] 2. Construct a hybrid measurement matrix by combining the ambient noise reference signal and the intrinsic mode function. Then, solve the hybrid measurement matrix based on the independent component analysis algorithm to separate the statistically independent source signals.

[0071] The VMD-decomposed IMF (as the observed signal X) is concatenated with the ambient noise reference signal (as the reference signal N) by column to construct a mixing matrix Z = [X, N]. Assume that the observed signal is a linear mixture of independent source signals: Z = A·S, where A is the mixing matrix and S is the independent source signal.

[0072] The FastICA algorithm is used to maximize the non-Gaussianity of the source signal and solve the unmixing matrix W (W = A ―1 ), separating statistically independent source signals Centering and whitening processing: eliminate the signal mean and decorrelation; iterative optimization: calculate the independent components through negative entropy approximation and update the unmixing vector until convergence.

[0073] 3. Filter the effective components of the leakage current according to the energy distribution characteristics and frequency characteristics of the source signal, and eliminate the noise source components that meet the correlation threshold with the ambient noise reference signal.

[0074] Energy distribution screening: Calculates the energy contribution of each source signal and retains components with significantly higher energy than background noise (e.g., energy contribution >80%). Frequency characteristics screening: Analyzes the source signal spectrum through FFT and retains components within the preset 10kHz-1MHz high-frequency range as effective leakage current components. Correlation rejection: Calculates the cross-correlation between the source signal and the ambient noise reference signal. If the correlation coefficient exceeds a set threshold, it is determined to be a noise component and rejected. Here, the threshold is the coefficient corresponding to high correlation.

[0075] S13. Extracting the time-frequency characteristics of the effective component of the leakage current after separation, and extracting the local temperature gradient characteristics of the infrared thermal imaging data and the spectral energy characteristics of the ultrasonic signal to generate a multimodal feature vector;

[0076] In this embodiment, extracting the time-frequency characteristics of the effective component of the leakage current after separation includes the following steps:

[0077] 1. The effective component of the leakage current is input into a preset multi-branch convolutional neural network. The first branch uses a one-dimensional convolution kernel to extract time series features, capturing the periodic changes of the leakage current. The second branch uses a two-dimensional convolution kernel to process the time-frequency spectrum, using an adaptive long strip convolution kernel to extract frequency band variation features vertically and temporal continuity features horizontally.

[0078] The first branch extracts temporal features, using a one-dimensional convolution kernel (e.g., length = 64, stride = 8) sliding along the time axis to capture periodic variations in the leakage current (e.g., the repetition frequency of partial discharge pulses). A ReLU activation function is used, and a maximum pooling layer (pooling size = 4) compresses the temporal dimension to output a temporal feature vector. The second branch extracts time-frequency features, performing a time-frequency transformation on the leakage current signal to generate a time-frequency spectrum (a two-dimensional time-frequency matrix). Using a two-dimensional convolution kernel (e.g., a 3×5 strip kernel), the kernel extracts frequency band variation features (e.g., high-frequency energy concentration) vertically (frequency axis) and temporal continuity features (e.g., pulse duration) horizontally (time axis). Adaptive kernel shape adjustment (e.g., dynamic adjustment of the kernel aspect ratio) is used to adapt to the time-frequency distribution characteristics of different discharge patterns.

[0079] 2. Embed a learnable frequency-domain constraint layer in the multi-branch convolutional neural network, initialize the convolution kernel parameters to the wavelet basis function of the preset partial discharge frequency band, and enhance the feature extraction of high-frequency discharge components;

[0080] Initializing the convolution kernel parameters to a wavelet basis function for a preset partial discharge frequency band (e.g., 10kHz-1MHz) forces the network to focus on the target frequency band. During training, the convolution kernel parameters are fine-tuned through backpropagation to balance prior knowledge with data-driven learning capabilities.

[0081] 3. Dynamically activate the output weights of different convolution channels through a gating mechanism, screen features that meet the preset correlation threshold with the time-frequency characteristics of the current signal, and generate a time-frequency feature vector that characterizes the partial discharge behavior.

[0082] A gating unit (such as the dynamic weight generator of a GRU) is introduced after the output of the convolutional channel to calculate the channel activation weight based on the time-frequency characteristics of the current input signal (such as pulse density and energy fluctuation). Only channel features with weights higher than a preset threshold are retained, suppressing interference from irrelevant or low signal-to-noise ratio features. The output result is a time-frequency feature vector (dimension = 128) that characterizes the partial discharge behavior.

[0083] In this embodiment, generating a multimodal feature vector includes the following steps:

[0084] 1. Perform channel attention weighting on the time-frequency characteristics of the effective component of the leakage current to suppress power frequency noise interference and enhance the significant characteristics of high-frequency discharge pulses;

[0085] A channel attention module is used to calculate the global average pooling value of each channel feature and generate a channel weight vector. Channels corresponding to high-frequency discharge pulses (e.g., center frequency > 10kHz) are given higher weights (weight coefficient = 0.8-1.0) and power-frequency noise channels are suppressed (weight coefficient = 0.2-0.5). The weighted time-frequency feature vector (dimension = 128) is output.

[0086] 2. Perform spatial attention modeling on the local temperature gradient characteristics of infrared thermal imaging data to extract the statistical distribution characteristics of the temperature gradient direction and amplitude in the hotspot area and eliminate background thermal interference;

[0087] The infrared thermal image is spatially divided into blocks (e.g., an 8×8 grid), and the temperature gradient amplitude and direction of each area are calculated. A spatial attention map is constructed: regions with larger temperature gradient amplitudes and more consistent directions (e.g., hotspots) are assigned higher attention weights (weight coefficient = 0.7-1.0). The thermal image feature vector (dimension = 64) is output to characterize the spatial distribution characteristics of temperature anomalies.

[0088] 3. Perform multi-scale decomposition of the spectral energy characteristics of the ultrasonic signal, extract the energy fluctuation characteristics within the preset frequency band, and quantify the coupling relationship between mechanical vibration and discharge;

[0089] Perform multiscale wavelet packet decomposition (e.g., 3-layer decomposition) on the ultrasonic signal to extract the energy contribution of different frequency bands (e.g., low-frequency mechanical vibration band 0-20kHz, high-frequency discharge band 20kHz-1MHz). Calculate the energy fluctuation characteristics (e.g., variance, peak-to-peak value) of each frequency band to quantify the coupling relationship between mechanical vibration and discharge. Output the ultrasonic feature vector (dimension = 32).

[0090] 4. Concatenate the weighted time-frequency features, temperature gradient features, and spectral energy features according to the preset dimensions to generate a preliminary multimodal feature vector.

[0091] The weighted time-frequency features (128 dimensions), thermal imaging features (64 dimensions), and ultrasonic features (32 dimensions) are aligned by dimension (e.g., zero-padding or dimensionality reduction to a uniform length); they are concatenated into a multimodal feature vector (total dimension = 224) in a preset dimensional order (electrical → thermal → mechanical); and feature compression and dimensionality reduction are performed through a lightweight fusion network (e.g., a 1×1 convolutional layer) to output the final multimodal feature vector (dimension = 64).

[0092] The above steps achieve efficient extraction of leakage current time-frequency features and structured fusion of multimodal features through the collaborative design of multi-branch convolutional neural networks, attention mechanisms and multi-scale analysis, solving the problems of feature redundancy and modal mismatch in traditional methods, and providing a reliable data basis for the accurate diagnosis of insulation defects in power equipment.

[0093] S14. Use the attention mechanism to analyze the spatiotemporal correlation of multimodal feature vectors, dynamically assign weight coefficients to each modality for fusion, and generate comprehensive fault features;

[0094] S141. Analyze the spatiotemporal correlation of the multimodal feature vectors using the attention mechanism, dynamically assign weight coefficients to each modality for fusion, and generate a comprehensive fault feature, including the following steps:

[0095] 1. Calculate the cross-modal correlation weights between the electrical, thermal, and mechanical modalities through a pre-set cross-modal spatiotemporal attention network;

[0096] The pre-configured attention network here models the spatiotemporal characteristics of the electrical, thermal, and mechanical modes, respectively. The electrical mode uses a temporal attention module to analyze the temporal clustering of discharge pulses, such as periodic discharge intervals, and a frequency-domain attention module to focus on high-frequency energy distribution, such as the 10kHz-1MHz frequency band. The thermal mode uses a spatial attention module to locate areas of temperature anomalies, such as hot spots on the surface of insulators, and a temporal attention module to analyze temperature trends, such as the rate of temperature rise. The mechanical mode uses a spectral energy attention module to identify frequency points with sudden energy increases, such as high-frequency energy peaks in the ultrasonic spectrum.

[0097] 2. Adjust the weight coefficient of each modal according to the real-time operating conditions and the probability distribution of the fault type, and generate comprehensive fault characteristics through weighted fusion.

[0098] Based on real-time operating conditions and fault type classification results, including partial discharge, overheating, and mechanical damage, the weight coefficients of each modality are dynamically adjusted. For example, when partial discharge is detected, the weight of the electrical modality is increased; when the ambient temperature rises, the weight of the thermal modality is increased. The adjusted weight coefficients are multiplied by the characteristics of each modality, and a weighted sum is used to generate a comprehensive fault signature, strengthening the spatiotemporal correlation between the electrical, thermal, and mechanical modalities.

[0099] S142. Preset cross-modal spatiotemporal attention network, including:

[0100] For electrical modal features, the temporal attention module is used to capture the temporal clustering of discharge pulses, and the frequency domain attention module is used to focus on the energy distribution in the high-frequency band.

[0101] For thermal modal features, the spatial attention module is used to locate temperature anomaly areas, and the temporal attention module is used to analyze temperature evolution trends.

[0102] For mechanical modal features, the frequency points with energy surges are identified through the spectrum energy attention module.

[0103] The specific implementation of the cross-modal spatiotemporal attention network involves the differentiated design of attention modules for each modality. For the electrical modality, the temporal attention module analyzes pulse timing patterns using a sliding window, while the frequency-domain attention module enhances high-frequency components using band masks. For the thermal modality, the spatial attention module locates hotspots using a convolutional network, while the temporal attention module analyzes temperature evolution through causal convolution. For the mechanical modality, the spectral energy attention module uses energy thresholds to filter key frequencies.

[0104] S143. Calculating cross-modal correlation weights between the electrical modality, the thermal modality, and the mechanical modality using a preset cross-modal spatiotemporal attention network, including:

[0105] The electrical modal features, thermal modal features, and mechanical modal features are grouped by sensor type, and the attention weights within the groups and between key groups are calculated;

[0106] Introducing causal convolutional layers into temporal attention calculations to constrain weight distribution to rely solely on historical temporal information;

[0107] Through lightweight matrix compression technology, the weight matrix of the fully connected layer is decomposed into a low-rank matrix product, and the attention weight is binarized to retain only the correlation features of the key modalities.

[0108] Specifically, grouped attention calculation: the electrical, thermal, and mechanical modes are grouped according to the sensor type, and only the attention weights within the group (such as electrical-electrical) and between key groups (such as electrical-thermal, electrical-mechanical) are calculated to reduce the amount of calculation; causal convolution constraint: a causal convolution layer is introduced in the temporal attention module to ensure that the weight distribution depends only on historical data to avoid future information interfering with real-time monitoring; lightweight and binarization: the weight matrix of the fully connected layer is decomposed into a low-rank matrix to reduce the number of parameters; continuous weights are learned in the training phase, and are binarized to 0 / 1 through the threshold during deployment, retaining only key related features.

[0109] S144. Analyzing the spatiotemporal correlation of the multimodal feature vectors using an attention mechanism, dynamically assigning weight coefficients to each modality for fusion, and generating a comprehensive fault feature, further comprising:

[0110] 1. Perform spatiotemporal alignment on the high-frequency pulse amplitude of the electrical mode, the local temperature gradient amplitude of the thermal mode, and the frequency points of the spectral energy surge of the mechanical mode, and assign weight coefficients to each mode based on their correlation characteristics;

[0111] 2. The aligned multimodal features are fused to generate a comprehensive fault feature vector that is adapted to the edge computing model input.

[0112] Specifically, time and space alignment: align the timestamps of the high-frequency pulse amplitude of the electrical mode, the local temperature gradient amplitude of the thermal mode, map the spatial coordinates to the insulator sheet number, and perform time synchronization alignment on the frequency points of the spectral energy surge of the mechanical mode to ensure the consistency of multimodal features in the time and space dimensions; weight allocation and fusion: according to the cross-modal correlation weights, such as electric-thermal weight = 0.6, electric-mechanical weight = 0.3, perform weighted summation on the aligned features to generate a comprehensive fault feature vector; edge computing adaptation: reduce the dimensionality and standardize the fused features, and output 64-dimensional low-dimensional features that adapt to the input of the lightweight diagnostic model to meet the real-time requirements of edge devices.

[0113] S15. Analyze the insulation components based on the comprehensive fault characteristics and the preset edge computing device to obtain the target defect location, and output the defect location information and warning signals to the monitoring terminal in real time.

[0114] This implementation includes the following steps:

[0115] 1. Deploy a lightweight diagnostic model in the edge computing device, input comprehensive fault characteristics into the model to predict defect probability, and output a defect probability distribution map for each area of ​​the insulation component;

[0116] 2. Based on the defect probability distribution map and the preset spatial mapping relationship between the insulating components, determine the areas where the probability exceeds the dynamic threshold as candidate defect areas;

[0117] 3. Match the candidate defect areas through comprehensive feature vectors to obtain the target defect location.

[0118] Through a lightweight diagnostic model deployed on edge computing devices, here the diagnostic model is a pruned convolutional neural network, which predicts the defect probability of the comprehensive fault characteristics and generates a defect probability distribution map for each area of ​​the insulating component. Based on the preset spatial mapping relationship of the insulating components, which includes the preset mapping relationship between the insulator piece number and the bushing geometric coordinates, the probability threshold is dynamically adjusted to screen out candidate defect areas with excessive probability. Further multimodal feature consistency verification is carried out, in which the candidate areas are required to simultaneously meet the requirements of high-frequency pulse increase of leakage current, abnormal local temperature rise of infrared thermal images, and sudden energy mutation of ultrasonic spectrum. Combined with the spatiotemporal correlation matching of timestamp synchronization and spatial coordinate mapping, the target defect position is located. Finally, the defect location information and warning signals are transmitted to the monitoring terminal in real time through a low-latency communication protocol, supporting rapid response of operation and maintenance personnel.

[0119] It is understandable that those skilled in the art can, under the guidance of the above embodiments, combine various implementation methods in the above embodiments to obtain technical solutions of multiple implementation methods.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for monitoring the state of electric equipment based on a non-contact leakage current sensor, characterized in that: include: The non-contact leakage current sensor collects leakage current signals from the surface of insulating components of power equipment, and acquires infrared thermal imaging data and ultrasonic signals to form a multimodal monitoring data set. Performing variational modal decomposition on the leakage current signal, and obtaining a plurality of intrinsic modal functions by optimizing the bandwidth and center frequency of each modal to separate the power frequency noise and the preset frequency component of partial discharge; and separating the effective component of the leakage current and the independent noise source from the intrinsic modal functions based on a blind source separation algorithm; Extract the time-frequency characteristics of the effective component of the leakage current after separation, and extract the local temperature gradient characteristics of the infrared thermal imaging data and the spectral energy characteristics of the ultrasonic signal to generate a multimodal feature vector; Utilizing the attention mechanism to analyze the spatiotemporal correlation of the multimodal feature vectors, dynamically assigning weight coefficients to each modality for fusion, and generating comprehensive fault features; Based on the comprehensive fault feature vector and the preset edge computing device, the insulating component is analyzed to obtain the target defect location, and the defect location information and early warning signal are output to the monitoring terminal in real time.

2. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 1, characterized in that: The variational modal decomposition is performed on the leakage current signal, and multiple natural mode functions are obtained by separating the power frequency noise and the preset frequency components of the partial discharge by optimizing the bandwidth and center frequency of each mode, including: A variational model is constructed to decompose the leakage current signal into multiple narrowband intrinsic mode functions by constraining the bandwidth and center frequency of each intrinsic mode function. Determine the number of decomposed modes based on the spectrum characteristics of the leakage current signal and initialize the center frequency of each mode; The bandwidth constraint parameters and center frequency of each mode are adjusted through an iterative optimization algorithm, so that the modal center frequency corresponding to the power frequency noise converges to the power frequency range, and the modal center frequency corresponding to the high-frequency component of partial discharge converges to the preset high-frequency range.

3. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 2, characterized in that: The separation of the effective component of the leakage current and the independent noise source from the intrinsic mode function based on the blind source separation algorithm includes: Environmental noise reference signals are collected in real time by deploying environmental noise sensors around power equipment; Constructing a hybrid measurement matrix using the ambient noise reference signal and the intrinsic mode function, and solving the hybrid measurement matrix based on an independent component analysis algorithm to separate statistically independent source signals; The effective components of the leakage current are screened according to the energy distribution characteristics and frequency characteristics of the source signal, and the noise source components that meet the correlation threshold with the environmental noise reference signal are eliminated.

4. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 1, characterized in that: The step of extracting the time-frequency characteristics of the effective component of the leakage current after separation includes: The effective component of the leakage current is input into a preset multi-branch convolutional neural network; wherein the first branch uses a one-dimensional convolution kernel to extract time series features to capture the periodic changes of the leakage current; the second branch uses a two-dimensional convolution kernel to process the time-frequency spectrum, and uses an adaptive long strip convolution kernel to extract frequency band change features vertically and time continuity features horizontally; A learnable frequency domain constraint layer is embedded in the multi-branch convolutional neural network, and convolution kernel parameters are initialized to wavelet basis functions of a preset partial discharge frequency band to enhance feature extraction of high-frequency discharge components; The output weights of different convolution channels are dynamically activated through the gating mechanism, and the features that meet the preset correlation threshold with the time-frequency characteristics of the current signal are screened to generate the time-frequency feature vector that characterizes the partial discharge behavior.

5. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 4, characterized in that: The extraction of the time-frequency characteristics of the effective component of the leakage current after separation, and the extraction of the local temperature gradient characteristics of the infrared thermal imaging data and the spectrum energy characteristics of the ultrasonic signal to generate a multimodal feature vector includes: Performing channel attention weighting on the time-frequency characteristics of the effective component of the leakage current to suppress power frequency noise interference and enhance the significant characteristics of the high-frequency discharge pulse; Performing spatial attention modeling on the local temperature gradient characteristics of the infrared thermal imaging data, extracting the temperature gradient direction and amplitude statistical distribution characteristics of the hotspot area, and eliminating background thermal interference; Performing multi-scale decomposition on the spectrum energy characteristics of the ultrasonic signal, extracting the energy fluctuation characteristics within a preset frequency band, and quantifying the coupling relationship between mechanical vibration and discharge; The weighted time-frequency features, temperature gradient features, and spectrum energy features are concatenated according to a preset dimension to generate a preliminary multimodal feature vector.

6. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 1, characterized in that: The attention mechanism is used to analyze the spatiotemporal correlation of the multimodal feature vectors, dynamically assign weight coefficients to each modality for fusion, and generate comprehensive fault features, including: The cross-modal correlation weights between electrical, thermal, and mechanical modalities are calculated through a preset cross-modal spatiotemporal attention network. The weight coefficient of each modal is adjusted according to the real-time operating conditions and the probability distribution of fault types, and comprehensive fault characteristics are generated through weighted fusion.

7. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 6, characterized in that: The preset cross-modal spatiotemporal attention network includes: For electrical modal features, the temporal attention module is used to capture the temporal clustering of discharge pulses, and the frequency domain attention module is used to focus on the energy distribution in the high-frequency band. For thermal modal features, the spatial attention module is used to locate temperature anomaly areas, and the temporal attention module is used to analyze temperature evolution trends. For mechanical modal features, the frequency points with energy surges are identified through the spectrum energy attention module.

8. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 7, wherein the step of calculating the cross-modal correlation weights among the electrical modality, the thermal modality, and the mechanical modality using a preset cross-modal spatiotemporal attention network comprises: The electrical modal features, thermal modal features, and mechanical modal features are grouped by sensor type, and the attention weights within the groups and between key groups are calculated; Introducing causal convolutional layers into temporal attention calculations to constrain weight distribution to rely solely on historical temporal information; Through lightweight matrix compression technology, the weight matrix of the fully connected layer is decomposed into a low-rank matrix product, and the attention weight is binarized to retain only the correlation features of the key modalities.

9. The method for monitoring the state of electric equipment based on a non-contact leakage current sensor according to claim 8, characterized in that: The method of analyzing the spatiotemporal correlation of the multimodal feature vectors by using the attention mechanism, dynamically allocating weight coefficients of each modality for fusion, and generating a comprehensive fault feature further includes: Performing spatiotemporal alignment on the high-frequency pulse amplitude of the electrical mode, the local temperature gradient amplitude of the thermal mode, and the frequency points of the spectral energy surge of the mechanical mode, and assigning a weight coefficient to each mode based on the correlation characteristics; The aligned multimodal features are fused to generate a comprehensive fault feature vector that is adapted to the edge computing model input.

10. The method for monitoring the state of electric power equipment based on a non-contact leakage current sensor according to claim 1, wherein the step of analyzing the insulating component based on the comprehensive fault feature vector and a preset edge computing device to obtain a target defect location comprises: Deploying a lightweight diagnostic model in the edge computing device, inputting the comprehensive fault characteristics into the lightweight diagnostic model to predict defect probability, and outputting a defect probability distribution map of each area of ​​the insulation component; Based on the defect probability distribution map and a preset spatial mapping relationship between the insulating components, determining an area whose probability exceeds a dynamic threshold as a candidate defect area; The candidate defect areas are matched using the comprehensive feature vector to obtain a target defect location.

Images