Fiber bragg grating sound wave intelligent sensing and detecting method for partial discharge of cable joint

Abstract

The invention provides a cable joint partial discharge fiber bragg grating sound wave intelligent sensing and detection method, and relates to the technical field of cable partial discharge detection, and the method comprises the steps: obtaining multi-channel time-domain spectral response data of a detected cable joint through a fiber bragg grating sensing array, and demodulating the data to obtain a multi-dimensional feature vector group; decomposing the waveform signal into narrowband intrinsic mode components, extracting energy density through Hilbert transform and executing singular value decomposition, and screening dominant mode components to reconstruct the waveform signal to obtain stress wave time sequence characteristics; calculating a waveform high-order cross-correlation tensor as a hyperedge weight, extracting a time-frequency amplitude feature vector as an initial node feature, constructing a space-time hypergraph, and generating a node embedding vector through graph convolution; and initializing a Gaussian noise field based on a node embedding vector and setting condition information, executing reverse sampling denoising to generate a three-dimensional space probability density field to identify an initial positioning coordinate, and iteratively correcting the positioning coordinate and a wave velocity value until a residual error converges.

Description

Technical Field

[0001] This invention relates to the field of partial discharge detection technology for cables, and in particular to a method for intelligent sensing and detection of partial discharge in cable joints using fiber optic gratings and acoustic waves. Background Technology

[0002] As a crucial component of the power system, power cables directly impact the safety and stability of the power grid. Cable joints, being the weakest link in cable lines, are prone to insulation defects due to factors such as manufacturing processes, installation quality, and operating environment, leading to partial discharge. Partial discharge is a primary indicator of cable insulation degradation; persistent partial discharge gradually erodes the insulation material, ultimately causing insulation breakdown. Therefore, effective monitoring and precise location of partial discharge at cable joints are of paramount importance for preventing cable faults and ensuring the safe operation of the power grid.

[0003] Currently, cable partial discharge detection technologies mainly include electrical testing methods, ultrasonic testing methods, and ultra-high frequency testing methods. Fiber optic grating sensing technology, as a new type of acoustic wave detection method, has advantages such as inherent safety, resistance to electromagnetic interference, and the ability to achieve distributed measurement, and is gradually being applied to the field of cable partial discharge monitoring.

[0004] However, existing fiber optic grating acoustic wave detection and localization technologies for partial discharge of cable joints still have problems such as insufficient suppression of complex background noise and multi-source interference, failure to fully explore multi-channel spatiotemporal correlation information leading to insufficient feature expression, and weak adaptability to wave velocity uncertainty and propagation path complexity, resulting in large localization errors and making it difficult to meet the engineering requirements for accurate localization of partial discharge of cable joints. Summary of the Invention

[0005] This invention provides a method for intelligent sensing and detection of partial discharge in cable joints using fiber Bragg gratings, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of the present invention provides a method for intelligent sensing and detection of partial discharge fiber Bragg grating acoustic waves in cable joints, comprising: The multi-channel time-domain spectral response data of the cable joint under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix and the dominant mode components are selected to reconstruct the waveform signal to obtain the stress wave time series characteristics. The higher-order cross-correlation tensor of the waveform of the stress wave time series features is used as the hyperedge weight. The time-frequency amplitude feature vector in the stress wave time series features is used as the initial node feature. A spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature. The initial node feature is aggregated through graph convolution in the spatiotemporal hypergraph to generate a hyperedge representation. The hyperedge representation is aggregated and the node embedding vector is obtained through order neighborhood aggregation. The Gaussian noise field is initialized based on the node embedding vector and condition information is set. Inverse sampling denoising is performed based on the condition information to generate an initial three-dimensional spatial probability density field and local maxima points are identified as initial positioning coordinates. The theoretical propagation time is calculated based on the preset wave velocity value and the residual is calculated in combination with the stress wave time series characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

[0007] In one alternative implementation, The multi-channel time-domain spectral response data of the cable joint under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector set, including: The grating period change caused by stress wave generated by partial discharge is collected by a fiber optic grating sensor array uniformly distributed along the circumference of the cable joint under test. The Bragg wavelength drift of each grating node is recorded in real time as a function of time to obtain multi-channel time-domain spectral response data. Baseline drift correction is performed on the multi-channel time-domain spectral response data. Wavelength abrupt changes exceeding the standard deviation of background noise in each channel are identified as the start time of the effective signal. Wavelength drift sequence within the effective signal time window is extracted. The wavelength drift of each sampling point in the wavelength drift sequence is converted into stress wave amplitude value through the preset sensitivity coefficient, and the stress wave time domain waveform signal corresponding to each grating node is constructed. Combined with the spatial coordinate information of each grating node, a spatiotemporally correlated waveform signal group is formed. Peak amplitude is extracted by peak detection of the time-domain waveform of each channel in the waveform signal group. The arrival time difference of each channel relative to the reference channel is determined by cross-correlation analysis and the arrival time of each channel is calculated by combining the absolute arrival time of the reference channel. Short-time Fourier transform is performed on the time-domain waveform of each channel to obtain the time spectrum and the spectral centroid is calculated. The waveform signal group, the arrival time, the peak amplitude and the spectral centroid are combined into a multidimensional feature vector group.

[0008] In one alternative implementation, The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components, and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix, and the dominant mode components are selected to reconstruct the waveform signal, obtaining the stress wave time series features, including: Initialize the candidate mode components corresponding to the waveform signal in the multidimensional feature vector group and minimize the spectral bandwidth to obtain the preliminary decomposition result. Determine the frequency trajectory based on the preliminary decomposition result, calculate the similarity of the frequency trajectories between different channels and determine similar mode pairs by combining the preset similarity threshold, and perform cross-correlation analysis and time compensation on the similar mode pairs to obtain narrowband intrinsic mode components. The narrowband intrinsic mode components are subjected to Hilbert transform to obtain analytic signals, and the instantaneous amplitude is extracted as the instantaneous energy density. The instantaneous energy density is arranged according to the time sampling points to form a time series and organized to obtain an energy density matrix. Singular value decomposition is performed on the energy density matrix to obtain a sequence of singular values ​​and corresponding left and right vector matrices. A time pattern is extracted from the left vector matrix and a time metric is calculated. A spatial pattern is extracted from the right vector matrix and a spatial metric is calculated. A smoothness score is calculated based on the time and spatial metrics, and a screening index is determined. The singular value sequence is screened based on the screening index to determine the dominant singular values ​​and reconstruct a denoised energy density matrix. An inverse Hilbert transform is performed on the denoised energy density matrix to obtain a time-domain waveform. The time-domain waveform is summed and superimposed, and time-frequency amplitude feature parameters are extracted to obtain stress wave time series features.

[0009] In one alternative implementation, The higher-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves is used as the hyperedge weight. The time-frequency amplitude feature vector in the time-series features of stress waves is used as the initial node feature. The spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature, including: The time-frequency amplitude feature vector in the stress wave time series features is divided into multiple time segments according to the time window. The time-frequency amplitude feature vector in each time segment is constructed as a graph node. Three graph nodes are randomly selected from the graph nodes of the current time segment to obtain a candidate hyperedge node group. The third-order tensor outer product operation is performed on the time-frequency amplitude feature vector corresponding to the candidate hyperedge node group and the tensor norm is extracted as the cooperative strength. The cooperative strength is normalized to obtain the hyperedge weight. The time-domain peak amplitude, dominant frequency, and energy centroid time position are extracted from the stress wave time-series features and spliced ​​together to form the initial node features of the graph nodes; A hyperedge correlation matrix is ​​constructed based on the hyperedge weights, and a node feature matrix is ​​constructed based on the initial node feature vectors. A node indirect matrix is ​​obtained by multiplying the transpose of the hyperedge correlation matrix with the hyperedge correlation matrix. The node indirect matrix is ​​multiplied with the node feature matrix and a nonlinear transformation is performed to obtain updated node features. The hyperedge correlation matrices corresponding to each time segment are concatenated along the time dimension to obtain a spatiotemporal hyperedge matrix. A spatiotemporal hypergraph is constructed by combining the updated node features.

[0010] In one alternative implementation, In the spatiotemporal hypergraph, the initial node features are aggregated through graph convolution to generate hyperedge representations. Aggregating these hyperedge representations and then aggregating them through order-neighbor clusters yields node embedding vectors, including: Obtain the graph nodes in the spatiotemporal hypergraph, calculate the feature covariance matrix of the initial node features corresponding to each graph node, and extract the principal component directions as the intrinsic manifold basis of the hyperedge. Project the initial node features onto the intrinsic manifold basis to obtain the manifold coordinate representation and calculate the manifold distance between graph nodes. Based on the manifold distance, construct the adaptive adjacency weights within the hyperedge and perform manifold-aware weighted aggregation on the initial node features. Combine the hyperedge weights with multiplication gating to obtain the hyperedge representation. Extract the set of associated hyperedges corresponding to each graph node and calculate the mutual information between the hyperedge representation corresponding to each associated hyperedge and the hyperedge representation corresponding to the current graph node. Based on the mutual information, filter the hyperedge representations in the set of associated hyperedges to obtain the first-level node features. Based on the first-level node features, a Laplacian matrix is ​​constructed and a graph filter is obtained by extracting low-frequency feature vectors through spectral decomposition. The first-level node features corresponding to each graph node are then subjected to frequency domain filtering and polynomial expansion through the graph filter to obtain multi-order neighborhood aggregation coefficients. Based on the multi-order neighborhood aggregation coefficients, the hyperedge representation is weighted and aggregated to obtain the node embedding vector.

[0011] In one alternative implementation, Initializing a Gaussian noise field based on the node embedding vector and setting condition information, performing inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identifying local maxima points as initial positioning coordinates includes: The node embedding vector is subjected to feature encoding and spatial broadcasting to obtain spatial conditional features. A three-dimensional Gaussian noise field following a standard normal distribution is initialized and concatenated with the corresponding spatial conditional features to obtain a conditional noise field. The total time step is set and the current time step is initialized. The conditional noise field and the time encoding corresponding to the current time step are subjected to convolution transformation to obtain noise components. The noise components are subtracted from the conditional noise field representation and a preset scaling factor is added to obtain a denoised field representation. The denoised field representation is used as the conditional noise field of the next time step and the time step is decremented. This process is repeated until the time step is zero to obtain the initial three-dimensional spatial probability density field. Based on a preset probability density threshold, the grid points in the initial three-dimensional spatial probability density field are filtered to obtain an initial grid point set. Initial grid points with probability density values ​​greater than all adjacent grid points are identified as candidate maxima. The gradient magnitude between the probability density value of the candidate maxima and the probability density values ​​of the corresponding adjacent grid points is calculated. The candidate maxima are sorted in descending order according to the gradient magnitude, and the first candidate maxima is selected to extract the coordinates in three-dimensional space to obtain the initial positioning coordinates.

[0012] In one alternative implementation, The theoretical propagation time is calculated based on a preset wave velocity value, and the residual is calculated in conjunction with the stress wave timing characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than a preset residual threshold to obtain the corrected positioning coordinates. Calculate the spatial distance between the initial positioning coordinates and the positions of each sensor, calculate the theoretical propagation time in combination with the preset wave velocity value, extract the actual arrival time from the stress wave timing characteristics, and calculate the time difference between the theoretical propagation time and the actual arrival time as the time residual. Construct a state vector containing the three-dimensional coordinate components of the initial positioning coordinates and the preset wave velocity value, determine whether the time residual is less than a preset time threshold, and if it is less, output the initial positioning coordinates in the state vector as the corrected positioning coordinates. If the value is not less than the threshold value, a random perturbation is applied to the state vector, and the time residual is recalculated to obtain the Jacobian matrix. The gradient vector is calculated based on the transpose of the Jacobian matrix and the time residual. A damping matrix is ​​constructed based on the Jacobian matrix and a preset damping coefficient. A system of linear equations is constructed based on the damping matrix and the gradient vector, and the correction increment is obtained by solving the system. The state vector is updated based on the correction increment to obtain the updated state vector. The root mean square of the current time residual is calculated based on the updated state vector as the root mean square value of the updated residual, and the damping coefficient is corrected. This process is repeated until the root mean square value of the updated residual is less than a preset residual threshold value. The corrected positioning coordinates are then obtained and output.

[0013] A second aspect of the present invention provides a fiber Bragg grating acoustic wave intelligent sensing and detection system for partial discharge of cable joints, comprising: The first unit is used to acquire multi-channel time-domain spectral response data corresponding to the cable joint under test through a fiber optic grating sensing array and demodulate it based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The second unit is used to decompose the waveform signal in the multidimensional feature vector group into narrowband intrinsic mode components and extract the energy density through Hilbert transform to construct an energy density matrix. The unit then performs singular value decomposition on the energy density matrix and selects the dominant mode components to reconstruct the waveform signal to obtain the stress wave time series characteristics. The third unit uses the high-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves as the hyperedge weights, takes the time-frequency amplitude feature vector in the time-series features of stress waves as the initial node features, constructs a spatiotemporal hypergraph based on the hyperedge weights and the initial node features, generates a hyperedge representation by aggregating the initial node features in the spatiotemporal hypergraph through graph convolution, and obtains a node embedding vector by aggregating the hyperedge representation and aggregating the order neighborhood. The fourth unit is used to initialize a Gaussian noise field based on the node embedding vector and set condition information, perform inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identify local maxima points as initial positioning coordinates, calculate the theoretical propagation time based on the preset wave velocity value and calculate the residual in combination with the stress wave time series characteristics, and iteratively correct the initial positioning coordinates and the wave velocity value until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

[0014] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this invention, multi-channel time-domain spectral response data is acquired through a fiber optic grating sensing array and demodulated using sensitivity coefficients. Energy density features are extracted using narrowband intrinsic mode decomposition and Hilbert transform. Then, singular value decomposition is used to screen dominant mode components and reconstruct the waveform signal. This effectively suppresses noise interference, enhances the characteristics of stress wave signals generated by partial discharge, and improves the detection sensitivity and signal-to-noise ratio of weak acoustic signals. It can accurately capture early signs of partial discharge. A spatiotemporal hypergraph structure is used to model the temporal characteristics of stress waves, with high-order cross-correlation tensors of the waveform used as hyperedge weights. Node embedding vectors are extracted through graph convolution and multi-order neighborhood aggregation mechanisms. By fully exploring the spatiotemporal correlation and high-order coupling relationships between multi-channel sensor data, the propagation characteristics of partial discharge acoustic signals can be more comprehensively characterized, providing richer feature information for subsequent accurate positioning. Based on the node embedding vector, a Gaussian noise field is initialized and a three-dimensional spatial probability density field is generated through inverse sampling denoising guided by conditional information. The positioning coordinates and wave velocity values ​​are corrected by residual iteration combining theoretical propagation time and actual stress wave time sequence characteristics. This effectively handles the problems of wave velocity uncertainty and propagation path complexity, significantly improving the spatial positioning accuracy of partial discharge sources and providing a reliable basis for fault diagnosis and maintenance decisions of cable joints. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent sensing and detection method for partial discharge fiber optic grating acoustic waves in cable joints according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the positioning logic of the fiber optic grating acoustic wave intelligent sensing and detection method for partial discharge of cable joints according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the intelligent sensing and detection method for partial discharge fiber optic grating acoustic waves in cable joints according to an embodiment of the present invention. Figure 1 As shown, the method includes: The multi-channel time-domain spectral response data of the cable joint under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix and the dominant mode components are selected to reconstruct the waveform signal to obtain the stress wave time series characteristics. The higher-order cross-correlation tensor of the waveform of the stress wave time series features is used as the hyperedge weight. The time-frequency amplitude feature vector in the stress wave time series features is used as the initial node feature. A spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature. The initial node feature is aggregated through graph convolution in the spatiotemporal hypergraph to generate a hyperedge representation. The hyperedge representation is aggregated and the node embedding vector is obtained through order neighborhood aggregation. The Gaussian noise field is initialized based on the node embedding vector and condition information is set. Inverse sampling denoising is performed based on the condition information to generate an initial three-dimensional spatial probability density field and local maxima points are identified as initial positioning coordinates. The theoretical propagation time is calculated based on the preset wave velocity value and the residual is calculated in combination with the stress wave time series characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

[0021] In one alternative implementation, The multi-channel time-domain spectral response data of the cable joint under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector set, including: The grating period change caused by stress wave generated by partial discharge is collected by a fiber optic grating sensor array uniformly distributed along the circumference of the cable joint under test. The Bragg wavelength drift of each grating node is recorded in real time as a function of time to obtain multi-channel time-domain spectral response data. Baseline drift correction is performed on the multi-channel time-domain spectral response data. Wavelength abrupt changes exceeding the standard deviation of background noise in each channel are identified as the start time of the effective signal. Wavelength drift sequence within the effective signal time window is extracted. The wavelength drift of each sampling point in the wavelength drift sequence is converted into stress wave amplitude value through the preset sensitivity coefficient, and the stress wave time domain waveform signal corresponding to each grating node is constructed. Combined with the spatial coordinate information of each grating node, a spatiotemporally correlated waveform signal group is formed. Peak amplitude is extracted by peak detection of the time-domain waveform of each channel in the waveform signal group. The arrival time difference of each channel relative to the reference channel is determined by cross-correlation analysis and the arrival time of each channel is calculated by combining the absolute arrival time of the reference channel. Short-time Fourier transform is performed on the time-domain waveform of each channel to obtain the time spectrum and the spectral centroid is calculated. The waveform signal group, the arrival time, the peak amplitude and the spectral centroid are combined into a multidimensional feature vector group.

[0022] Multiple fiber Bragg grating sensors are uniformly arranged around the circumference of the cable joint to form a sensing array. The sensitivity coefficient of the fiber Bragg grating sensors is predetermined through calibration experiments, with a typical value of 0.78 pm / kPa. The sensing array includes eight fiber Bragg grating nodes located at 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° on the surface of the cable joint. Each fiber Bragg grating has a center wavelength of 1550 nm, a reflectivity of 99%, and a wavelength resolution of 1 pm. When partial discharge occurs inside the cable joint, the resulting stress wave propagates to the surface of the cable joint, causing a periodic change in the fiber Bragg gratings, resulting in a drift in the Bragg wavelength.

[0023] A high-speed spectrometer was used to acquire the Bragg wavelength information of each grating node in real time, with a sampling frequency of 500 kHz and an acquisition time of 200 ms. The wavelength drift of each grating node was recorded as a function of time to obtain multi-channel time-domain spectral response data. For example, in a partial discharge event, the raw wavelength drift data of the eight channels were recorded as data matrices, which included the changes in wavelength drift values ​​over time.

[0024] Baseline drift correction was performed on the acquired multi-channel time-domain spectral response data. Baseline drift is mainly caused by factors such as changes in ambient temperature and wavelength shift of the light source. Baseline correction employed a sliding window median filtering method with a window length of 2 ms. For each channel's data, the median within the sliding window was calculated as a baseline estimate. This baseline estimate was then subtracted from the original data to obtain the corrected wavelength drift data.

[0025] Valid signals are identified based on background noise levels. The standard deviation σ of the first 20ms of data for each channel is calculated as a background noise index, with a threshold of 3σ. The wavelength shift point that first exceeds this threshold in each channel is detected, and this moment is marked as the start time of the valid signal. Starting from the start time of the valid signal, the data for the next 50ms is extracted as the valid signal time window. For example, in a certain test, the standard deviation of background noise in channel 1 was 0.65pm, the threshold was set to 1.95pm, and the first wavelength drift value exceeding the threshold of 2.13pm was detected at 27.36ms. This moment was marked as the start time of the valid signal.

[0026] The wavelength drift sequence is converted into stress wave amplitude values. Based on a preset sensitivity coefficient of 0.78 pm / kPa, the wavelength drift at each sampling point is divided by this coefficient to obtain the corresponding stress wave amplitude value, in kPa. The time-domain waveform signal of the stress wave corresponding to each grating node is constructed, and combined with the spatial coordinate information of each grating node to form a spatiotemporally correlated waveform signal group. For example, the wavelength drift value of 2.13 pm in channel 1 is converted into a stress wave amplitude value of 2.73 kPa.

[0027] Peak detection is performed on the time-domain waveforms of each channel in the waveform signal group. A local maximum search algorithm is used to identify peak points in the waveform and extract their amplitude values. The minimum peak interval is set to 1 ms to avoid noise interference. For example, a peak amplitude of 5.86 kPa is detected at 32.56 ms in channel 1.

[0028] Cross-correlation analysis is used to determine the arrival time difference of each channel relative to the reference channel. The channel with the highest signal-to-noise ratio is selected as the reference channel, and the cross-correlation function between the waveforms of the other channels and the reference channel is calculated. The peak position of the cross-correlation function corresponds to the time delay between the two channels. Combined with the absolute arrival time of the reference channel, the arrival time of each channel is calculated. For example, taking channel 3 as the reference channel, its absolute arrival time is 28.12 ms, and the time delay of channel 1 relative to channel 3 is -0.76 ms, then the arrival time of channel 1 is 27.36 ms.

[0029] Short-time Fourier transform (SFT) is performed on the time-domain waveforms of each channel to obtain the time spectrum. The SFT parameters are set as follows: window length 10ms, overlap rate 50%, Hanning window. The calculated time spectrum contains three-dimensional information of time, frequency, and energy. The spectral centroid is calculated for the spectrum at each time step to characterize the dominant frequency characteristic of the signal. The spectral centroid is calculated by multiplying each frequency component by its energy and then dividing by the total energy. For example, the spectral centroid of channel 1 at 32.56ms is 125kHz.

[0030] The waveform signal group, arrival time, peak amplitude, and spectral centroid are combined into a multi-dimensional feature vector group. For the 8 channels, an 8×4 dimensional feature matrix is ​​constructed, with each row representing one channel and containing arrival time, peak amplitude, spectral centroid, and channel location information. For example, the feature vector of channel 1 is [27.36ms, 5.86kPa, 125kHz, 0°].

[0031] In this embodiment, by utilizing grating sensitivity to accurately map wavelength drift to stress wave amplitude and constructing a spatially consistent spatiotemporally correlated waveform group, the physical propagation process of stress waves is reconstructed, improving the accuracy of multi-point collaborative detection. Through a multi-channel arrival time difference extraction method based on cross-correlation analysis, the robustness of waveform arrival time calculation under waveform distortion conditions is effectively enhanced, improving the accuracy of partial discharge event localization. By fusing short-time Fourier transform time-frequency feature extraction and spectral centroid calculation, the energy distribution and frequency structure of the partial discharge signal are more comprehensively characterized, improving the distinguishability of different discharge types.

[0032] In one alternative implementation, The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components, and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix, and the dominant mode components are selected to reconstruct the waveform signal, obtaining the stress wave time series features, including: Initialize the candidate mode components corresponding to the waveform signal in the multidimensional feature vector group and minimize the spectral bandwidth to obtain the preliminary decomposition result. Determine the frequency trajectory based on the preliminary decomposition result, calculate the similarity of the frequency trajectories between different channels and determine similar mode pairs by combining the preset similarity threshold, and perform cross-correlation analysis and time compensation on the similar mode pairs to obtain narrowband intrinsic mode components. The narrowband intrinsic mode components are subjected to Hilbert transform to obtain analytic signals, and the instantaneous amplitude is extracted as the instantaneous energy density. The instantaneous energy density is arranged according to the time sampling points to form a time series and organized to obtain an energy density matrix. Singular value decomposition is performed on the energy density matrix to obtain a sequence of singular values ​​and corresponding left and right vector matrices. A time pattern is extracted from the left vector matrix and a time metric is calculated. A spatial pattern is extracted from the right vector matrix and a spatial metric is calculated. A smoothness score is calculated based on the time and spatial metrics, and a screening index is determined. The singular value sequence is screened based on the screening index to determine the dominant singular values ​​and reconstruct a denoised energy density matrix. An inverse Hilbert transform is performed on the denoised energy density matrix to obtain a time-domain waveform. The time-domain waveform is summed and superimposed, and time-frequency amplitude feature parameters are extracted to obtain stress wave time series features.

[0033] The multidimensional feature vector set contains waveform signals from eight channels, each corresponding to a fiber optic grating sensing node at a different location circumferentially on the cable connector. These channels typically contain multiple frequency components, requiring detailed analysis to extract effective features. Candidate modal components are initialized for the waveform signals in the multidimensional feature vector set. A variational mode decomposition method is employed, with a mode number of 5, a penalty factor of 2000, and a tolerance of 10. -7The initial decomposition results are obtained by minimizing the spectral bandwidth of each mode. Taking the waveform signal of channel 1 as an example, its initial decomposition yields 5 modal components with center frequencies of 35kHz, 75kHz, 125kHz, 165kHz, and 210kHz. Based on the initial decomposition results, the frequency trajectory is determined. For each modal component, the corresponding frequency-time curve is calculated using a short-time Fourier transform with a window length of 2ms and an overlap rate of 75%. The similarity of the frequency trajectories between different channels is calculated using cosine similarity, with similarity values ​​ranging from 0 to 1; a higher value indicates higher similarity. A preset similarity threshold of 0.85 is used. If the similarity of the frequency trajectories of the same modal component in two channels is greater than the threshold, they are identified as similar modal pairs. For example, the similarity of the third modal component between channel 1 and channel 2 is 0.92, thus they are identified as similar modal pairs.

[0034] Cross-correlation analysis is performed on similar mode pairs to calculate time delays. For each similar mode pair, a cross-correlation function is calculated, and the time offset corresponding to the maximum value of the cross-correlation function is used as the time delay between the two modes. For example, the time delay of the third mode component between channel 1 and channel 2 is 0.35ms. Time compensation is performed based on the time delay, shifting the lagging mode component in time to align it with the mode component of the reference channel. After time compensation, the aligned mode components are averaged to obtain the narrowband intrinsic mode components. For the signals from the eight channels of the cable connector, four main narrowband intrinsic mode components are extracted, with frequency ranges of 30-50kHz, 70-90kHz, 110-140kHz, and 160-180kHz, respectively.

[0035] A Hilbert transform is performed on the narrowband intrinsic mode components to obtain the analytic signal. The Hilbert transform generates orthogonal components of the original signal by delaying its phase by 90 degrees, constructing a complex analytic signal. The instantaneous amplitude is extracted from the analytic signal and calculated as its magnitude, serving as the instantaneous energy density. The instantaneous energy density is arranged into a time series based on the time sampling points, corresponding to the four narrowband intrinsic mode components and eight channels, resulting in an energy density matrix of 32 rows (4 modes × 8 channels) × 1000 columns (number of sampling points).

[0036] Singular value decomposition (SVD) is performed on the energy density matrix to obtain a sequence of singular values ​​and corresponding left and right vector matrices. Taking a partial discharge test as an example, SVD yields 32 singular values. The first five singular values ​​are 156.24, 98.75, 45.32, 21.89, and 10.76, respectively, after which the singular values ​​decrease rapidly. Temporal patterns are extracted based on the left vector matrix, with each column representing a temporal pattern. The continuity of the temporal patterns is calculated, and the time metric is defined as the average absolute difference between the elements of the left vector column vectors. A smaller time metric indicates a smoother temporal pattern. For the first five temporal patterns, the time metrics are 0.012, 0.025, 0.047, 0.089, and 0.153, respectively. Spatial patterns are extracted based on the right vector matrix, with each row representing a spatial pattern. The concentration of spatial patterns is calculated, and the spatial metric is defined as the variance of the right vector row vectors. A larger spatial metric indicates a more concentrated spatial pattern. The spatial metrics for the first five spatial models are 0.175, 0.132, 0.087, 0.054, and 0.028, respectively.

[0037] Smoothness scores were calculated based on temporal and spatial metrics, with the scoring formula being the spatial metric divided by the temporal metric. A higher score indicates that the pattern corresponding to that singular value is more likely to be a valid signal. The smoothness scores for the first five singular values ​​were 14.58, 5.28, 1.85, 0.61, and 0.18, respectively. A screening threshold of 1.0 was set to filter the singular value sequence, identifying the top three dominant singular values. The energy density matrix was reconstructed using the dominant singular values ​​and their corresponding left and right vectors, resulting in a denoised energy density matrix. The reconstruction process preserved valid signal components while filtering out noise components.

[0038] An inverse Hilbert transform is performed on the noise-reduced energy density matrix to recover the time-domain waveform. The inverse Hilbert transform recovers the original signal by extracting the real part of the analytic signal. The recovered time-domain waveforms are summed and superimposed, adding the time-domain waveforms of different modal components to obtain the complete time-domain signal. Time-frequency amplitude characteristic parameters, including peak time, peak amplitude, rise time, duration, dominant frequency, and bandwidth, are extracted from the superimposed time-domain waveform. For example, in a partial discharge test, the extracted characteristic parameters are: peak time 32.5 ms, peak amplitude 6.2 kPa, rise time 0.8 ms, duration 12.5 ms, dominant frequency 125 kHz, and bandwidth 45 kHz.

[0039] In this embodiment, by minimizing the spectral bandwidth to obtain preliminary mode decomposition results and combining multi-channel frequency trajectory similarity to screen similar modes, high-precision extraction of narrowband intrinsic mode components of partial discharge stress waves is achieved. This ensures that the decomposition results maintain consistency and physical comparability across multiple channels, significantly improving the problems of traditional single-channel decomposition being susceptible to noise interference and severe mode aliasing. By performing Hilbert transform on the narrowband mode components to construct an instantaneous energy density matrix, and extracting temporal and spatial modes based on singular value decomposition, and then using smoothness scoring to screen dominant singular values ​​to complete energy density matrix denoising, background noise and random disturbances can be effectively suppressed, making the true discharge energy distribution clearer in the time and spatial dimensions, and significantly enhancing the detection capability of weak discharges. By reconstructing the time-domain waveform through inverse Hilbert transform and superimposing it to extract time-frequency amplitude features, the stress wave time sequence features are made more stable and continuous, and the feature parameters more accurately characterize the discharge type, intensity, and propagation path.

[0040] In one alternative implementation, The higher-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves is used as the hyperedge weight. The time-frequency amplitude feature vector in the time-series features of stress waves is used as the initial node feature. The spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature, including: The time-frequency amplitude feature vector in the stress wave time series features is divided into multiple time segments according to the time window. The time-frequency amplitude feature vector in each time segment is constructed as a graph node. Three graph nodes are randomly selected from the graph nodes of the current time segment to obtain a candidate hyperedge node group. The third-order tensor outer product operation is performed on the time-frequency amplitude feature vector corresponding to the candidate hyperedge node group and the tensor norm is extracted as the cooperative strength. The cooperative strength is normalized to obtain the hyperedge weight. The time-domain peak amplitude, dominant frequency, and energy centroid time position are extracted from the stress wave time-series features and spliced ​​together to form the initial node features of the graph nodes; A hyperedge correlation matrix is ​​constructed based on the hyperedge weights, and a node feature matrix is ​​constructed based on the initial node feature vectors. A node indirect matrix is ​​obtained by multiplying the transpose of the hyperedge correlation matrix with the hyperedge correlation matrix. The node indirect matrix is ​​multiplied with the node feature matrix and a nonlinear transformation is performed to obtain updated node features. The hyperedge correlation matrices corresponding to each time segment are concatenated along the time dimension to obtain a spatiotemporal hyperedge matrix. A spatiotemporal hypergraph is constructed by combining the updated node features.

[0041] The time-frequency amplitude feature vector in the stress wave time series characteristics is divided into multiple time segments according to a time window, with a window length set to 10ms and an overlap rate of 50% between adjacent windows. For a 200ms acquisition data segment, 39 time segments can be obtained. Within each time segment, the time-frequency amplitude feature vector is extracted to construct graph nodes. The time-frequency amplitude feature vector includes parameters such as peak time, peak amplitude, rise time, duration, dominant frequency, and bandwidth. For example, in a partial discharge test of a cable joint, the 5th time segment contains 8 time-frequency amplitude feature vectors, corresponding to the signal characteristics of 8 fiber Bragg grating sensor channels. These 8 time-frequency amplitude feature vectors are respectively constructed as graph nodes, labeled as N1 to N8.

[0042] Candidate superedge node groups are obtained by randomly selecting three graph nodes from the graph nodes in the current time segment. A superedge is an edge connecting three or more nodes, representing a higher-order relationship between nodes. Using a random combination method, the number of combinations of selecting three nodes from N nodes is equal to the number of combinations of N-choose-3. For example, for 8 nodes, 56 three-node combinations can be obtained. In practical applications, distance or relevance thresholds can be set for filtering to reduce computational load. In the previous example, assume that 15 candidate superedge node groups are retained after filtering, denoted as E1 to E15. For example, E1 consists of nodes N1, N3, and N6, and E2 consists of nodes N2, N4, and N7.

[0043] Perform a third-order tensor outer product operation on the time-frequency amplitude feature vectors corresponding to the candidate hyperedge node groups. The tensor outer product can capture higher-order interactions between multiple vectors. For each three-node combination, extract the corresponding time-frequency amplitude feature vector and perform a third-order tensor outer product calculation. In specific implementation, the three feature vectors can be expanded into column vectors respectively, and the outer product of these three vectors can be calculated to obtain a third-order tensor. For example, for E1(N1, N3, N6), extract the corresponding time-frequency amplitude feature vectors v1, v3, and v6, each with a dimension of 6, and calculate the third-order tensor outer product of v1, v3, and v6 to obtain a 6×6×6 third-order tensor T1.

[0044] The tensor norm is extracted from the calculated third-order tensor as the cooperation strength. The tensor norm measures the "size" of the tensor, reflecting the strength of the cooperation between nodes. The Frobenius norm is used, calculated as the square root of the sum of squares of all elements in the tensor. For the third-order tensor T1 corresponding to E1, its Frobenius norm yields a cooperation strength value of 125.7. This process is repeated for all candidate hyperedge node groups to obtain a set of cooperation strength values. For example, the cooperation strengths of the 15 candidate hyperedges are: 125.7, 98.3, 85.2, 76.9, 70.5, 68.2, 65.8, 62.1, 58.9, 55.3, 52.6, 48.7, 45.2, 42.8, and 39.1.

[0045] The collaboration strengths are normalized to obtain the hyperedge weights. The normalization method uses maximum-minimum scaling to map the collaboration strength values ​​to the [0, 1] interval. For the collaboration strengths of the aforementioned 15 candidate hyperedges, the normalized hyperedge weights are as follows: 1.00, 0.68, 0.53, 0.44, 0.36, 0.33, 0.31, 0.27, 0.23, 0.19, 0.16, 0.11, 0.07, 0.04, 0.00.

[0046] The peak amplitude, dominant frequency, and energy centroid time position in the stress wave time series characteristics are extracted and concatenated to obtain the initial node features of the graph nodes. The peak amplitude in the time domain reflects the discharge intensity, the dominant frequency reflects the discharge type characteristics, and the energy centroid time position reflects the signal energy distribution. For the aforementioned 8 graph nodes, their initial feature vectors are extracted. For example, the initial feature vector of node N1 is [5.8kPa, 125kHz, 32.5ms], and the initial feature vector of node N2 is [4.2kPa, 135kHz, 33.1ms].

[0047] Construct a hyperedge incidence matrix based on the hyperedge weights. The hyperedge incidence matrix is ​​a matrix of the number of nodes multiplied by the number of hyperedges, where each element represents the relationship between a node and a hyperedge. If node i belongs to hyperedge j, the value of the matrix element (i, j) is the weight of that hyperedge; otherwise, it is 0. For the previous example, construct an 8×15 hyperedge incidence matrix H. For instance, if hyperedge E1 consists of nodes N1, N3, and N6 with a weight of 1.00, then the elements H(1, 1) = H(3, 1) = H(6, 1) = 1.00 in matrix H, and the elements in the first column of all other rows are 0.

[0048] A node feature matrix is ​​constructed based on the initial node feature vectors. The node feature matrix is ​​a matrix of the number of nodes × the feature dimension, with each row representing the feature vector of a node. For the aforementioned 8 nodes, each node has a feature dimension of 3, and an 8×3 node feature matrix X is constructed. For example, the first row of X is the feature vector of node N1 [5.8kPa, 125kHz, 32.5ms].

[0049] The node indirection matrix is ​​obtained by multiplying the transpose of the hyperedge incidence matrix H by the hyperedge incidence matrix H. The node indirection matrix reflects the indirect connections between nodes established through hyperedges. Specifically, it is calculated by multiplying the transpose of the hyperedge incidence matrix H by the hyperedge incidence matrix H. T Multiplying by H yields a square matrix of node number × node number. For the example above, calculate an 8 × 15 matrix H. T Multiplying it by the 15×8 matrix H yields an 8×8 node indirection matrix A.

[0050] The node indirect matrix is ​​multiplied by the node feature matrix, and a nonlinear transformation is performed to obtain the updated node features. The 8×8 node indirect matrix A is multiplied by the 8×3 node feature matrix X to obtain an 8×3 matrix. The ReLU nonlinear transformation function is then applied to obtain the updated node feature matrix X'. For example, the updated feature vector of node N1 becomes [6.2 kPa, 128 kHz, 32.3 ms], capturing the interaction information with other nodes.

[0051] The spatiotemporal hyperedge matrix is ​​obtained by concatenating the hyperedge association matrices corresponding to each time segment along the time dimension. For the 39 divided time segments, hyperedge association matrices H1 to H39 are constructed respectively, and these matrices are concatenated along the time dimension to form a spatiotemporal hyperedge matrix with a three-dimensional tensor structure. The spatiotemporal hyperedge matrix can represent the association relationship between nodes and hyperedges in different time segments, capturing the dynamic change characteristics in time series data.

[0052] A spatiotemporal hypergraph is constructed by combining updated node features. The spatiotemporal hypergraph contains node feature information and hyperedge connections, comprehensively describing the spatiotemporal structure and higher-order correlations of the partial discharge signal from the cable joint. Each node in the spatiotemporal hypergraph represents the time-frequency amplitude feature within a time segment, the hyperedges between nodes represent the cooperative relationship between features, and the hyperedge weights reflect the cooperative strength.

[0053] In this embodiment, by performing a third-order tensor outer product on the time-frequency amplitude feature vector within the same time segment and measuring the cooperative strength between the three nodes using the tensor norm, the nonlinear coupling relationship of the partial discharge stress wave in multiple feature dimensions can be characterized. This makes the feature relationship expression richer and more consistent with the physical characteristics of the complex coupling propagation of stress waves. By normalizing the cooperative strength to the hyperedge weight and constructing the hyperedge correlation matrix, the local structure in which multiple nodes participate can be captured in the hypergraph structure, improving the sensitivity to abnormal feature patterns, weak discharge feature combinations, and non-stationary energy distribution. Compared with ordinary graph structures, the ability to express complex relationships is significantly enhanced. Through the nonlinear fusion of the node feature matrix and the node indirect matrix, high-order feature propagation and updating based on the hypergraph structure are realized, enabling the node features to comprehensively reflect key attributes such as the dominant frequency, peak amplitude, and energy time distribution of the partial discharge signal.

[0054] In one alternative implementation, In the spatiotemporal hypergraph, the initial node features are aggregated through graph convolution to generate hyperedge representations. Aggregating these hyperedge representations and then aggregating them through order-neighbor clusters yields node embedding vectors, including: Obtain the graph nodes in the spatiotemporal hypergraph, calculate the feature covariance matrix of the initial node features corresponding to each graph node, and extract the principal component directions as the intrinsic manifold basis of the hyperedge. Project the initial node features onto the intrinsic manifold basis to obtain the manifold coordinate representation and calculate the manifold distance between graph nodes. Based on the manifold distance, construct the adaptive adjacency weights within the hyperedge and perform manifold-aware weighted aggregation on the initial node features. Combine the hyperedge weights with multiplication gating to obtain the hyperedge representation. Extract the set of associated hyperedges corresponding to each graph node and calculate the mutual information between the hyperedge representation corresponding to each associated hyperedge and the hyperedge representation corresponding to the current graph node. Based on the mutual information, filter the hyperedge representations in the set of associated hyperedges to obtain the first-level node features. Based on the first-level node features, a Laplacian matrix is ​​constructed and a graph filter is obtained by extracting low-frequency feature vectors through spectral decomposition. The first-level node features corresponding to each graph node are then subjected to frequency domain filtering and polynomial expansion through the graph filter to obtain multi-order neighborhood aggregation coefficients. Based on the multi-order neighborhood aggregation coefficients, the hyperedge representation is weighted and aggregated to obtain the node embedding vector.

[0055] Graph nodes in the spatiotemporal hypergraph are obtained. Based on the previously constructed spatiotemporal hypergraph, all graph nodes contained therein are extracted. For example, in the monitoring of partial discharge at cable joints, 312 graph nodes are extracted from the spatiotemporal hypergraph constructed from 39 time segments. Each node corresponds to a time-frequency amplitude feature vector within a time segment. The feature covariance matrix corresponding to the initial node features of each graph node is calculated using a sample covariance estimation method. For node N1, the initial feature vector is [5.8 kPa, 125 kHz, 32.5 ms], which, together with the feature vectors of adjacent nodes, constitutes a sample set. The 3×3 covariance matrix of the sample is calculated. The principal component directions of the feature covariance matrix are extracted using eigenvalue decomposition, selecting the feature vector corresponding to the largest eigenvalue as the principal component direction. For example, the principal component direction corresponding to node N1 is [0.85, 0.37, 0.38], representing the main change direction of the partial discharge signal in this region. This principal component direction is used as the intrinsic manifold basis of the hyperedge for subsequent manifold learning.

[0056] The initial node features are projected onto the intrinsic manifold basis to obtain manifold coordinate representations. The inner product of the initial node feature vectors and the principal component directions is calculated. For example, the initial feature vector of node N1 [5.8 kPa, 125 kHz, 32.5 ms] is projected onto the principal component directions [0.85, 0.37, 0.38], resulting in a manifold coordinate value of 98.6. For nodes N1, N3, and N6 contained in hyperedge E1, their manifold coordinate values ​​are calculated, resulting in [98.6, 82.3, 89.5]. The manifold distances between graph nodes are calculated, using Euclidean distances represented in manifold coordinates. For example, the manifold distance between nodes N1 and N3 is 16.3, the manifold distance between nodes N1 and N6 is 9.1, and the manifold distance between nodes N3 and N6 is 7.2.

[0057] Adaptive adjacency weights within hyperedges are constructed based on manifold distance, using a Gaussian kernel function with the kernel width parameter set to the average manifold distance. For example, the average manifold distance between nodes within hyperedge E1 is 10.9, resulting in an adjacency weight of 0.31 for nodes N1 and N3, 0.67 for nodes N1 and N6, and 0.78 for nodes N3 and N6. Manifold-aware weighted aggregation is then performed on the initial node features, using the adaptive adjacency weights to calculate a weighted average of the feature vectors of nodes within the hyperedge. For instance, after manifold-aware weighted aggregation of the initial feature vectors of the three nodes within hyperedge E1, the aggregated feature vector is obtained as [5.3 kPa, 130 kHz, 33.1 ms].

[0058] Multiplicative gating is applied using hyperedge weights, multiplying the hyperedge weights by the aggregated feature vector to obtain the hyperedge representation. For example, hyperedge E1 has a weight of 1.00, and its hyperedge representation is [5.3 kPa, 130 kHz, 33.1 ms]; hyperedge E2 has a weight of 0.68, and its aggregated feature vector is [4.5 kPa, 128 kHz, 33.5 ms]. After multiplicative gating, the hyperedge representation is [3.1 kPa, 87.0 kHz, 22.8 ms]. Through multiplicative gating, important hyperedges are enhanced while minor hyperedges are suppressed.

[0059] Extract the set of associated hyperedges for each graph node, i.e., all hyperedges containing that node. For example, the set of hyperedges associated with node N1 includes E1, E5, E8, and E12, and the set of hyperedges associated with node N3 includes E1, E3, E7, and E9. Calculate the mutual information between the hyperedge representation corresponding to each associated hyperedge and the hyperedge representation corresponding to the current graph node. The mutual information calculation is based on the kernel density estimation method, using a radial basis function kernel, with a bandwidth parameter set to 0.5. For example, the mutual information between the hyperedges E1, E5, E8, and E12 associated with node N1 and the hyperedge representation of node N1 are 0.85, 0.62, 0.43, and 0.29, respectively.

[0060] The hyperedge representations in the associated hyperedge set are filtered based on mutual information, with a mutual information threshold of 0.5. Hyperedges exceeding the threshold are retained, while those below are filtered out. For example, among the hyperedges associated with node N1, E1 and E5 have a mutual information greater than 0.5 and are retained; E8 and E12 have a mutual information less than 0.5 and are filtered out. The retained hyperedge representations are averaged to obtain the first-level node features. For example, the first-level node features of node N1 are the average of the hyperedge representations of E1 and E5 [5.0 kPa, 127 kHz, 32.8 ms]. This mutual information filtering improves the signal-to-noise ratio of the feature representations and filters out irrelevant information.

[0061] The Laplacian matrix is ​​constructed based on the features of the first-level nodes. The Laplacian matrix is ​​used to analyze the structural properties of the graph. The construction method involves calculating the similarity between nodes based on the first-level node features, generating an adjacency matrix, calculating the degree matrix (the diagonal matrix of the row sums of the adjacency matrices), and subtracting the adjacency matrix from the degree matrix to obtain the Laplacian matrix. For a graph with 312 nodes, a 312×312 Laplacian matrix is ​​obtained.

[0062] Low-frequency eigenvectors of the Laplacian matrix are extracted through spectral decomposition to obtain a graph filter. Spectral decomposition calculates the eigenvalues ​​and eigenvectors of the Laplacian matrix, and the eigenvectors corresponding to the k smallest eigenvalues ​​(k is set to 10) are selected to construct the graph filter. The low-frequency eigenvectors correspond to the global structure in the graph and can capture the main modes of the discharge signal. The first-level node features corresponding to each graph node are frequency-domain filtered using the graph filter. This is done by projecting the node features onto the eigenvector space and approximating the calculations using Chebyshev polynomials. For 312 nodes, the eigenvectors of each node, after frequency-domain filtering, yield 10 frequency-domain coefficients.

[0063] Polynomial expansion yields multi-order neighborhood aggregation coefficients. The frequency-domain filtered coefficients are then transformed into spatial-domain multi-order neighborhood aggregation coefficients through the inverse Chebyshev polynomial transform. The polynomial order is set to 3, corresponding to the aggregation weights of first-, second-, and third-order neighbors. For example, the multi-order neighborhood aggregation coefficients of node N1 are [0.45, 0.32, 0.23], indicating that the aggregation weights for first-, second-, and third-order neighbors are 0.45, 0.32, and 0.23, respectively.

[0064] The node embedding vector is obtained by weighted aggregation of hyperedge representations based on multi-order neighborhood aggregation coefficients. For each node, the hyperedge representations of its first-order, second-order, and third-order neighboring nodes are collected and weighted and summed according to the multi-order neighborhood aggregation coefficients. For example, the first-order neighboring nodes of node N1 are N3, N6, etc., the second-order neighboring nodes are N2, N4, etc., and the third-order neighboring nodes are N5, N7, etc. The hyperedge representations of the nodes are extracted respectively, and weighted and summed according to the aggregation coefficients [0.45, 0.32, 0.23] to obtain the embedding vector of node N1 [4.8kPa, 126kHz, 32.7ms, 0.37, 0.65]. The first three components come from the original feature space, and the last two components are new feature dimensions obtained through graph structure learning, which can express the position information of the node in the graph topology.

[0065] In this embodiment, by calculating the covariance matrix of graph node features and extracting the principal component directions as the intrinsic manifold basis of hyperedges, the feature aggregation process can adaptively align the main change directions of partial discharge signals in the high-dimensional feature space, significantly enhancing the fidelity of feature representation for complex time-frequency amplitude features. This makes the distance measurement between nodes more consistent with the actual physical laws of stress wave evolution. By projecting the initial node features onto the manifold basis and calculating the manifold distance between nodes, and then adaptively constructing the hyperedge adjacency weights based on the manifold distance between nodes, nonlinear weighted aggregation of feature relationships is achieved. This effectively suppresses weakly correlated or noise-dominated features, while strengthening feature patterns with physical consistency, thus improving the accuracy of hyperedge representation. By calculating the mutual information between graph nodes and their associated hyperedges and selecting key hyperedges with high mutual information, redundant associations are effectively filtered out, and the most information-contributing partial discharge feature combinations are retained. This significantly improves the targeting of feature selection, enhances the efficiency and reliability of representation learning, and has a clear advantage in preserving weak discharge features in complex backgrounds.

[0066] In one alternative implementation, Initializing a Gaussian noise field based on the node embedding vector and setting condition information, performing inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identifying local maxima points as initial positioning coordinates includes: The node embedding vector is subjected to feature encoding and spatial broadcasting to obtain spatial conditional features. A three-dimensional Gaussian noise field following a standard normal distribution is initialized and concatenated with the corresponding spatial conditional features to obtain a conditional noise field. The total time step is set and the current time step is initialized. The conditional noise field and the time encoding corresponding to the current time step are subjected to convolution transformation to obtain noise components. The noise components are subtracted from the conditional noise field representation and a preset scaling factor is added to obtain a denoised field representation. The denoised field representation is used as the conditional noise field of the next time step and the time step is decremented. This process is repeated until the time step is zero to obtain the initial three-dimensional spatial probability density field. Based on a preset probability density threshold, the grid points in the initial three-dimensional spatial probability density field are filtered to obtain an initial grid point set. Initial grid points with probability density values ​​greater than all adjacent grid points are identified as candidate maxima. The gradient magnitude between the probability density value of the candidate maxima and the probability density values ​​of the corresponding adjacent grid points is calculated. The candidate maxima are sorted in descending order according to the gradient magnitude, and the first candidate maxima is selected to extract the coordinates in three-dimensional space to obtain the initial positioning coordinates.

[0067] Feature encoding and spatial broadcasting are performed on the node embedding vectors to obtain spatial conditional features. Feature encoding employs a multilayer perceptron structure containing two fully connected layers: a hidden layer dimension of 128 and an output layer dimension of 64, with the LeakyReLU activation function. For example, the embedding vector of node N1 [4.8kPa, 126kHz, 32.7ms, 0.37, 0.65] yields a 64-dimensional feature vector after feature encoding. The spatial broadcasting operation copies the encoded feature vector onto a predefined 3D grid with a resolution of 64×64×64, covering the entire monitoring space of the cable joint (600mm×400mm×400mm). For instance, the encoded 64-dimensional feature vector, after spatial broadcasting, results in a 64×64×64×64 four-dimensional tensor, indicating that each point in space has the same 64-dimensional feature representation.

[0068] A three-dimensional Gaussian noise field following a standard normal distribution is initialized, with the same dimensions as the spatial grid, i.e., 64×64×64. Noise values ​​are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 1. For example, the generated noise field has a noise value of 0.73 at coordinates (32, 16, 48) and a noise value of -0.51 at coordinates (10, 42, 25). The noise field is then concatenated with the corresponding spatial conditional features to obtain a conditional noise field. The concatenation operation is performed along the feature dimension, merging the 64-dimensional spatial conditional features with the 1-dimensional noise values ​​to obtain a conditional noise field with a shape of 64×64×64×65. The first 64 channels of the last dimension represent the conditional features, and the last channel represents the noise value.

[0069] The total time step is set to 1000, representing the number of iterations in the diffusion process. The current time step is initialized to 1000. The time steps decrease from large to small, corresponding to the noise removal process from large to small. A convolutional transformation is performed on the conditional noise field and the time code corresponding to the current time step to obtain the noise component. The time code uses a sinusoidal positional encoding method, mapping the time step to a 128-dimensional vector representation. The convolutional transformation uses a 3D convolutional structure, including residual blocks and an attention mechanism, with a kernel size of 3×3×3, a stride of 1, and padding of 1. For example, at time step 1000, performing a convolutional transformation on a conditional noise field with a shape of 64×64×64×65 yields a noise component with a shape of 64×64×64×1.

[0070] The denoised field representation is obtained by subtracting the noise component from the conditional noise field representation and adding a preset scaling factor. The scaling factor is dynamically adjusted according to the current time step and is calculated by dividing the current time step by the square root of the total time steps and multiplying by the preset factor of 0.01. For example, at time step 1000, the scaling factor is 0.01×(1000 / 1000)^0.5=0.01; at time step 500, the scaling factor is 0.01×(500 / 1000)^0.5≈0.007. For the conditional noise field value of 0.73 and the corresponding noise component of 0.12 at coordinates (32, 16, 48), the subtraction operation yields 0.61. After adding the scaling factor of 0.01, the denoised field representation is 0.616.

[0071] The denoised field representation is used as the conditional noise field for the next time step, and the time step is decreased, for example, from 1000 to 999. The denoised field representation with a shape of 64×64×64×1 is concatenated with the conditional feature with a shape of 64×64×64×64 to obtain a new conditional noise field for processing in the next time step. The aforementioned denoising process is repeated until the time step is zero, resulting in an initial three-dimensional spatial probability density field. After 1000 iterations, the noise is gradually removed, and the final three-dimensional spatial probability density field has a shape of 64×64×64, representing the probability density of a local discharge source at each location in the cable joint space.

[0072] The initial set of grid points is obtained by filtering the grid points in the initial three-dimensional probability density field based on a preset probability density threshold. The threshold is set to 70% of the maximum value of the probability density field. For example, if the maximum probability density value is 0.87 and it appears at coordinates (35, 22, 41), then the threshold is 0.87 × 70% ≈ 0.61. All points in the 64 × 64 × 64 grid are traversed, and grid points with probability density values ​​greater than 0.61 are selected to form the initial set of grid points. In the actual case, approximately 280 grid points were selected, distributed around the local discharge power supply of the cable joint.

[0073] Initial grid points with probability density values ​​greater than all their neighboring grid points are identified as candidate maxima. A neighboring grid point is defined as one of the 26 points in 3D space adjacent to the current grid point (excluding itself within a 3×3×3 neighborhood). For example, for a grid point at coordinates (35, 22, 41), the probability density values ​​of its 26 neighboring points are checked. If the probability density value of this point (0.87) is greater than the probability density values ​​of all its neighbors, it is marked as a candidate maxima. In the actual case, five candidate maxima were identified, with coordinates (35, 22, 41), (34, 21, 42), (36, 23, 40), (33, 24, 41), and (37, 21, 39), and corresponding probability density values ​​of 0.87, 0.85, 0.82, 0.79, and 0.76, respectively.

[0074] Calculate the gradient magnitude between the probability density values ​​of candidate maxima and their corresponding neighboring grid points. The gradient magnitude is calculated using the central difference method, calculating the gradients in the x, y, and z directions respectively, and then taking the Euclidean norm. For example, for a candidate maximum point at coordinates (35, 22, 41), the gradient in the x-direction is (P(36, 22, 41) - P(34, 22, 41)) / 2 = (0.83 - 0.82) / 2 = 0.005, the gradient in the y-direction is (P(35, 23, 41) - P(35, 21, 41)) / 2 = (0.84 - 0.83) / 2 = 0.005, and the gradient in the z-direction is (P(35, 22, 42) - P(35, 22, 40)) / 2 = (0.84 - 0.81) / 2 = 0.015. The gradient magnitude is (0.005^2 + 0.005^2 + 0.015^2)^0.5 ≈ 0.017. The gradient magnitudes were calculated for the five candidate maxima, and the results were 0.017, 0.015, 0.014, 0.012 and 0.011.

[0075] Candidate maxima are sorted in descending order based on gradient magnitude, and the first candidate maxima is selected. Its coordinates in 3D space are extracted to obtain the initial positioning coordinates. After sorting by gradient magnitude, the five candidate maxima are (35, 22, 41), (34, 21, 42), (36, 23, 40), (33, 24, 41), and (37, 21, 39). The first candidate maxima (35, 22, 41), with corresponding physical coordinates of (328.1 mm, 206.3 mm, 384.4 mm), is selected as the initial positioning coordinates for the partial discharge source.

[0076] In this embodiment, by performing feature encoding and spatial broadcasting on the node embedding vectors, the high-dimensional semantic features contained in the spatiotemporal hypergraph are mapped to three-dimensional space. This imbues the diffusion process with location correlation and physical constraints of partial discharge events, making the spatial distribution inference more consistent with the actual stress wave propagation law. By initializing a standard normal three-dimensional noise field and performing conditional denoising iterations over multiple time steps, noise is gradually eliminated using convolutional transformations and temporal encoding while preserving the spatial structure dominated by node embeddings. This achieves a reconstruction process from noise to structure, exhibiting stronger noise resistance and nonlinear fitting capabilities. Even under low signal-to-noise ratio conditions, it can maintain clear and stable probability peak regions. By selecting grid points based on probability density thresholds and employing a joint sorting mechanism of "probability maxima points + gradient magnitude," the problem of susceptibility to local noise peak interference is effectively avoided, making the extracted candidate maxima points more physically reasonable. The introduction of gradient magnitude can reduce spurious peaks and noise peaks, improving the sensitivity to the actual discharge source location.

[0077] In one alternative implementation, The theoretical propagation time is calculated based on a preset wave velocity value, and the residual is calculated in conjunction with the stress wave timing characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than a preset residual threshold to obtain the corrected positioning coordinates. Calculate the spatial distance between the initial positioning coordinates and the positions of each sensor, calculate the theoretical propagation time in combination with the preset wave velocity value, extract the actual arrival time from the stress wave timing characteristics, and calculate the time difference between the theoretical propagation time and the actual arrival time as the time residual. Construct a state vector containing the three-dimensional coordinate components of the initial positioning coordinates and the preset wave velocity value, determine whether the time residual is less than a preset time threshold, and if it is less, output the initial positioning coordinates in the state vector as the corrected positioning coordinates. If the value is not less than the threshold value, a random perturbation is applied to the state vector, and the time residual is recalculated to obtain the Jacobian matrix. The gradient vector is calculated based on the transpose of the Jacobian matrix and the time residual. A damping matrix is ​​constructed based on the Jacobian matrix and a preset damping coefficient. A system of linear equations is constructed based on the damping matrix and the gradient vector, and the correction increment is obtained by solving the system. The state vector is updated based on the correction increment to obtain the updated state vector. The root mean square of the current time residual is calculated based on the updated state vector as the root mean square value of the updated residual, and the damping coefficient is corrected. This process is repeated until the root mean square value of the updated residual is less than a preset residual threshold value. The corrected positioning coordinates are then obtained and output.

[0078] The spatial distance between the initial positioning coordinates and the positions of each sensor is calculated using a three-dimensional Euclidean distance calculation method. It is assumed that a total of 8 fiber optic acoustic sensors are arranged, with spatial coordinates as follows: sensor 1 (100mm, 50mm, 100mm), sensor 2 (500mm, 50mm, 100mm), sensor 3 (500mm, 350mm, 100mm), sensor 4 (100mm, 350mm, 100mm), sensor 5 (100mm, 50mm, 300mm), sensor 6 (500mm, 50mm, 300mm), sensor 7 (500mm, 350mm, 300mm), and sensor 8 (100mm, 350mm, 300mm). For the initial positioning coordinates (328.1mm, 206.3mm, 384.4mm), the spatial distances to each sensor are calculated, resulting in the following distances: first distance 357.6mm, second distance 325.4mm, third distance 320.7mm, fourth distance 353.3mm, fifth distance 289.8mm, sixth distance 251.1mm, seventh distance 242.5mm, and eighth distance 282.0mm.

[0079] The theoretical propagation time was calculated based on a preset wave velocity value, initially set at 1500 m / s, which is the typical propagation speed of sound waves in the insulation medium of a cable joint. The theoretical propagation time was calculated by dividing the spatial distance by the wave velocity. For the aforementioned eight sensors, the theoretical propagation times were: first propagation time 238.4 μs, second propagation time 216.9 μs, third propagation time 213.8 μs, fourth propagation time 235.5 μs, fifth propagation time 193.2 μs, sixth propagation time 167.4 μs, seventh propagation time 161.7 μs, and eighth propagation time 188.0 μs. The actual arrival time was extracted from the stress wave timing characteristics, and a threshold cross-validation method combined with energy centroid positioning was used. Based on the stress wave timing characteristics obtained from the previous processing, the actual arrival times of each sensor are extracted as follows: first arrival time 240.2 μs, second arrival time 215.5 μs, third arrival time 216.4 μs, fourth arrival time 233.8 μs, fifth arrival time 191.5 μs, sixth arrival time 169.2 μs, seventh arrival time 159.8 μs, and eighth arrival time 190.3 μs.

[0080] The time difference between the theoretical propagation time and the actual arrival time is calculated as the time residual. The time residual is calculated by subtracting the theoretical propagation time from the actual arrival time. For the eight sensors, the time residuals are as follows: First residual 1.8 μs, second residual -1.4 μs, third residual 2.6 μs, fourth residual -1.7 μs, fifth residual -1.7 μs, sixth residual 1.8 μs, seventh residual -1.9 μs, and eighth residual 2.3 μs. The time residual reflects the positioning accuracy; the smaller the residual, the more accurate the positioning.

[0081] A state vector is constructed, containing three-dimensional coordinate components of the initial positioning coordinates and a preset wave velocity value. The state vector is a four-dimensional vector, with the first three components being spatial coordinates and the last component being the wave velocity value. For the initial positioning coordinates (328.1mm, 206.3mm, 384.4mm) and a wave velocity of 1500m / s, the state vector is represented as [328.1, 206.3, 384.4, 1500]. It is then determined whether the time residual is less than a preset time threshold, set to 1.5μs. The root mean square value of the time residuals from the eight sensors is calculated, yielding a result of 1.84μs, which is greater than the preset threshold of 1.5μs, indicating that further correction of the positioning coordinates is needed.

[0082] A random perturbation is applied to the state vector, and the time residuals are recalculated to obtain the Jacobian matrix. The amplitude of the random perturbation is set to 0.1% of the coordinate components and 0.1% of the wave velocity. A perturbation of +0.328 mm is applied to the abscissa, and the state vector becomes [328.428, 206.3, 384.4, 1500]. The time residuals of the eight sensors are recalculated, yielding a first residual of 1.7 μs, a second residual of -1.5 μs, a third residual of 2.5 μs, and so on. The first column of the Jacobian matrix is ​​calculated as (first residual change / 0.328), (second residual change / 0.328), etc., resulting in [-0.3049, -0.3049, -0.3049, -0.3049, -0.3049, -0.3049, -0.3049, -0.3049, -0.3049]. Perturbations are applied to the ordinate, elevation coordinate, and wave velocity value respectively to obtain the remaining columns of the Jacobian matrix. Finally, an 8×4 Jacobian matrix is ​​obtained, which represents the sensitivity of the time residual to each component of the state vector.

[0083] The gradient vector is obtained based on the transpose of the Jacobian matrix and the time residual. The gradient vector is calculated by multiplying the transpose of the Jacobian matrix by the time residual vector. For an 8×4 Jacobian matrix and an 8-dimensional time residual vector, a 4-dimensional gradient vector is obtained. In actual calculations, the gradient vector is [0.243, 0.187, 0.315, -0.002], representing the adjustment direction and magnitude of each state component. A damping matrix is ​​constructed based on the Jacobian matrix and a preset damping coefficient, initially set to 0.01. The damping matrix is ​​calculated by multiplying the transpose of the Jacobian matrix by the Jacobian matrix, then adding the identity matrix multiplied by the damping coefficient. For an 8×4 Jacobian matrix and a damping coefficient of 0.01, a 4×4 damping matrix is ​​obtained.

[0084] A system of linear equations is constructed based on the damping matrix and gradient vector, and the correction increment is obtained by solving it. The linear equations are the result of multiplying the damping matrix by the correction increment, which equals the gradient vector, where the correction increment is the variable to be solved. The system of linear equations is solved using Gaussian elimination, yielding the correction increment [-0.827, -0.631, -1.058, 6.735]. The updated state vector is obtained by updating the state vector based on the correction increment, which is the original state vector plus the correction increment. For the original state vector [328.1, 206.3, 384.4, 1500] and the correction increment [-0.827, -0.631, -1.058, 6.735], the updated state vector [327.273, 205.669, 383.342, 1506.735] is obtained.

[0085] The root mean square (RMS) of the current time residual is calculated based on the updated state vector and used as the updated residual RMS value, with the damping coefficient adjusted accordingly. The theoretical propagation time and time residual are recalculated using the updated state vector to obtain the updated residual vector. The calculated RMS value of the updated residual vector is 1.59 μs, which is less than the original residual RMS value of 1.84 μs but greater than the threshold of 1.5 μs, requiring further iteration. If the updated residual RMS value decreases, the damping coefficient is decreased (multiplied by 0.5); if it increases, the damping coefficient is increased (multiplied by 2). Due to the decrease in residual value, the damping coefficient is updated to 0.005.

[0086] The iteration is repeated until the root mean square value of the updated residual is less than the preset residual threshold of 1.5 μs. In practical applications, after 5 iterations, the state vector converges to [326.8, 204.7, 381.5, 1512.3], with a corresponding root mean square value of 1.32 μs, which is less than the threshold of 1.5 μs. The corrected positioning coordinates (326.8 mm, 204.7 mm, 381.5 mm) and the corrected wave velocity value of 1512.3 m / s are obtained as the final positioning result output.

[0087] In this embodiment, the theoretical propagation time is obtained by calculating the spatial distance from the initial positioning coordinates to each sensor and combining it with the preset wave velocity. The time residual is then compared with the actual arrival time to obtain the time residual, thereby realizing the physical consistency verification of the spatial positioning results. This can promptly detect the offset caused by noise, local peak misjudgment, or spatial probability density field error, so that the positioning results no longer depend solely on the spatial probability distribution and can simultaneously meet the physical constraints of wave propagation, thus improving the reliability of the positioning results. By introducing random perturbation and Jacobian matrix calculation, the system can maintain sensitivity to the error direction even in local nonlinear regions or when the initial value deviation is large, avoiding getting trapped in local minima or experiencing convergence stagnation. By constructing a damping matrix based on the Jacobian matrix and damping coefficient to solve for the increment, the optimization process can still maintain stable convergence even when there is a lot of noise and complex local curvature.

[0088] Figure 2 This is a flowchart illustrating the positioning logic of the fiber optic grating acoustic wave intelligent sensing and detection method for partial discharge of cable joints according to an embodiment of the present invention.

[0089] A second aspect of the present invention provides a fiber Bragg grating acoustic wave intelligent sensing and detection system for partial discharge of cable joints, comprising: The first unit is used to acquire multi-channel time-domain spectral response data corresponding to the cable joint under test through a fiber optic grating sensing array and demodulate it based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The second unit is used to decompose the waveform signal in the multidimensional feature vector group into narrowband intrinsic mode components and extract the energy density through Hilbert transform to construct an energy density matrix. The unit then performs singular value decomposition on the energy density matrix and selects the dominant mode components to reconstruct the waveform signal to obtain the stress wave time series characteristics. The third unit uses the high-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves as the hyperedge weights, takes the time-frequency amplitude feature vector in the time-series features of stress waves as the initial node features, constructs a spatiotemporal hypergraph based on the hyperedge weights and the initial node features, generates a hyperedge representation by aggregating the initial node features in the spatiotemporal hypergraph through graph convolution, and obtains a node embedding vector by aggregating the hyperedge representation and aggregating the order neighborhood. The fourth unit is used to initialize a Gaussian noise field based on the node embedding vector and set condition information, perform inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identify local maxima points as initial positioning coordinates, calculate the theoretical propagation time based on the preset wave velocity value and calculate the residual in combination with the stress wave time series characteristics, and iteratively correct the initial positioning coordinates and the wave velocity value until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

[0090] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0091] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0092] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent sensing and detection of partial discharge in cable joints using fiber optic gratings and acoustic waves, characterized in that, include: The multi-channel time-domain spectral response data of the cable joint under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix and the dominant mode components are selected to reconstruct the waveform signal to obtain the stress wave time series characteristics. The higher-order cross-correlation tensor of the waveform of the stress wave time series features is used as the hyperedge weight. The time-frequency amplitude feature vector in the stress wave time series features is used as the initial node feature. A spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature. The initial node feature is aggregated through graph convolution in the spatiotemporal hypergraph to generate a hyperedge representation. The hyperedge representation is aggregated and the node embedding vector is obtained through order neighborhood aggregation. The Gaussian noise field is initialized based on the node embedding vector and condition information is set. Inverse sampling denoising is performed based on the condition information to generate an initial three-dimensional spatial probability density field and local maxima points are identified as initial positioning coordinates. The theoretical propagation time is calculated based on the preset wave velocity value and the residual is calculated in combination with the stress wave time series characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

2. The method according to claim 1, characterized in that, The multi-channel time-domain spectral response data of the cable connector under test is acquired by a fiber optic grating sensor array and demodulated based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector set, including: The grating period change caused by stress wave generated by partial discharge is collected by a fiber optic grating sensor array uniformly distributed along the circumference of the cable joint under test. The Bragg wavelength drift of each grating node is recorded in real time as a function of time to obtain multi-channel time-domain spectral response data. Baseline drift correction is performed on the multi-channel time-domain spectral response data. Wavelength abrupt changes exceeding the standard deviation of background noise in each channel are identified as the start time of the effective signal. Wavelength drift sequence within the effective signal time window is extracted. The wavelength drift of each sampling point in the wavelength drift sequence is converted into stress wave amplitude value through the preset sensitivity coefficient, and the stress wave time domain waveform signal corresponding to each grating node is constructed. Combined with the spatial coordinate information of each grating node, a spatiotemporally correlated waveform signal group is formed. Peak amplitude is extracted by peak detection of the time-domain waveform of each channel in the waveform signal group. The arrival time difference of each channel relative to the reference channel is determined by cross-correlation analysis and the arrival time of each channel is calculated by combining the absolute arrival time of the reference channel. Short-time Fourier transform is performed on the time-domain waveform of each channel to obtain the time spectrum and the spectral centroid is calculated. The waveform signal group, the arrival time, the peak amplitude and the spectral centroid are combined into a multidimensional feature vector group.

3. The method according to claim 1, characterized in that, The waveform signal in the multidimensional feature vector group is decomposed into narrowband intrinsic mode components, and the energy density is extracted by Hilbert transform to construct an energy density matrix. Singular value decomposition is performed on the energy density matrix, and the dominant mode components are selected to reconstruct the waveform signal, obtaining the stress wave time series features, including: Initialize the candidate mode components corresponding to the waveform signal in the multidimensional feature vector group and minimize the spectral bandwidth to obtain the preliminary decomposition result. Determine the frequency trajectory based on the preliminary decomposition result, calculate the similarity of the frequency trajectories between different channels and determine similar mode pairs by combining the preset similarity threshold, and perform cross-correlation analysis and time compensation on the similar mode pairs to obtain narrowband intrinsic mode components. The narrowband intrinsic mode components are subjected to Hilbert transform to obtain analytic signals, and the instantaneous amplitude is extracted as the instantaneous energy density. The instantaneous energy density is arranged according to the time sampling points to form a time series and organized to obtain an energy density matrix. Singular value decomposition is performed on the energy density matrix to obtain a sequence of singular values ​​and corresponding left and right vector matrices. A time pattern is extracted from the left vector matrix and a time metric is calculated. A spatial pattern is extracted from the right vector matrix and a spatial metric is calculated. A smoothness score is calculated based on the time and spatial metrics, and a screening index is determined. The singular value sequence is screened based on the screening index to determine the dominant singular values ​​and reconstruct a denoised energy density matrix. An inverse Hilbert transform is performed on the denoised energy density matrix to obtain a time-domain waveform. The time-domain waveform is summed and superimposed, and time-frequency amplitude feature parameters are extracted to obtain stress wave time series features.

4. The method according to claim 1, characterized in that, The higher-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves is used as the hyperedge weight. The time-frequency amplitude feature vector in the time-series features of stress waves is used as the initial node feature. The spatiotemporal hypergraph is constructed based on the hyperedge weight and the initial node feature, including: The time-frequency amplitude feature vector in the stress wave time series features is divided into multiple time segments according to the time window. The time-frequency amplitude feature vector in each time segment is constructed as a graph node. Three graph nodes are randomly selected from the graph nodes of the current time segment to obtain a candidate hyperedge node group. The third-order tensor outer product operation is performed on the time-frequency amplitude feature vector corresponding to the candidate hyperedge node group and the tensor norm is extracted as the cooperative strength. The cooperative strength is normalized to obtain the hyperedge weight. The time-domain peak amplitude, dominant frequency, and energy centroid time position are extracted from the stress wave time-series features and spliced ​​together to form the initial node features of the graph nodes; A hyperedge correlation matrix is ​​constructed based on the hyperedge weights, and a node feature matrix is ​​constructed based on the initial node feature vectors. A node indirect matrix is ​​obtained by multiplying the transpose of the hyperedge correlation matrix with the hyperedge correlation matrix. The node indirect matrix is ​​multiplied with the node feature matrix and a nonlinear transformation is performed to obtain updated node features. The hyperedge correlation matrices corresponding to each time segment are concatenated along the time dimension to obtain a spatiotemporal hyperedge matrix. A spatiotemporal hypergraph is constructed by combining the updated node features.

5. The method according to claim 1, characterized in that, In the spatiotemporal hypergraph, the initial node features are aggregated through graph convolution to generate hyperedge representations. Aggregating these hyperedge representations and then aggregating them through order-neighbor aggregation yields node embedding vectors, including: Obtain the graph nodes in the spatiotemporal hypergraph, calculate the feature covariance matrix of the initial node features corresponding to each graph node, and extract the principal component directions as the intrinsic manifold basis of the hyperedge. Project the initial node features onto the intrinsic manifold basis to obtain the manifold coordinate representation and calculate the manifold distance between graph nodes. Based on the manifold distance, construct the adaptive adjacency weights within the hyperedge and perform manifold-aware weighted aggregation on the initial node features. Combine the hyperedge weights with multiplication gating to obtain the hyperedge representation. Extract the set of associated hyperedges corresponding to each graph node and calculate the mutual information between the hyperedge representation corresponding to each associated hyperedge and the hyperedge representation corresponding to the current graph node. Based on the mutual information, filter the hyperedge representations in the set of associated hyperedges to obtain the first-level node features. Based on the first-level node features, a Laplacian matrix is ​​constructed and a graph filter is obtained by extracting low-frequency feature vectors through spectral decomposition. The first-level node features corresponding to each graph node are then subjected to frequency domain filtering and polynomial expansion through the graph filter to obtain multi-order neighborhood aggregation coefficients. Based on the multi-order neighborhood aggregation coefficients, the hyperedge representation is weighted and aggregated to obtain the node embedding vector.

6. The method according to claim 1, characterized in that, Initializing a Gaussian noise field based on the node embedding vector and setting condition information, performing inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identifying local maxima points as initial positioning coordinates includes: The node embedding vector is subjected to feature encoding and spatial broadcasting to obtain spatial conditional features. A three-dimensional Gaussian noise field following a standard normal distribution is initialized and concatenated with the corresponding spatial conditional features to obtain a conditional noise field. The total time step is set and the current time step is initialized. The conditional noise field and the time encoding corresponding to the current time step are subjected to convolution transformation to obtain noise components. The noise components are subtracted from the conditional noise field representation and a preset scaling factor is added to obtain a denoised field representation. The denoised field representation is used as the conditional noise field of the next time step and the time step is decremented. This process is repeated until the time step is zero to obtain the initial three-dimensional spatial probability density field. Based on a preset probability density threshold, the grid points in the initial three-dimensional spatial probability density field are filtered to obtain an initial grid point set. Initial grid points with probability density values ​​greater than all adjacent grid points are identified as candidate maxima. The gradient magnitude between the probability density value of the candidate maxima and the probability density values ​​of the corresponding adjacent grid points is calculated. The candidate maxima are sorted in descending order according to the gradient magnitude, and the first candidate maxima is selected to extract the coordinates in three-dimensional space to obtain the initial positioning coordinates.

7. The method according to claim 1, characterized in that, The theoretical propagation time is calculated based on a preset wave velocity value, and the residual is calculated in conjunction with the stress wave timing characteristics. The initial positioning coordinates and the wave velocity value are iteratively corrected until the root mean square of the residual is less than a preset residual threshold to obtain the corrected positioning coordinates. Calculate the spatial distance between the initial positioning coordinates and the positions of each sensor, calculate the theoretical propagation time in combination with the preset wave velocity value, extract the actual arrival time from the stress wave timing characteristics, and calculate the time difference between the theoretical propagation time and the actual arrival time as the time residual. Construct a state vector containing the three-dimensional coordinate components of the initial positioning coordinates and the preset wave velocity value, determine whether the time residual is less than a preset time threshold, and if it is less, output the initial positioning coordinates in the state vector as the corrected positioning coordinates. If the value is not less than the threshold value, a random perturbation is applied to the state vector, and the time residual is recalculated to obtain the Jacobian matrix. The gradient vector is calculated based on the transpose of the Jacobian matrix and the time residual. A damping matrix is ​​constructed based on the Jacobian matrix and a preset damping coefficient. A system of linear equations is constructed based on the damping matrix and the gradient vector, and the correction increment is obtained by solving the system. The state vector is updated based on the correction increment to obtain the updated state vector. The root mean square of the current time residual is calculated based on the updated state vector as the root mean square value of the updated residual, and the damping coefficient is corrected. This process is repeated until the root mean square value of the updated residual is less than a preset residual threshold value. The corrected positioning coordinates are then obtained and output.

8. A fiber optic grating acoustic wave intelligent sensing and detection system for partial discharge of cable joints, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire multi-channel time-domain spectral response data corresponding to the cable joint under test through a fiber optic grating sensing array and demodulate it based on a preset sensitivity coefficient to obtain a multi-dimensional feature vector group. The second unit is used to decompose the waveform signal in the multidimensional feature vector group into narrowband intrinsic mode components and extract the energy density through Hilbert transform to construct an energy density matrix. The unit then performs singular value decomposition on the energy density matrix and selects the dominant mode components to reconstruct the waveform signal to obtain stress wave time series features. The third unit uses the high-order cross-correlation tensor of the waveform used to calculate the time-series features of stress waves as the hyperedge weights, takes the time-frequency amplitude feature vector in the time-series features of stress waves as the initial node features, constructs a spatiotemporal hypergraph based on the hyperedge weights and the initial node features, generates a hyperedge representation by aggregating the initial node features in the spatiotemporal hypergraph through graph convolution, and obtains a node embedding vector by aggregating the hyperedge representation and aggregating the order neighborhood. The fourth unit is used to initialize a Gaussian noise field based on the node embedding vector and set condition information, perform inverse sampling denoising based on the condition information to generate an initial three-dimensional spatial probability density field and identify local maxima points as initial positioning coordinates, calculate the theoretical propagation time based on the preset wave velocity value and calculate the residual in combination with the stress wave time series characteristics, and iteratively correct the initial positioning coordinates and the wave velocity value until the root mean square of the residual is less than the preset residual threshold to obtain the corrected positioning coordinates.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images