Fault detection and early warning method and system for insulating layer of oil-immersed high-voltage wiring harness

Abstract

The invention provides a fault detection and early warning method and system for an insulating layer of an oil-immersed high-voltage wiring harness, and relates to the technical field of power equipment detection, and the method comprises the steps: obtaining a multi-physical field coupling signal, carrying out the time-frequency decomposition, obtaining degradation feature distribution, extracting a feature response amplitude, calculating a gradient change rate, and recognizing an abnormal feature cluster; deducing residual life distribution of the insulating medium according to the degradation rate and generating a risk index; according to the risk index, the early warning level and the intervention time window are determined, early detection and early warning of the fault of the insulating layer of the oil-immersed high-voltage wire harness can be realized, and the operation safety of power equipment is improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, and in particular to a method and system for detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses. Background Technology

[0002] As an important component of the power system, oil-immersed high-voltage line harnesses have a significant impact on the safe and stable operation of the entire system. The insulation layer of oil-immersed high-voltage line harnesses will gradually age and deteriorate due to the combined effects of multiple factors such as electric field, thermal field, and mechanical stress, which may lead to insulation breakdown faults, causing serious safety accidents and economic losses. Traditional oil-immersed high-voltage wiring harness insulation condition monitoring mainly relies on periodic offline testing or simple online monitoring methods. Offline testing usually requires shutdown operations, which not only affects the normal operation of the system, but also cannot reflect the actual degradation of the equipment under operating conditions in real time. Online monitoring mainly relies on single parameter monitoring, such as partial discharge or temperature, and lacks comprehensive consideration of multi-physics coupling effects. It still has problems such as difficulty in accurately locating the insulation degradation process, lack of multi-dimensional analysis capability of insulation degradation characteristics, difficulty in effectively identifying early and hidden faults, and lack of accurate assessment of remaining life and risk classification mechanism. It is also impossible to dynamically assess the remaining life of the insulation layer based on factors such as degradation rate and operating conditions. Summary of the Invention

[0003] This invention provides a method and system for detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of the present invention provides a method for detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses, comprising: The multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions is obtained and multi-scale time-frequency decomposition is performed to obtain the degradation feature distribution. The degradation feature distribution is spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. Extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain the fault candidate set. For each abnormal region in the fault candidate set, the degradation rate is calculated based on the gradient change rate and the remaining lifetime distribution of the insulating medium is deduced. The remaining lifetime distribution is then coupled with the dynamic load curve of the operating condition in a time series to generate a hierarchical risk quantification index. Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and the warning information is output.

[0005] In one alternative implementation, The degradation feature distribution is obtained by acquiring the multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions and performing multi-scale time-frequency decomposition. The degradation feature distribution is then spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution, including: The electric field and temperature components of the oil-immersed high-voltage line harness under operating conditions are obtained and combined to obtain a multi-physics field coupled signal; The multi-physics coupling signal is decomposed into time-frequency components at multiple time window scales and multiple frequency resolutions. Feature components reflecting the degradation state of the insulating medium at each scale are extracted. The information entropy difference between feature components at different scales is calculated. Based on the information entropy difference, the feature components at each scale are weighted and fused to obtain the degradation feature distribution. Obtain the geometric topology location information of the oil-immersed high-voltage harness, match the location identifier in the degradation feature distribution with the geometric topology location information to obtain a location transformation matrix, and map the degradation feature distribution to the corresponding geometric topology location based on the location transformation matrix to obtain an aligned degradation feature distribution.

[0006] In one alternative implementation, Extract the feature response amplitude at each position in the alignment degradation feature distribution, calculate the gradient rate of change of the feature response amplitude in the spatial dimension, and match the gradient rate of change with the aging evolution path constraints to obtain the initial candidate positions, including: Traverse the spatial locations in the alignment degradation feature distribution and extract the corresponding feature response amplitudes; Calculate the local geometric topological curvature corresponding to each spatial location and initialize the sampling direction based on the local geometric topological curvature. Construct an adaptive direction sampling set in the multi-scale neighborhood corresponding to the current location based on the sampling direction. Calculate the directional derivative between the feature response amplitude of the current location and the feature response amplitude of the neighborhood locations on different sampling directions based on the adaptive direction sampling set. Extract the gradient component corresponding to the dominant adaptive direction based on the directional derivative and use the gradient component as the gradient change rate of the current location in the spatial dimension. Extract the multidimensional temporal evolution trajectory corresponding to the gradient change rate from the pre-acquired historical degradation data and map it to the continuous cohomology space. Calculate the topological invariants of the multidimensional temporal evolution trajectory in the continuous cohomology space and construct historical topological signatures based on the topological invariants. Calculate the similarity between different historical topological signatures and determine similar historical topological signatures. Aggregate multidimensional temporal evolution trajectories with similar historical topological signatures and obtain aging evolution path constraints by identifying invariant manifolds. Calculate the topological signature corresponding to each spatial location, calculate the Wasserstein distance metric between the topological signature and the aging evolution path constraint, and filter the initial candidate locations by combining the results with a pre-set adaptive distance threshold.

[0007] In one alternative implementation, A multidimensional feature space is constructed based on the feature response amplitudes of the initial candidate locations. Abnormal feature clusters are identified in this multidimensional feature space using a clustering separation algorithm. The anomalous feature clusters are then aggregated to obtain a fault candidate set, including: Extract the feature response magnitude and local manifold deviation corresponding to the initial candidate position and combine them to form an extended feature vector. Construct a multidimensional feature space based on the extended feature vector. In the multidimensional feature space, a nearest neighbor association graph is constructed between the initial candidate positions. The first manifold geodesic distance corresponding to each edge in the nearest neighbor association graph is calculated. The local density field intensity of each initial candidate position is calculated based on the first manifold geodesic distance. Density peak points are identified based on the local density field intensity and the density peak points are used as cluster centers. The second manifold geodesic distance from other initial candidate positions to each density peak point is calculated. The initial candidate positions are assigned to the candidate clusters corresponding to the nearest density peak points based on the second manifold geodesic distance. For each candidate cluster, the local manifold deviation of the initial candidate position in the candidate cluster is extracted, the first statistical moment feature corresponding to the local manifold deviation in the current candidate cluster is calculated, the second statistical moment feature corresponding to the local manifold deviation of the pre-constructed normal evolution mode is extracted, the distribution distance between the first statistical moment feature and the second statistical moment feature is calculated, the candidate clusters whose distribution distance exceeds a preset distance threshold are identified as anomalous feature clusters, and the initial candidate positions in the anomalous feature clusters are aggregated to obtain a fault candidate set.

[0008] In one alternative implementation, For each anomalous region in the fault candidate set, calculating the degradation rate based on the gradient change rate and extrapolating the remaining lifetime distribution of the insulating medium includes: The temperature field distribution and electric field distribution corresponding to the abnormal region are obtained, and the temperature field strength value and electric field strength value at each location are determined. The thermal sensitivity mapping relationship of the temperature field strength value to the gradient change rate and the electrical sensitivity mapping relationship of the electric field strength value to the gradient change rate are determined. Based on the thermal sensitivity mapping relationship and the electrical sensitivity mapping relationship, the gradient change rate is decoupled to obtain the decoupled gradient change rate. The baseline gradient change rate corresponding to the abnormal region at the historical detection time is obtained. The dynamic evolution trajectory of the decoupled gradient change rate relative to the baseline gradient change rate is determined. The nonlinear acceleration factor in the dynamic evolution trajectory is extracted and quantified to obtain the degradation rate. The critical failure threshold of the insulating medium is extracted from the pre-acquired historical degradation data. The state space distance of the decoupling gradient change rate evolving towards the critical failure threshold is determined. The theoretical failure time is obtained based on the state space distance and the degradation rate. The decoupling gradient change rate corresponding to the historical degradation data is extracted and Fourier spectrum decomposition is performed to obtain the dominant frequency component. The amplitude envelope is calculated based on the dominant frequency component, and the time-domain broadening coefficient is extracted to extend the uncertainty of the theoretical failure time, thereby obtaining the failure time confidence interval. The remaining lifetime of the abnormal region is obtained by calculating the expected value. The remaining lifetime of all abnormal regions is summarized to obtain the remaining lifetime distribution of the insulating medium.

[0009] In one alternative implementation, The remaining life distribution is coupled with the dynamic load curve of the operating condition in a time series calculation to generate hierarchical risk quantification indicators, including: Obtain the dynamic load curve of the operating condition and extract the time-series characteristic parameters corresponding to load fluctuations; Establish a time-domain mapping relationship between the remaining lifetime of each anomalous region in the remaining lifetime distribution and the time-series characteristic parameters. Based on the time-domain mapping relationship, determine the dynamic loss amount of the dynamic load curve on the remaining lifetime. Superimpose the dynamic loss amount onto the current degradation state of the corresponding anomalous region in the remaining lifetime distribution to obtain the load-coupled degradation evolution trajectory. Extract the probability density function of the degradation state reaching the critical failure threshold from the load-coupled degradation evolution trajectory. Calculate the time-domain failure probability distribution based on the probability density function. The time-domain failure probability distribution is mapped to phase space and a phase space trajectory is constructed. The rate of change of curvature of the trajectory curvature in the phase space trajectory is extracted. Based on the rate of change of curvature, the inflection point of risk evolution is identified. The time-domain failure probability distribution is divided into multiple risk evolution stages by using the inflection point as the boundary. The variance of the time-domain failure probability distribution in each risk evolution stage is calculated. Based on the variance, the risk quantification index of each stage is calculated. The risk quantification indexes of all risk evolution stages are summarized to generate a hierarchical risk quantification index.

[0010] In one alternative implementation, Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and warning information is output, including: Obtain the preset safety margin constraint and extract the risk level classification threshold. Compare the risk quantification index of each risk evolution stage in the hierarchical risk quantification index with the risk level classification threshold to determine the fault warning level corresponding to each risk evolution stage. Extract the time domain range of the risk evolution stage corresponding to the fault warning level, take the start time of the time domain range as the intervention start time, take the end time of the time domain range as the intervention end time, and determine the intervention time window corresponding to the fault warning level based on the intervention start time and the intervention end time. The fault warning level, the intervention time window, and the risk quantification index are encapsulated to generate warning information and output.

[0011] A second aspect of the present invention provides a fault detection and early warning system for the insulation layer of oil-immersed high-voltage wire harnesses, comprising: The first unit is used to acquire the multi-physics field coupling signal of the oil-immersed high-voltage line harness in operation and perform multi-scale time-frequency decomposition to obtain the degradation feature distribution, and spatially align the degradation feature distribution with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. The second unit is used to extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain a fault candidate set. The third unit is used to calculate the degradation rate and deduce the remaining lifetime distribution of the insulating medium for each abnormal region in the fault candidate set based on the gradient change rate, and to perform time-series coupling calculation of the remaining lifetime distribution with the dynamic load curve of the operating condition to generate hierarchical risk quantification indicators. The fourth unit is used to determine the fault warning level and the corresponding intervention time window based on the risk quantification index and the preset safety margin constraint, and output the warning information.

[0012] 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.

[0013] 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.

[0014] This invention achieves precise localization of insulation degradation characteristics in oil-immersed high-voltage line harnesses through time-frequency decomposition and spatial alignment of multi-physics field coupled signals, improving the accuracy and sensitivity of fault identification. By establishing a fault candidate screening mechanism based on gradient change rate and aging evolution path constraints, combined with a clustering and separation algorithm for multi-dimensional feature space, it can effectively distinguish between normal insulation degradation and abnormal fault modes, reducing false alarm rates and enhancing the reliability and targeting of detection. Based on the time-series coupling calculation of remaining lifetime distribution and dynamic load curves, hierarchical risk quantification indicators are generated, achieving a technological leap from simple fault detection to predictive maintenance. This provides a scientific basis for operation and maintenance decisions, extends equipment lifespan, reduces the risk of sudden faults, and improves the safe operation level of the power system. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method for detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the remaining life assessment of the oil-immersed high-voltage wire harness insulation layer fault detection and early warning method according to an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] Figure 1 This is a flowchart illustrating the method for detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses according to an embodiment of the present invention. Figure 1 As shown, the method includes: The multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions is obtained and multi-scale time-frequency decomposition is performed to obtain the degradation feature distribution. The degradation feature distribution is spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. Extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain the fault candidate set. For each abnormal region in the fault candidate set, the degradation rate is calculated based on the gradient change rate and the remaining lifetime distribution of the insulating medium is deduced. The remaining lifetime distribution is then coupled with the dynamic load curve of the operating condition in a time series to generate a hierarchical risk quantification index. Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and the warning information is output.

[0019] In one alternative implementation, The degradation feature distribution is obtained by acquiring the multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions and performing multi-scale time-frequency decomposition. The degradation feature distribution is then spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution, including: The electric field and temperature components of the oil-immersed high-voltage line harness under operating conditions are obtained and combined to obtain a multi-physics field coupled signal; The multi-physics coupling signal is decomposed into time-frequency components at multiple time window scales and multiple frequency resolutions. Feature components reflecting the degradation state of the insulating medium at each scale are extracted. The information entropy difference between feature components at different scales is calculated. Based on the information entropy difference, the feature components at each scale are weighted and fused to obtain the degradation feature distribution. Obtain the geometric topology location information of the oil-immersed high-voltage harness, match the location identifier in the degradation feature distribution with the geometric topology location information to obtain a location transformation matrix, and map the degradation feature distribution to the corresponding geometric topology location based on the location transformation matrix to obtain an aligned degradation feature distribution.

[0020] The electric field component of the oil-immersed high-voltage wiring harness under operating conditions was acquired using a distributed sensor array. The sensor array consisted of at least eight electric field strength sensors, uniformly distributed around the harness. Each sensor recorded local electric field strength changes at a sampling frequency of 100 Hz. The temperature component was acquired by using an infrared thermal imager to thermally image the surface of the harness at a resolution of 640 × 480 pixels, a temperature measurement accuracy of ±0.5℃, and a sampling frequency of 10 Hz. The electric field and temperature components were combined into a multi-physics coupled signal, which was then processed using time alignment and data normalization. Specifically, the sampling rate of the temperature component is increased to 100Hz through linear interpolation to match the sampling rate of the electric field component. Then, the electric field and temperature components are standardized to normalize their numerical ranges to the [0, 1] interval. The standardized electric field and temperature components are then linearly combined with a weight ratio of 6:4 to form a multi-physics coupling signal. This weight ratio is based on statistical analysis of a large amount of experimental data. Specifically, comparative experiments were conducted on 500 samples of oil-immersed high-voltage wiring harnesses under different operating conditions to analyze the sensitivity of the physical field coupling signal to insulation degradation characteristics under different weight ratios. Based on the principle of maximum information entropy, a weight ratio of 6:4 was calculated to maximize the retention of effective information from both physical field signals. Therefore, the weight ratio was set to 6:4. It should be noted that the weight ratio can be modified according to actual needs due to different application scenarios and actual production tasks.

[0021] Multi-scale time-frequency decomposition was performed on the multi-physics coupled signal using wavelet packet transform. The Daubechies wavelet basis function was selected as the mother wavelet, and the decomposition level was set to 5 levels, forming 32 frequency bands. The time window scales were set to 100ms, 500ms, 1s, 5s, and 10s, corresponding to frequency resolutions of 10Hz, 2Hz, 1Hz, 0.2Hz, and 0.1Hz. At each scale, feature components reflecting the degradation state of the insulating medium were extracted. These feature components include energy concentration, frequency center shift, and phase consistency. Energy concentration represents the degree of concentration of signal energy in a specific frequency band; frequency center shift represents the shift of the signal spectrum center relative to the normal state; and phase consistency represents the stability of the signal phase changes.

[0022] The information entropy difference between feature components at different scales is calculated using the relative entropy calculation method. For each feature component, the corresponding probability distribution density function is first calculated, and then the corresponding information entropy value is calculated. The probability distribution density function of the feature component is obtained through kernel density estimation, using a Gaussian kernel function with a bandwidth parameter set to 0.15. The information entropy difference is defined as the absolute difference in information entropy between two adjacent scale feature components. For example, for the energy concentration feature, the information entropy is 2.35 at the 100ms scale and 2.87 at the 500ms scale; therefore, the information entropy difference between these two scales is 0.52.

[0023] Based on the information entropy difference, feature components at each scale are weighted and fused using an adaptive weight allocation strategy. A larger information entropy difference indicates more effective information contained in the feature component at that scale, and thus a larger weight is assigned. Specifically, the weight calculation uses the Softmax function to convert the information entropy difference into weight values ​​within the interval [0, 1], ensuring that the sum of all weights is 1. For the aforementioned example, if the information entropy differences corresponding to the five time window scales are 0.52, 0.38, 0.45, 0.29, and 0.33, respectively, then the calculated weights are 0.28, 0.19, 0.24, 0.13, and 0.16, respectively. The feature components at each scale are then linearly combined according to their corresponding weights to obtain the degraded feature distribution.

[0024] The geometric topological location information of the oil-immersed high-voltage power line harness was acquired, and a high-precision geometric model was established using 3D laser scanning technology. The scanning resolution was 0.5 mm, and the point cloud data density was no less than 100 points per square centimeter. The acquired point cloud data was denoised and meshed to construct a 3D surface model of the power line harness. The location identifiers in the degradation feature distribution, including sensor numbers and relative coordinate information, needed to be matched with the geometric topological location information.

[0025] The calculation of the position transformation matrix is ​​based on the principle of rigid body transformation, including translation and rotation. The translation part is determined by the coordinate difference of corresponding points, and the rotation part is solved by the least squares method to obtain the optimal rotation matrix. Specifically, at least four feature points are selected, which exist simultaneously in the location markers of the degraded feature distribution and the geometric topology model. The coordinate correspondence is established, and the parameters of the transformation matrix are solved. The transformation matrix is ​​a 4×4 matrix, containing a 3×3 rotation submatrix and a 3×1 translation vector.

[0026] The degradation feature distribution is mapped to the corresponding geometric topology position based on the position transformation matrix. For each point in the degradation feature distribution, its corresponding position coordinates in the geometric topology model are calculated through the transformation matrix. For the meshed surface model, the nearest neighbor interpolation method is used to assign the degradation feature values ​​to the mesh nodes to obtain the aligned degradation feature distribution.

[0027] In this embodiment, by performing time-frequency decomposition at multiple time window scales and multiple frequency resolutions, and combining the information entropy difference to achieve weighted fusion of feature components, the ability to distinguish and the stability of insulation degradation features are effectively improved. This overcomes the problem of insufficient identification accuracy of single-scale analysis under noise interference. By spatially matching the distribution of degradation features with the geometric topology of the wire harness and completing feature mapping, the alignment of degradation features with the actual physical structure is achieved. This enables accurate spatial positioning of potential insulation degradation areas, providing more reliable technical support for online condition assessment and preventive maintenance of oil-immersed high-voltage wire harnesses.

[0028] In one alternative implementation, Extract the feature response amplitude at each position in the alignment degradation feature distribution, calculate the gradient rate of change of the feature response amplitude in the spatial dimension, and match the gradient rate of change with the aging evolution path constraints to obtain the initial candidate positions, including: Traverse the spatial locations in the alignment degradation feature distribution and extract the corresponding feature response amplitudes; Calculate the local geometric topological curvature corresponding to each spatial location and initialize the sampling direction based on the local geometric topological curvature. Construct an adaptive direction sampling set in the multi-scale neighborhood corresponding to the current location based on the sampling direction. Calculate the directional derivative between the feature response amplitude of the current location and the feature response amplitude of the neighborhood locations on different sampling directions based on the adaptive direction sampling set. Extract the gradient component corresponding to the dominant adaptive direction based on the directional derivative and use the gradient component as the gradient change rate of the current location in the spatial dimension. Extract the multidimensional temporal evolution trajectory corresponding to the gradient change rate from the pre-acquired historical degradation data and map it to the continuous cohomology space. Calculate the topological invariants of the multidimensional temporal evolution trajectory in the continuous cohomology space and construct historical topological signatures based on the topological invariants. Calculate the similarity between different historical topological signatures and determine similar historical topological signatures. Aggregate multidimensional temporal evolution trajectories with similar historical topological signatures and obtain aging evolution path constraints by identifying invariant manifolds. Calculate the topological signature corresponding to each spatial location, calculate the Wasserstein distance metric between the topological signature and the aging evolution path constraint, and filter the initial candidate locations by combining the results with a pre-set adaptive distance threshold.

[0029] For the obtained alignment degradation characteristic distribution of the oil-immersed high-voltage wire harness, a grid traversal method was used to extract the spatial locations and their corresponding characteristic response amplitudes. The surface of the oil-immersed high-voltage wire harness was divided into 2 mm × 2 mm grid cells, with the center of each grid cell serving as a spatial location, resulting in approximately 12,000 discrete sampling points. The characteristic response amplitude at each spatial location represents the severity of insulation degradation at that location, with values ​​ranging from 0 to 1, where 0 indicates no degradation and 1 indicates complete degradation. The extraction process used a trilinear interpolation algorithm to ensure the accuracy of the characteristic response amplitude calculations at the grid node locations.

[0030] The local geometric topological curvature corresponding to each spatial location is calculated using a discrete differential geometry method. Based on point cloud data within a 5 mm radius neighborhood around each location, a quadratic surface is fitted, and the principal curvatures and principal directions are solved. The local geometric topological curvature is characterized by both mean curvature and Gaussian curvature. The mean curvature is the arithmetic mean of the two principal curvatures, and the Gaussian curvature is the product of the two principal curvatures. At a point in the bending part of the wire harness, the calculated mean curvature is 0.15 per millimeter, and the Gaussian curvature is 0.02 per square millimeter, indicating that there are obvious curved surface characteristics at the current location. When determining the initial sampling direction based on the local geometric topological curvature, the two principal directions corresponding to the principal curvatures are selected as the initial sampling directions. For the aforementioned example point, the two principal directions are [0.82, 0.57, 0.03] and [-0.57, 0.82, 0.04], and these two vectors are unit orthogonal vectors.

[0031] An adaptive direction sampling set is constructed using an adaptive angle partitioning strategy. Based on the main direction, 24 sampling directions are generated within the range of 0 to 360 degrees with a step size of 15 degrees, forming a complete direction sampling set. For regions with large curvature (average curvature greater than 0.2 mm), a smaller angle step size of 10 degrees is used to generate 36 sampling directions, improving direction resolution. Within each sampling direction, points within a multi-scale neighborhood are selected centered on the current position for analysis. The multi-scale neighborhood includes three scales: 2 mm, 5 mm, and 10 mm, corresponding to the micro, meso, and macroscopic features of the wire bundle surface.

[0032] When calculating the directional derivative, for each sampling direction, the ratio of the difference between the feature response amplitude at the current position and the feature response amplitude of the neighboring position in that direction to the distance is calculated to obtain the directional derivative in that direction. For example, for a certain sampling point, in the direction [0.82, 0.57, 0.03], the feature response amplitude at the current position is 0.65, and the feature response amplitude of the neighboring point 2 mm away is 0.72, then the directional derivative in that direction is 0.035. Analyzing the directional derivatives in all directions, the direction with the largest absolute value is extracted as the dominant directional direction. If, among the 24 sampling directions, the directional derivative of 0.035 corresponding to the direction [0.82, 0.57, 0.03] is the maximum value, then that direction is the dominant directional direction, and the corresponding directional derivative is the gradient change rate at that position.

[0033] A multidimensional temporal evolution trajectory corresponding to the gradient change rate was extracted from historical degradation data. This was achieved using a historical monitoring database containing monitoring data of oil-immersed high-voltage line harnesses at different degradation stages, spanning two years with a sampling interval of one week. For each historical monitoring location, the gradient change rate over time was extracted to form a multidimensional temporal evolution trajectory. For example, the gradient change rate sequence for a certain monitoring point over 10 consecutive weeks was [0.015, 0.018, 0.022, 0.027, 0.032, 0.037, 0.043, 0.049, 0.056, 0.064], indicating that the degradation rate at this point showed an accelerating trend.

[0034] Multidimensional temporal evolution trajectories are mapped to a persistent cohomology space. Persistent cohomology is a mathematical tool for analyzing the topological properties of data, capable of capturing persistent topological structures within the data. The temporal trajectory is treated as a point cloud in a high-dimensional space, and topological structures at different scales are constructed using Vietoris-Rips complexes to calculate persistent cohomology groups. For each temporal trajectory, 0-dimensional and 1-dimensional persistent cohomology groups are calculated, corresponding to the persistence of connected components and loop structures, respectively. The filtering parameter range for the persistent cohomology calculation is set to [0, 0.2], with a step size of 0.01.

[0035] The topological invariants of a multidimensional temporal evolution trajectory in the persistent homology space are calculated, and statistical features of the persistent homology graph are extracted, including the number of persistent intervals, the length of the longest persistent interval, the mean length of the persistent intervals, and the variance. For example, the 0-dimensional persistent homology graph of a certain temporal trajectory contains 3 persistent intervals, the longest persistent interval has a length of 0.15, the mean length is 0.09, and the variance is 0.0025. These features together constitute the topological signature of the trajectory, used to characterize the topological properties of the temporal data.

[0036] The similarity between different historical topological signatures is calculated using cosine similarity. For two topological signature vectors, their dot product is divided by the product of their corresponding norms to obtain the similarity value. A similarity threshold of 0.85 is set; when the similarity between two topological signatures is greater than this threshold, they are considered similar historical topological signatures. Multidimensional temporal evolution trajectories with similar historical topological signatures are aggregated, and hierarchical clustering is used to group similar trajectories into the same category. By identifying common evolutionary patterns in each category of trajectories, invariant manifolds are extracted to obtain aging evolution path constraints. Specifically, the aging evolution path constraints are implemented by combining mathematical expression calculation with cluster extraction. First, hierarchical clustering is performed on the multidimensional temporal trajectories of similar historical topological signatures. Principal component analysis is performed on the trajectories within each cluster to extract common feature vectors. Finally, an nth-order polynomial function is fitted using the least squares method as the mathematical representation of the invariant manifold. The resulting nth-order polynomial function is the specific mathematical expression for the aging evolution path constraints.

[0037] The Wasserstein distance metric between the topological signature and aging evolution path constraints corresponding to each spatial location is calculated, comparing the topological signature of the current location with various types of aging evolution path constraints. Wasserstein distance is a measure of the difference between probability distributions, effectively capturing structural differences between topological features. The distance calculation uses the EMD algorithm, where n is the feature dimension, and is filtered using a pre-set adaptive distance threshold. The adaptive distance threshold is dynamically determined based on the statistical characteristics of historical data, with a basic range of [0.2, 0.5]. When the Wasserstein distance between the topological signature and a certain type of aging evolution path constraint is less than the threshold, the location is determined as an initial candidate location. For example, the Wasserstein distance between the topological signature of a certain location and the path constraint of the rapid degradation type is 0.18, which is less than the set threshold of 0.25; therefore, this location is selected as an initial candidate location.

[0038] In this embodiment, by traversing spatial locations in the aligned degradation feature distribution and combining local geometric topological curvature for directional adaptive sampling, the local deformation features in the complex geometric structure of the harness can be utilized to achieve directional sensitivity analysis of the feature response amplitude in the spatial dimension, improving the expressive power of degradation features in terms of spatial resolution and directional consistency. By calculating the directional derivatives of different sampling directions and extracting the gradient components corresponding to the dominant evolutionary directions, the ability to capture the evolution trend of local degradation regions is effectively strengthened, overcoming the problem of insufficient identification accuracy in complex structural regions. By mapping the gradient change rate of historical degradation data to the continuous homology space and extracting topological invariants to construct a topological signature, aging patterns with stable topological structures can be identified in high-dimensional temporal evolution features. Based on the similarity of the topological signature, clustering is performed and invariant manifolds are extracted to establish aging evolution path constraints. Typical aging modes can be identified based on the global stability of the topological structure, significantly enhancing the robustness and generalization of the model to different degradation mechanisms. By calculating the Wasserstein distance between the spatial location topological signature and the aging evolution path constraints and combining it with adaptive threshold screening, the degradation state can be comprehensively evaluated at the spatiotemporal and topological levels.

[0039] In one alternative implementation, A multidimensional feature space is constructed based on the feature response amplitudes of the initial candidate locations. Abnormal feature clusters are identified in this multidimensional feature space using a clustering separation algorithm. The anomalous feature clusters are then aggregated to obtain a fault candidate set, including: Extract the feature response magnitude and local manifold deviation corresponding to the initial candidate position and combine them to form an extended feature vector. Construct a multidimensional feature space based on the extended feature vector. In the multidimensional feature space, a nearest neighbor association graph is constructed between the initial candidate positions. The first manifold geodesic distance corresponding to each edge in the nearest neighbor association graph is calculated. The local density field intensity of each initial candidate position is calculated based on the first manifold geodesic distance. Density peak points are identified based on the local density field intensity and the density peak points are used as cluster centers. The second manifold geodesic distance from other initial candidate positions to each density peak point is calculated. The initial candidate positions are assigned to the candidate clusters corresponding to the nearest density peak points based on the second manifold geodesic distance. For each candidate cluster, the local manifold deviation of the initial candidate position in the candidate cluster is extracted, the first statistical moment feature corresponding to the local manifold deviation in the current candidate cluster is calculated, the second statistical moment feature corresponding to the local manifold deviation of the pre-constructed normal evolution mode is extracted, the distribution distance between the first statistical moment feature and the second statistical moment feature is calculated, the candidate clusters whose distribution distance exceeds a preset distance threshold are identified as anomalous feature clusters, and the initial candidate positions in the anomalous feature clusters are aggregated to obtain a fault candidate set.

[0040] The characteristic response amplitude corresponding to the initial candidate positions is extracted. The characteristic response amplitude characterizes the degree of performance degradation of the insulation layer material, and the value ranges from 0 to 1. The local manifold deviation of each initial candidate position is calculated. The local manifold deviation is an index that measures the degree to which the distribution of points in a local area deviates from its manifold. The specific calculation method is based on sampling points within a 15 mm range around each candidate position. A local tangent plane is fitted by principal component analysis, and the mean Euclidean distance from each point to the tangent plane is calculated as the local manifold deviation. For example, for an initial candidate position in an oil-immersed high-voltage harness, the extracted characteristic response amplitude is 0.78, and the calculated local manifold deviation is 0.23, indicating that there is significant insulation degradation and significant local deformation at this position. The characteristic response amplitude and the local manifold deviation are combined to form an extended feature vector [0.78, 0.23]. This vector contains information on both the degree of insulation degradation and local geometric deformation. A two-dimensional feature space is constructed based on the extended feature vectors of all initial candidate positions, where each point corresponds to an initial candidate position.

[0041] In the constructed two-dimensional feature space, the K-nearest neighbor algorithm is used to build a nearest neighbor graph among the initial candidate positions. For example, for each initial candidate position, the K positions in the feature space with the closest Euclidean distance are identified as its nearest neighbors. The value of K is dynamically adjusted based on the total number of initial candidate positions, typically ranging from 5% to 10%. For instance, if the total number of initial candidate positions is 200, then K is between 10 and 20. For every two adjacent initial candidate positions, an undirected edge is established in the nearest neighbor graph.

[0042] The geodesic distance of the first manifold corresponding to each edge in the nearest neighbor graph is calculated. Considering the manifold structure of the feature space, the geodesic distance refers to the shortest path distance measured along the manifold surface, which reflects the intrinsic structure of the data better than Euclidean distance. Specifically, Dijkstra's algorithm is used to calculate the shortest path length between any two points in the nearest neighbor graph as the geodesic distance. For example, for two initial candidate positions with feature vectors [0.78, 0.23] and [0.81, 0.26], the Euclidean distance in the feature space is 0.05, but the geodesic distance calculated after considering the manifold structure is 0.09, reflecting the actual distance along the manifold path.

[0043] The local density field intensity of each initial candidate location is calculated based on the geodesic distance of the first manifold. The local density field intensity represents the density of the initial candidate locations in the feature space. The calculation method involves counting the number of locations within a specific geodesic distance centered on the current location and then smoothing the result using a Gaussian kernel function. Specifically, a cutoff distance of 0.15 is set. For each initial candidate location, the number of other locations with a geodesic distance less than 0.15 from its first manifold is counted, and a Gaussian kernel with a bandwidth parameter of 0.05 is applied for weighting. For example, if the calculated local density field intensity of an initial candidate location is 8.7, it indicates that there are many initial candidate locations around this location, potentially representing the core region of an anomaly cluster.

[0044] Density peak points are identified based on local density field intensity using a local maximum detection algorithm. The local density field intensity of each initial candidate location is compared with that of all its nearest neighbors. If the density field intensity at the current location is greater than that of all its nearest neighbors, it is marked as a density peak point. To avoid too many fragmented peak points, a minimum density threshold condition is set, requiring the local density field intensity of the density peak point to be no less than 5.0. In practical applications, analysis of 200 initial candidate locations identified four density peak points with local density field intensities of 12.3, 9.8, 8.7, and 7.5, corresponding to the centers of four possible abnormal regions in the oil-immersed high-voltage harness.

[0045] Using density peak points as cluster centers, the second manifold geodesic distances from other initial candidate locations to each density peak point are calculated. The calculation method for the second manifold geodesic distance is the same as for the first manifold geodesic distance, employing Dijkstra's algorithm to calculate the shortest path length on the nearest neighbor graph. For each non-peak initial candidate location, its second manifold geodesic distances to all density peak points are calculated, and the density peak point with the shortest distance is selected as the assignment target. For example, if the second manifold geodesic distances from a certain initial candidate location to the four density peak points are 0.28, 0.17, 0.33, and 0.42, this initial candidate location will be assigned to the candidate cluster corresponding to the second density peak point. All 200 initial candidate locations are assigned to four candidate clusters, forming the complete clustering result.

[0046] For each formed candidate cluster, the local manifold deviation of all initial candidate positions within the cluster is extracted, and its first statistical moment feature is calculated. The statistical moment feature is a set of statistics that describe the distribution shape of the data, including mean, variance, skewness, and kurtosis. Specifically, for a candidate cluster containing 25 initial candidate positions, the mean of its local manifold deviation is 0.29, the variance is 0.0064, the skewness is 0.87, and the kurtosis is 3.54, which together constitute the first statistical moment feature vector of the candidate cluster [0.29, 0.0064, 0.87, 3.54].

[0047] The second statistical moment feature of the local manifold deviation corresponding to the pre-constructed normal evolution model is calculated based on oil-immersed high-voltage wiring harness samples confirmed to be in normal condition from historical monitoring data. By analyzing the local manifold deviation distribution of multiple normal samples, reference values ​​of the statistical moment feature are obtained. For example, the statistical moment feature of the local manifold deviation of the normal evolution model has a mean of 0.12, a variance of 0.0025, a skewness of 0.31, and a kurtosis of 2.87, forming the second statistical moment feature vector [0.12, 0.0025, 0.31, 2.87].

[0048] The distribution distance between the first and second statistical moment features is calculated, and the Bhattacharyya distance is used to measure the similarity between the two distributions. The larger the Bhattacharyya distance value, the greater the difference between the two distributions. For the candidate cluster in the previous example, the calculated Bhattacharyya distance between its first statistical moment feature and the second statistical moment feature of the normal evolutionary pattern is 1.87.

[0049] Candidate clusters whose distribution distance exceeds a preset distance threshold are identified as anomalous clusters. The preset distance threshold is set to 1.5 based on historical data statistics. When the distribution distance of a candidate cluster is greater than 1.5, the cluster is determined to be an anomalous cluster. In the aforementioned example, the distribution distance of the candidate cluster, 1.87, is greater than the threshold of 1.5, and therefore it is identified as an anomalous cluster.

[0050] In this embodiment, by simultaneously extracting the feature response amplitude and local manifold deviation at the initial candidate positions and forming an extended feature vector, a more comprehensive characterization of the degradation state is achieved, effectively improving the discriminativeness and sensitivity of the degradation feature expression. The nearest neighbor association graph constructed based on the extended feature vector and the manifold geodesic distance calculation enable the similarity measurement between candidate positions to better fit the real geometric relationship of the wire bundle in the nonlinear manifold space, significantly improving the structural consistency and physical rationality of the clustering results. By identifying the density peak point based on the local density field intensity and using it as the cluster center, the clustering bias problem caused by parameter sensitivity and uncertainty of the number of categories is effectively overcome. By calculating the statistical moment feature of the local manifold deviation within the candidate cluster and comparing the distribution distance with the normal evolution pattern, abnormal feature clusters can be automatically identified based on the statistical distribution difference, thereby accurately distinguishing between normal aging and potential fault degradation states.

[0051] In one alternative implementation, For each anomalous region in the fault candidate set, calculating the degradation rate based on the gradient change rate and extrapolating the remaining lifetime distribution of the insulating medium includes: The temperature field distribution and electric field distribution corresponding to the abnormal region are obtained, and the temperature field strength value and electric field strength value at each location are determined. The thermal sensitivity mapping relationship of the temperature field strength value to the gradient change rate and the electrical sensitivity mapping relationship of the electric field strength value to the gradient change rate are determined. Based on the thermal sensitivity mapping relationship and the electrical sensitivity mapping relationship, the gradient change rate is decoupled to obtain the decoupled gradient change rate. The baseline gradient change rate corresponding to the abnormal region at the historical detection time is obtained. The dynamic evolution trajectory of the decoupled gradient change rate relative to the baseline gradient change rate is determined. The nonlinear acceleration factor in the dynamic evolution trajectory is extracted and quantified to obtain the degradation rate. The critical failure threshold of the insulating medium is extracted from the pre-acquired historical degradation data. The state space distance of the decoupling gradient change rate evolving towards the critical failure threshold is determined. The theoretical failure time is obtained based on the state space distance and the degradation rate. The decoupling gradient change rate corresponding to the historical degradation data is extracted and Fourier spectrum decomposition is performed to obtain the dominant frequency component. The amplitude envelope is calculated based on the dominant frequency component, and the time-domain broadening coefficient is extracted to extend the uncertainty of the theoretical failure time, thereby obtaining the failure time confidence interval. The remaining lifetime of the abnormal region is obtained by calculating the expected value. The remaining lifetime of all abnormal regions is summarized to obtain the remaining lifetime distribution of the insulating medium.

[0052] The temperature field distribution corresponding to the abnormal area was obtained by scanning the abnormal area of ​​the oil-immersed high-voltage wiring harness using infrared thermal imaging technology to acquire a two-dimensional temperature distribution map. The spatial resolution of the temperature field distribution was 1 mm, the temperature measurement accuracy was ±0.5℃, and the temperature measurement range was 30℃ to 120℃. For each identified abnormal area, the temperature values ​​of all locations within that area were extracted to form a complete temperature field distribution. Simultaneously, when acquiring the electric field distribution, an electric field sensing array was used to scan and measure the surface of the oil-immersed high-voltage wiring harness to obtain the spatial distribution of the electric field intensity. Non-contact sensors were used for electric field measurement, with a measurement accuracy of ±5% and a measurement range of 0.1 kV / cm to 15 kV / cm. Based on the temperature field distribution and electric field distribution, the temperature field strength and electric field strength values ​​at each location within the abnormal area were determined. For example, for a location within a certain abnormal area, the measured temperature field strength value was 85.7℃, and the electric field strength value was 7.2 kV / cm.

[0053] A piecewise linear regression method was used to determine the thermal sensitivity mapping relationship between temperature field strength and gradient rate of change. The gradient rate of change refers to the rate at which insulation properties change over time, reflecting the trend of insulation degradation. By analyzing the gradient rate of change under different temperature conditions in historical monitoring data, a mapping function between temperature and gradient rate of change was established. Specifically, the temperature range was divided into multiple intervals, and linear regression was used within each interval to determine the coefficient of the relationship between temperature field strength and gradient rate of change. For example, within the temperature range of 60℃ to 90℃, the thermal sensitivity coefficient was 0.0032, indicating that for every 1℃ increase in temperature, the gradient rate of change increased by 0.0032. An exponential regression method was used to determine the electrical sensitivity mapping relationship between electric field strength and gradient rate of change. By analyzing the gradient rate of change under different electric field strengths, a mapping function between electric field strength and gradient rate of change was established. Within the electric field strength range of 5 kV / cm to 10 kV / cm, the electrical sensitivity coefficient was 0.0025, indicating that for every 1 kV / cm increase in electric field strength, the gradient rate of change increased by 0.0025.

[0054] A multi-factor weighted decomposition method is employed to decouple the gradient rate of change based on thermal and electrical sensitivity mapping relationships. The measured gradient rate of change is decomposed into three components: a temperature contribution component, an electric field contribution component, and a material intrinsic component. During decoupling, the influence of temperature and electric field on the gradient rate of change is calculated using thermal and electrical sensitivity coefficients, with the remaining components treated as the material intrinsic component. For the aforementioned example location, the measured gradient rate of change is 0.085. Through decoupling analysis, the temperature contribution component is calculated to be 0.037, the electric field contribution component to be 0.023, and the material intrinsic component to be 0.025. The decoupled gradient rate of change is then represented by the material intrinsic component, 0.025.

[0055] The baseline gradient change rate of the abnormal region at historical detection times is obtained. The gradient change rate value of this abnormal region under normal operating conditions is extracted from the historical monitoring database. The baseline gradient change rate represents the rate of change of the insulating material during normal aging and serves as a reference value for assessing the degree of abnormality in the current state. For the analyzed abnormal region, the obtained baseline gradient change rate is 0.012, indicating that the current intrinsic gradient change rate of the material in this region is 0.025, significantly higher than the normal value. The dynamic evolution trajectory of the decoupling gradient change rate relative to the baseline gradient change rate is determined by comparing the decoupling gradient change rate with the baseline gradient change rate over multiple consecutive detection periods to form a time series. For example, in six consecutive months of detection, the decoupling gradient change rates of an abnormal region were 0.015, 0.018, 0.021, 0.025, 0.031, and 0.038, respectively, while the baseline gradient change rate remained around 0.012 during the same period. The difference between the two shows an accelerating increasing trend, forming a clear nonlinear evolution trajectory.

[0056] A polynomial fitting method is used to extract the nonlinear acceleration factor from the dynamic evolution trajectory. The time series data is fitted to a second- or third-order polynomial, and the coefficients of the higher-order terms in the polynomial are extracted as the nonlinear acceleration factor. For the time series data in the aforementioned example, in the fitted third-order polynomial, the coefficient of the third-order term is 0.0018, and the coefficient of the second-order term is 0.0012, which together constitute the nonlinear acceleration factor. The nonlinear acceleration factor is quantified to obtain the degradation rate, and the acceleration factor is compared with the baseline gradient change rate to calculate the relative acceleration ratio. For the aforementioned example, the calculated degradation rate is 2.47, indicating that the insulation degradation rate in this region is 2.47 times the normal aging rate.

[0057] The critical failure threshold of the insulation medium is extracted from pre-acquired historical degradation data. The gradient change rate of the same type of oil-immersed high-voltage line harness at failure is then queried from the test database. The critical failure threshold is a characteristic value of the insulation material before breakdown or failure, and is an important reference for predicting remaining life. By analyzing historical failure cases, the critical failure threshold for this type of oil-immersed high-voltage line harness is determined to be 0.085. The state-space distance of the decoupling gradient change rate evolving towards the critical failure threshold is determined, and the difference between the current decoupling gradient change rate and the critical failure threshold is calculated. For the aforementioned example, the current decoupling gradient change rate is 0.025, the critical failure threshold is 0.085, and the state-space distance is 0.060.

[0058] The theoretical failure time is obtained by calculating the state-space distance and degradation rate, using a distance-by-rate method. Considering the nonlinear characteristics of the degradation process, a correction factor is introduced to adjust the calculation results. In the aforementioned example, the state-space distance is 0.060, the degradation rate is 2.47, and the calculated theoretical failure time is 9.7 months. This means that if the degradation rate remains constant, the insulation layer in this region will reach failure after 9.7 months. The correction factor is a variable that dynamically changes with the decoupling gradient rate and temperature / electric field strength, rather than a fixed constant. For example, the correction factor adopts an exponential function form: F = α·exp(β·G)·[1+γ·(T-T0)] 2 ], where G is the decoupling gradient change rate, T is the current temperature, T0 is the reference temperature, and α, β, and γ are preset fitting coefficients. Therefore, the correction factor will be dynamically adjusted according to the operating conditions to reflect the nonlinear characteristics of the degradation process.

[0059] The decoupled gradient change rate corresponding to historical degradation data was extracted and Fourier spectral decomposition was performed to obtain the dominant frequency components. A Fast Fourier Transform (FFT) algorithm was then applied to the historical time series data. Fourier spectral decomposition can reveal the implicit periodic variation patterns in the data. For the analyzed historical data, the dominant frequency obtained after spectral decomposition was 0.083 times / month, corresponding to a period of approximately 12 months, indicating an annual periodic variation in the insulation degradation process. Based on the dominant frequency components, the amplitude envelope was calculated, and the instantaneous amplitude of the signal was extracted using Hilbert transform, and the envelope curve was fitted. The amplitude envelope reflects the change pattern of signal intensity over time.

[0060] The time-domain broadening factor is extracted to extend the uncertainty of the theoretical failure time, and the ratio of the time-domain width of the amplitude envelope to the baseline width is calculated. The time-domain broadening factor represents the degree of uncertainty of the prediction result; the larger the value, the higher the uncertainty. For the aforementioned abnormal region, the calculated time-domain broadening factor is 1.35. Applying the time-domain broadening factor to extend the theoretical failure time, the failure time confidence interval is obtained as 7.2 to 13.1 months. The remaining lifetime of the abnormal region is calculated to be 10.2 months, which is then output as the final prediction result. When summarizing the remaining lifetimes of all abnormal regions to obtain the remaining lifetime distribution of the insulation medium, the minimum value principle is adopted, that is, the remaining lifetime of the entire oil-immersed high-voltage harness depends on the abnormal region with the shortest remaining lifetime. For the analyzed oil-immersed high-voltage harness, a total of 5 abnormal regions were detected, with remaining lifetimes of 10.2, 15.7, 18.3, 21.5, and 24.8 months, respectively. Therefore, the overall remaining lifetime of the oil-immersed high-voltage harness is 10.2 months.

[0061] In this embodiment, by synergistically decoupling the gradient change rate, the non-independent interference caused by the interaction of temperature and electric field is effectively eliminated, enabling the extracted decoupled gradient change rate to more realistically reflect the intrinsic degradation trend of the insulating material. This significantly improves the physical interpretability and index stability of the degradation assessment. By comparing the decoupled gradient change rate with the historical benchmark gradient change rate and extracting the nonlinear acceleration factor, the acceleration stage in the insulation degradation process can be dynamically identified, achieving a quantitative description of the degradation rate. By introducing the critical failure threshold of the insulating medium and calculating the state-space distance of the decoupled gradient change rate evolving towards the threshold, a spatiotemporal mapping relationship between the degradation rate and failure time is established, enabling predictable calculation of potential failure time. By combining Fourier spectral decomposition to extract the dominant frequency components and calculating the time-domain broadening coefficient based on the amplitude envelope, the influence of environmental disturbances and material randomness on failure time can be quantified, achieving an expansion of the uncertainty of the theoretical failure time.

[0062] In one alternative implementation, The remaining life distribution is coupled with the dynamic load curve of the operating condition in a time series calculation to generate hierarchical risk quantification indicators, including: Obtain the dynamic load curve of the operating condition and extract the time-series characteristic parameters corresponding to load fluctuations; Establish a time-domain mapping relationship between the remaining lifetime of each anomalous region in the remaining lifetime distribution and the time-series characteristic parameters. Based on the time-domain mapping relationship, determine the dynamic loss amount of the dynamic load curve on the remaining lifetime. Superimpose the dynamic loss amount onto the current degradation state of the corresponding anomalous region in the remaining lifetime distribution to obtain the load-coupled degradation evolution trajectory. Extract the probability density function of the degradation state reaching the critical failure threshold from the load-coupled degradation evolution trajectory. Calculate the time-domain failure probability distribution based on the probability density function. The time-domain failure probability distribution is mapped to phase space and a phase space trajectory is constructed. The rate of change of curvature of the trajectory curvature in the phase space trajectory is extracted. Based on the rate of change of curvature, the inflection point of risk evolution is identified. The time-domain failure probability distribution is divided into multiple risk evolution stages by using the inflection point as the boundary. The variance of the time-domain failure probability distribution in each risk evolution stage is calculated. Based on the variance, the risk quantification index of each stage is calculated. The risk quantification indexes of all risk evolution stages are summarized to generate a hierarchical risk quantification index.

[0063] Dynamic load curves of the operating conditions are obtained by continuously monitoring parameters such as current, voltage, and power of oil-immersed high-voltage line harnesses during actual operation. These curves record load changes at different times, reflecting the intensity and fluctuation characteristics of equipment use. A power acquisition module samples the operating power of the oil-immersed high-voltage line harnesses at 1-minute intervals, continuously monitoring for 24 hours to generate a dynamic load curve containing 1440 data points. Time-series characteristic parameters corresponding to load fluctuations are extracted, and time-domain analysis is performed on the dynamic load curves to extract characteristic parameters such as load average, fluctuation amplitude, fluctuation period, and peak duration. Taking a specific oil-immersed high-voltage line harness as an example, the extracted time-series characteristic parameters include: a load average of 85 kW, a fluctuation amplitude of 23 kW, a main fluctuation period of 120 minutes, and a peak duration of 45 minutes.

[0064] Multivariate regression analysis was used to establish the time-domain mapping relationship between the remaining lifetime and time-series characteristic parameters of each anomalous region in the remaining lifetime distribution. The accelerated effect of insulation degradation under different load characteristics in historical data was analyzed, and a quantitative relationship between load characteristics and lifetime loss was established. For each anomalous region, a mapping function was constructed, taking the load mean, fluctuation amplitude, fluctuation period, and peak duration as input variables, and outputting the corresponding lifetime loss coefficient. For example, for the anomalous region with a remaining lifetime of 10.2 months, the weighting coefficients of its time-domain mapping function were: load mean 0.028, fluctuation amplitude 0.035, fluctuation period -0.015, and peak duration 0.022. These coefficients indicate that the fluctuation amplitude has the greatest impact on lifetime loss, while the fluctuation period is negatively correlated; that is, the longer the fluctuation period, the smaller the lifetime loss.

[0065] The dynamic loss of the remaining lifetime is determined based on the time-domain mapping relationship, and the extracted time-series characteristic parameters are substituted into the mapping function for calculation. The dynamic loss represents the additional loss of the insulation layer lifetime under the current load condition. For the aforementioned example anomalous region, the calculated dynamic loss is 0.87 months / year, indicating that an additional 0.87 months of remaining lifetime will be lost each year under the current load condition. The dynamic loss is superimposed onto the current degradation state of the corresponding anomalous region in the remaining lifetime distribution, and a linear superposition method is used to consider the cumulative effect. For the anomalous region with a remaining lifetime of 10.2 months, the actual remaining lifetime after considering the load effect is adjusted to 9.5 months. Similarly, corresponding adjustments are made for other anomalous regions to obtain the remaining lifetime distribution considering the load effect.

[0066] When extracting the probability density function of reaching the critical failure threshold from the load-coupled degradation evolution trajectory, a kernel density estimation method is used. This involves placing a kernel function at each observation point and summing all kernel functions to form a smooth probability density estimate. Specifically, a Gaussian kernel function with a bandwidth of 0.8 is selected to perform kernel density estimation on the load-coupled degradation evolution trajectory, obtaining the probability density function at the critical failure time. For example, for the aforementioned anomalous region, the obtained probability density function is distributed over a time range of 7.8 to 11.2 months, with a peak at 9.5 months, indicating that this anomalous region is most likely to reach the critical failure state after 9.5 months. Based on the probability density function, the time-domain failure probability distribution is calculated, and the cumulative distribution function is obtained by integrating the probability density function. The time-domain failure probability distribution represents the probability of equipment failure at different time points and is core data for risk assessment. For this anomalous region, the calculated time-domain failure probability distribution shows: a failure probability of 15% within 8 months, 42% within 9 months, 78% within 10 months, and 96% within 11 months.

[0067] A delayed coordinate embedding method is used to map the temporal failure probability distribution to phase space and construct a phase space trajectory. Phase space is a multi-dimensional state space that can more comprehensively express the dynamic behavior of the system. In specific implementation, a delay time of 1 month and an embedding dimension of 3 are selected to transform the temporal failure probability distribution into a trajectory in three-dimensional phase space. The rate of change of curvature of the trajectory curvature in the phase space trajectory is extracted. First, the curvature value of the trajectory at each point is calculated, and then the rate of change of curvature over time is calculated. Curvature represents the degree of trajectory bending, and the rate of change of curvature reflects the acceleration of the system state change. For the aforementioned phase space trajectory, at time points of 1 month, 4 months, and 8 months, the rates of change of curvature are 0.05, 0.12, and 0.23, respectively, showing that the degree of trajectory bending gradually increases and the state change accelerates.

[0068] Based on the curvature change rate, inflection points in risk evolution are identified, and extreme points or abrupt changes in the curvature change rate are sought. Inflection points represent key turning points in the risk evolution process and are crucial for risk stage classification. For the aforementioned curvature change rate data, two inflection points were identified: 4 months and 8 months, at which the curvature change rate exhibits local and global maxima, respectively. Using these inflection points as boundaries, the time-domain failure probability distribution is divided into multiple intervals. Based on these inflection points, the time-domain failure probability distribution is divided into three risk evolution stages: 0-4 months (low-risk stage), 4-8 months (medium-risk stage), and 8-11 months (high-risk stage).

[0069] The variance of the time-domain failure probability distribution at each risk evolution stage is calculated. Statistical variance is calculated for the failure probability data within each stage. The variance reflects the dispersion of the failure probability distribution; a larger variance indicates higher uncertainty. For the three risk evolution stages mentioned above, the calculated variance values ​​are 0.0003, 0.0025, and 0.0087, respectively, showing that the uncertainty of risk gradually increases over time. When calculating the risk quantification index for each stage based on the variance values, the variance value is combined with the average failure probability of that stage to construct a comprehensive risk index. The risk quantification index comprehensively considers the magnitude and uncertainty of the failure probability, providing a quantitative basis for decision-making. For the three stages mentioned above, the calculated risk quantification indices are 0.15 (low risk), 0.45 (medium risk), and 0.82 (high risk), respectively, indicating that the operational risk of the equipment gradually increases over time. The calculation formula for the risk quantification index is: RI = P avg ·(1+k·σ 2 ), where RI is a risk quantification indicator, and P avg σ represents the average failure probability of this stage. 2 Here is the variance value, and k is the risk sensitivity coefficient, which has a default value of 2.

[0070] By summarizing risk quantification indicators from all stages of risk evolution, a hierarchical risk quantification index is generated, constructing a multi-level risk assessment structure. The hierarchical risk quantification index includes time and risk level dimensions, which can be used to guide maintenance decisions. For this oil-immersed high-voltage harness, the final generated hierarchical risk quantification index is as follows: Low-risk stage (0-4 months): Risk index 0.15, regular monitoring recommended; Medium-risk stage (4-8 months): Risk index 0.45, increased monitoring frequency recommended; High-risk stage (after 8 months): Risk index 0.82, shutdown for maintenance or replacement recommended.

[0071] In this embodiment, by introducing the dynamic load curve of the operating condition and extracting the time-series characteristic parameters of load fluctuation, the dynamic influence of external operating conditions can be fully considered when describing the insulation degradation process, improving the responsiveness and accuracy of lifetime prediction to changes in the actual operating environment. By establishing a time-domain mapping relationship between the remaining lifetime distribution and the load time-series characteristics, and calculating the dynamic loss amount superimposed on the degradation state, dynamic correction of the lifetime decay rate under load can be achieved, making the prediction results more consistent with the degradation law under actual operating conditions. By introducing the probability density function of the degradation state reaching the critical failure threshold in the load-coupled degradation evolution trajectory, the complex time evolution process can be transformed into a quantifiable failure probability distribution, thereby improving the interpretability of lifetime assessment and the accuracy of risk quantification. After mapping the failure probability distribution to phase space, the risk inflection point can be identified by the trajectory curvature change rate, which can effectively reveal the evolution characteristics of degradation risk from slow accumulation to rapid change, and improve the ability to identify the early risk acceleration stage.

[0072] Figure 2 This is a flowchart illustrating the remaining life assessment of the oil-immersed high-voltage wire harness insulation layer fault detection and early warning method according to an embodiment of the present invention.

[0073] In one alternative implementation, Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and warning information is output, including: Obtain the preset safety margin constraint and extract the risk level classification threshold. Compare the risk quantification index of each risk evolution stage in the hierarchical risk quantification index with the risk level classification threshold to determine the fault warning level corresponding to each risk evolution stage. Extract the time domain range of the risk evolution stage corresponding to the fault warning level, take the start time of the time domain range as the intervention start time, take the end time of the time domain range as the intervention end time, and determine the intervention time window corresponding to the fault warning level based on the intervention start time and the intervention end time. The fault warning level, the intervention time window, and the risk quantification index are encapsulated to generate warning information and output.

[0074] The preset safety margin constraint is obtained by retrieving the safety operation parameters set for oil-immersed high-voltage line harnesses in the system configuration. The safety margin constraint refers to the safety time margin reserved before the insulation layer reaches a critical failure state, ensuring sufficient time for maintenance intervention. The safety margin constraint varies depending on the operating environment and importance of the oil-immersed high-voltage line harness. For critical equipment in the power system, the safety margin constraint is typically set at 15% to 20% of the predicted remaining life, while for non-critical equipment, the safety margin constraint can be reduced to 10%. The safety margin constraint set for oil-immersed high-voltage line harnesses is 18%, meaning that at least 18% of the time is reserved for maintenance intervention before the predicted critical failure time.

[0075] A risk matrix method was used to extract risk level classification thresholds. Risks were graded based on two dimensions: risk probability and risk consequence. The risk level classification thresholds refer to the boundary values ​​between different risk levels, used to classify and judge risk quantification indicators. Based on industry standards and historical operating experience, risk levels were divided into four levels: low-risk (green alert), medium-risk (yellow alert), high-risk (orange alert), and extremely high-risk (red alert). The corresponding risk quantification indicator thresholds are: 0.3 (the boundary between low and medium risk), 0.6 (the boundary between medium and high risk), and 0.8 (the boundary between high and extremely high risk). These thresholds were derived from historical failure data analysis.

[0076] A threshold-based judgment method is used to compare the risk quantification index for each stage of risk evolution in the hierarchical risk quantification index with the risk level classification threshold. For each stage of risk evolution, the risk quantification index is compared with the risk level classification threshold to determine the risk level to which that stage belongs. The risk quantification index for the low-risk stage is 0.15, which is less than the first-level threshold of 0.3, so it is judged as a low-risk green alert; the risk quantification index for the medium-risk stage is 0.45, which is between the first-level threshold of 0.3 and the second-level threshold of 0.6, so it is judged as a medium-risk yellow alert; the risk quantification index for the high-risk stage is 0.82, which is greater than the third-level threshold of 0.8, so it is judged as an extremely high-risk red alert.

[0077] A mapping table is used to determine the fault warning level corresponding to each risk evolution stage. The fault warning level is a specific description of the risk level and is usually color-coded for easy and intuitive understanding. Based on the comparison results, each risk evolution stage is mapped to the corresponding fault warning level: the 0-4 month stage is a green warning level 1, indicating that the equipment is in good condition and can operate normally; the 4-8 month stage is a yellow warning level 2, indicating that the equipment shows early signs of deterioration and requires enhanced monitoring; the stage after 8 months is a red warning level 4, indicating that the equipment is severely deteriorated and a maintenance plan needs to be arranged immediately.

[0078] By querying the aforementioned time period division results, the time domain range of the risk evolution stage corresponding to the fault warning level is extracted. The time domain range refers to the start and end points of the risk evolution stage on the time axis, which is the basic data for determining the intervention time window. For the green warning stage, the time domain range is 0-4 months; for the yellow warning stage, the time domain range is 4-8 months; and for the red warning stage, the time domain range is 8-11 months.

[0079] The direct extraction method uses the starting time of the time domain as the intervention start time. The intervention start time refers to the point at which intervention measures need to be taken, which usually coincides with the start time of the risk evolution stage. For the green warning stage, the intervention start time is 0 months from the current time; for the yellow warning stage, the intervention start time is 4 months; and for the red warning stage, the intervention start time is 8 months.

[0080] Similarly, the direct extraction method is used to determine the termination time of the intervention within the time domain. The intervention termination time refers to the latest point in time when intervention measures must be completed; exceeding this time may lead to equipment failure. For the green alert stage, the intervention termination time is 4 months; for the yellow alert stage, it is 8 months; and for the red alert stage, it is 9.5 months. Considering an 18% safety margin constraint, the original 11-month termination time is shortened to 9.5 months.

[0081] An interval calculation method is used to determine the intervention time window corresponding to the fault warning level based on the intervention start and end times. The intervention time window refers to the time range within which maintenance intervention can be implemented, providing a time reference for maintenance planning. For the green warning stage, the intervention time window is 0-4 months, with a length of 4 months; for the yellow warning stage, the intervention time window is 4-8 months, with a length of 4 months; and for the red warning stage, the intervention time window is 8-9.5 months, with a length of 1.5 months. The length of the intervention time window reflects the available time margin for maintenance; the shorter the window, the higher the urgency of maintenance.

[0082] Data fusion technology is used to encapsulate fault warning levels, intervention time windows, and risk quantification indicators to generate warning information. This warning information is a comprehensive risk assessment result, containing key information such as warning level, intervention time, and quantitative risk indicators. In specific implementation, various data are organized into structured information packages for easy parsing and display. For example, the generated warning information includes: Equipment ID: OHV-2023-A158; Inspection date: June 15, 2023; Insulation area location: third connector of the wiring harness; Warning level: currently green, will enter yellow warning in 4 months, and red warning in 8 months; Risk quantification indicator: currently 0.15, 0.45 in 4 months, 0.82 in 8 months; Recommended intervention time window: Green warning: 0-4 months for routine inspection; Yellow warning: 4-8 months for specialized inspection; Red warning: 8-9.5 months for repair or replacement; Risk assessment explanation: Early signs of insulation degradation have appeared, recommending specialized inspection during the yellow warning stage to confirm the degradation status, and repair before the red warning stage.

[0083] In this embodiment, by introducing safety margin constraints into the risk assessment results and comparing the risk quantification indicators of each risk evolution stage using risk level classification thresholds, accurate risk level classification can be achieved. This enables the risk early warning system to have clear level boundaries and decision-making basis, improving the operability and consistency of risk assessment results. By determining the intervention time window according to the time domain range corresponding to the fault early warning level, an organic connection from risk identification to intervention timing planning is achieved. This allows for early intervention in the early stages of risk evolution, improving the timeliness and pertinence of fault prevention and control. The fault early warning level, intervention time window, and risk quantification indicators are encapsulated and output as structured early warning information, enabling the risk results to be directly used in the operation and maintenance decision-making system, realizing an automated closed loop from risk assessment to early warning response.

[0084] A second aspect of the present invention provides a fault detection and early warning system for the insulation layer of oil-immersed high-voltage wire harnesses, comprising: The first unit is used to acquire the multi-physics field coupling signal of the oil-immersed high-voltage line harness in operation and perform multi-scale time-frequency decomposition to obtain the degradation feature distribution, and spatially align the degradation feature distribution with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. The second unit is used to extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain a fault candidate set. The third unit is used to calculate the degradation rate and deduce the remaining lifetime distribution of the insulating medium for each abnormal region in the fault candidate set based on the gradient change rate, and to perform time-series coupling calculation of the remaining lifetime distribution with the dynamic load curve of the operating condition to generate hierarchical risk quantification indicators. The fourth unit is used to determine the fault warning level and the corresponding intervention time window based on the risk quantification index and the preset safety margin constraint, and output the warning information.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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 detecting and warning of insulation layer faults in oil-immersed high-voltage wire harnesses, characterized in that, include: The multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions is obtained and multi-scale time-frequency decomposition is performed to obtain the degradation feature distribution. The degradation feature distribution is spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. Extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain the fault candidate set. For each abnormal region in the fault candidate set, the degradation rate is calculated based on the gradient change rate and the remaining lifetime distribution of the insulating medium is deduced. The remaining lifetime distribution is then coupled with the dynamic load curve of the operating condition in a time series to generate a hierarchical risk quantification index. Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and the warning information is output.

2. The method according to claim 1, characterized in that, The degradation feature distribution is obtained by acquiring the multi-physics field coupling signal of the oil-immersed high-voltage line harness under operating conditions and performing multi-scale time-frequency decomposition. The degradation feature distribution is then spatially aligned with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution, including: The electric field and temperature components of the oil-immersed high-voltage line harness under operating conditions are obtained and combined to obtain a multi-physics field coupled signal; The multi-physics coupling signal is decomposed into time-frequency components at multiple time window scales and multiple frequency resolutions. Feature components reflecting the degradation state of the insulating medium at each scale are extracted. The information entropy difference between feature components at different scales is calculated. Based on the information entropy difference, the feature components at each scale are weighted and fused to obtain the degradation feature distribution. Obtain the geometric topology location information of the oil-immersed high-voltage harness, match the location identifier in the degradation feature distribution with the geometric topology location information to obtain a location transformation matrix, and map the degradation feature distribution to the corresponding geometric topology location based on the location transformation matrix to obtain an aligned degradation feature distribution.

3. The method according to claim 1, characterized in that, Extract the feature response amplitude at each position in the alignment degradation feature distribution, calculate the gradient rate of change of the feature response amplitude in the spatial dimension, and match the gradient rate of change with the aging evolution path constraints to obtain the initial candidate positions, including: Traverse the spatial locations in the alignment degradation feature distribution and extract the corresponding feature response amplitudes; Calculate the local geometric topological curvature corresponding to each spatial location and initialize the sampling direction based on the local geometric topological curvature. Construct an adaptive direction sampling set in the multi-scale neighborhood corresponding to the current location based on the sampling direction. Calculate the directional derivative between the feature response amplitude of the current location and the feature response amplitude of the neighborhood locations on different sampling directions based on the adaptive direction sampling set. Extract the gradient component corresponding to the dominant adaptive direction based on the directional derivative and use the gradient component as the gradient change rate of the current location in the spatial dimension. Extract the multidimensional temporal evolution trajectory corresponding to the gradient change rate from the pre-acquired historical degradation data and map it to the continuous cohomology space. Calculate the topological invariants of the multidimensional temporal evolution trajectory in the continuous cohomology space and construct historical topological signatures based on the topological invariants. Calculate the similarity between different historical topological signatures and determine similar historical topological signatures. Aggregate multidimensional temporal evolution trajectories with similar historical topological signatures and obtain aging evolution path constraints by identifying invariant manifolds. Calculate the topological signature corresponding to each spatial location, calculate the Wasserstein distance metric between the topological signature and the aging evolution path constraint, and filter the initial candidate locations by combining the results with a pre-set adaptive distance threshold.

4. The method according to claim 1, characterized in that, A multidimensional feature space is constructed based on the feature response amplitudes of the initial candidate locations. Abnormal feature clusters are identified in this multidimensional feature space using a clustering separation algorithm. The anomalous feature clusters are then aggregated to obtain a fault candidate set, including: Extract the feature response magnitude and local manifold deviation corresponding to the initial candidate position and combine them to form an extended feature vector. Construct a multidimensional feature space based on the extended feature vector. In the multidimensional feature space, a nearest neighbor association graph is constructed between the initial candidate positions. The first manifold geodesic distance corresponding to each edge in the nearest neighbor association graph is calculated. The local density field intensity of each initial candidate position is calculated based on the first manifold geodesic distance. Density peak points are identified based on the local density field intensity and the density peak points are used as cluster centers. The second manifold geodesic distance from other initial candidate positions to each density peak point is calculated. The initial candidate positions are assigned to the candidate clusters corresponding to the nearest density peak points based on the second manifold geodesic distance. For each candidate cluster, the local manifold deviation of the initial candidate position in the candidate cluster is extracted, the first statistical moment feature corresponding to the local manifold deviation in the current candidate cluster is calculated, the second statistical moment feature corresponding to the local manifold deviation of the pre-constructed normal evolution mode is extracted, the distribution distance between the first statistical moment feature and the second statistical moment feature is calculated, the candidate clusters whose distribution distance exceeds a preset distance threshold are identified as anomalous feature clusters, and the initial candidate positions in the anomalous feature clusters are aggregated to obtain a fault candidate set.

5. The method according to claim 1, characterized in that, For each anomalous region in the fault candidate set, calculating the degradation rate based on the gradient change rate and extrapolating the remaining lifetime distribution of the insulating medium includes: The temperature field distribution and electric field distribution corresponding to the abnormal region are obtained, and the temperature field strength value and electric field strength value at each location are determined. The thermal sensitivity mapping relationship of the temperature field strength value to the gradient change rate and the electrical sensitivity mapping relationship of the electric field strength value to the gradient change rate are determined. Based on the thermal sensitivity mapping relationship and the electrical sensitivity mapping relationship, the gradient change rate is decoupled to obtain the decoupled gradient change rate. The baseline gradient change rate corresponding to the abnormal region at the historical detection time is obtained. The dynamic evolution trajectory of the decoupled gradient change rate relative to the baseline gradient change rate is determined. The nonlinear acceleration factor in the dynamic evolution trajectory is extracted and quantified to obtain the degradation rate. The critical failure threshold of the insulating medium is extracted from the pre-acquired historical degradation data. The state space distance of the decoupling gradient change rate evolving towards the critical failure threshold is determined. The theoretical failure time is obtained based on the state space distance and the degradation rate. The decoupling gradient change rate corresponding to the historical degradation data is extracted and Fourier spectrum decomposition is performed to obtain the dominant frequency component. The amplitude envelope is calculated based on the dominant frequency component, and the time-domain broadening coefficient is extracted to extend the uncertainty of the theoretical failure time, thereby obtaining the failure time confidence interval. The remaining lifetime of the abnormal region is obtained by calculating the expected value. The remaining lifetime of all abnormal regions is summarized to obtain the remaining lifetime distribution of the insulating medium.

6. The method according to claim 1, characterized in that, The remaining life distribution is coupled with the dynamic load curve of the operating condition in a time series calculation to generate hierarchical risk quantification indicators, including: Obtain the dynamic load curve of the operating condition and extract the time-series characteristic parameters corresponding to load fluctuations; Establish a time-domain mapping relationship between the remaining lifetime of each anomalous region in the remaining lifetime distribution and the time-series characteristic parameters. Based on the time-domain mapping relationship, determine the dynamic loss amount of the dynamic load curve on the remaining lifetime. Superimpose the dynamic loss amount onto the current degradation state of the corresponding anomalous region in the remaining lifetime distribution to obtain the load-coupled degradation evolution trajectory. Extract the probability density function of the degradation state reaching the critical failure threshold from the load-coupled degradation evolution trajectory. Calculate the time-domain failure probability distribution based on the probability density function. The time-domain failure probability distribution is mapped to phase space and a phase space trajectory is constructed. The rate of change of curvature of the trajectory curvature in the phase space trajectory is extracted. Based on the rate of change of curvature, the inflection point of risk evolution is identified. The time-domain failure probability distribution is divided into multiple risk evolution stages by using the inflection point as the boundary. The variance of the time-domain failure probability distribution in each risk evolution stage is calculated. Based on the variance, the risk quantification index of each stage is calculated. The risk quantification indexes of all risk evolution stages are summarized to generate a hierarchical risk quantification index.

7. The method according to claim 1, characterized in that, Based on the aforementioned risk quantification indicators and preset safety margin constraints, the fault warning level and corresponding intervention time window are determined, and warning information is output, including: Obtain the preset safety margin constraint and extract the risk level classification threshold. Compare the risk quantification index of each risk evolution stage in the hierarchical risk quantification index with the risk level classification threshold to determine the fault warning level corresponding to each risk evolution stage. Extract the time domain range of the risk evolution stage corresponding to the fault warning level, take the start time of the time domain range as the intervention start time, take the end time of the time domain range as the intervention end time, and determine the intervention time window corresponding to the fault warning level based on the intervention start time and the intervention end time. The fault warning level, the intervention time window, and the risk quantification index are encapsulated to generate warning information and output.

8. An oil-immersed high-voltage wire harness insulation layer fault detection and early warning system, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire the multi-physics field coupling signal of the oil-immersed high-voltage line harness in operation and perform multi-scale time-frequency decomposition to obtain the degradation feature distribution, and spatially align the degradation feature distribution with the geometric topological position of the oil-immersed high-voltage line harness to obtain the aligned degradation feature distribution. The second unit is used to extract the feature response amplitude of each position in the alignment degradation feature distribution, calculate the gradient change rate of the feature response amplitude in the spatial dimension, match the gradient change rate with the aging evolution path constraint to obtain the initial candidate position, construct a multi-dimensional feature space based on the feature response amplitude of the initial candidate position, identify abnormal feature clusters in the multi-dimensional feature space through a clustering separation algorithm, and aggregate the abnormal feature clusters to obtain a fault candidate set. The third unit is used to calculate the degradation rate and deduce the remaining lifetime distribution of the insulating medium for each abnormal region in the fault candidate set based on the gradient change rate, and to perform time-series coupling calculation of the remaining lifetime distribution with the dynamic load curve of the operating condition to generate hierarchical risk quantification indicators. The fourth unit is used to determine the fault warning level and the corresponding intervention time window based on the risk quantification index and the preset safety margin constraint, and output the warning information.

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