A high-voltage equipment digital management and intelligent diagnosis system based on a graph convolution network

Abstract

This invention discloses a digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks, comprising: a data acquisition module for acquiring preprocessed multimodal data to generate node features; a graph construction module for generating an adjacency matrix and graph input; a graph convolution processing module for generating multi-order features using MixHop; a topology constraint module for applying homology constraints to generate topology features; a time-scale modeling module for window-based modeling to generate time-scale features; an energy constraint module for introducing power flow, partial discharge, and temperature rise to generate physical features; a contrastive learning and heterogeneous convolution module for fusing features; a classification and regression module; and an operation and maintenance instruction generation module for outputting status and instructions. This invention improves diagnostic accuracy and operational feasibility.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and management technology for power equipment, and in particular to a digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks. Background Technology

[0002] High-voltage equipment is a crucial component of the safe operation of power systems, and its operational status directly impacts grid stability. Traditional management methods rely on manual inspections and single-sensor analysis, making it difficult to achieve real-time and comprehensive status awareness. With the development of multimodal monitoring technology, current, voltage, temperature, and partial discharge data can be collected in real time, but existing methods still have shortcomings in fusion analysis and intelligent diagnosis.

[0003] Existing diagnostic models often employ methods such as support vector machines and convolutional neural networks, which typically neglect electrical connections and spatial topological characteristics between devices, leading to insufficient feature extraction. Some studies have introduced graph convolutional networks, but these often remain at a fixed order of neighborhood convolution, making it difficult to consider both short-term and long-term dependencies and limiting feature representation capabilities. Furthermore, existing methods lack mechanisms for preserving topological features such as connectivity, loops, and cavities, and their time-scale analysis is limited, failing to accurately distinguish between short-term fluctuations and long-term degradation.

[0004] Furthermore, most existing models are disconnected from physical mechanisms, failing to incorporate physical constraints such as power flow energy, partial discharge energy density, and temperature rise energy, which can easily lead to inconsistencies between diagnostic results and actual operation. During training, these models lack effective differentiation of convolutional order differences and employ a single convolutional operator, resulting in insufficient feature diversity.

[0005] In terms of results application, existing methods typically only output health status or simple health indices, failing to combine classification and regression results with operation and maintenance needs to generate maintenance priorities and decision instructions, resulting in a disconnect between diagnostic results and actual operation and maintenance.

[0006] Therefore, how to provide a digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks. This invention fully utilizes MixHop graph convolution, multi-scale topological homology analysis, time-scale modeling, physical consistency constraints, and the fusion of contrastive learning and heterogeneous convolution. It describes in detail the processing method for intelligently realizing the identification of the health status of high-voltage equipment, anomaly scoring, and degradation stage assessment. It has the advantages of comprehensive data modeling, sufficient feature representation, high diagnostic accuracy, and executable operation and maintenance decisions.

[0008] A digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks, according to an embodiment of the present invention, includes:

[0009] The data acquisition module is used to collect and preprocess multimodal monitoring data to generate node feature sequences;

[0010] The graph construction module is used to build a set of nodes and a set of edges, generate an adjacency matrix, and combine it with the node feature sequence to form graph input data;

[0011] The graph convolution processing module is used to perform convolution on the graph input data using the MixHop algorithm to generate a multi-level neighborhood feature set;

[0012] The topology constraint module is used to apply topological homology constraints on multi-order neighborhood feature sets to generate topology-preserving feature sets;

[0013] The time-scale modeling module is used to divide the topology-preserving feature set into time windows to model convolutional features and generate time-scale feature sets.

[0014] The energy constraint module is used to calculate the power flow energy, partial discharge energy density, and temperature rise energy in the time-scale feature set and then weight and adjust them to generate a physical consistency feature set.

[0015] The contrastive learning and heterogeneous convolution module is used to construct sample pairs on a physically consistent feature set and calculate the contrastive loss, which is then combined with heterogeneous convolution to generate a fused feature vector.

[0016] The classification and regression module is used to input the fused feature vector into the classification network to output the health status category and input it into the regression network to output the abnormality score and the degradation stage index.

[0017] The operation and maintenance instruction generation module is used to generate maintenance priority ranking and operation and maintenance decision instructions based on health status category, anomaly score and degradation stage indicators, and write them into the digital management system.

[0018] Optionally, modules can be integrated using the following methods:

[0019] Collect and preprocess multimodal monitoring data from high-voltage equipment to generate node feature sequences;

[0020] Construct a device graph structure, form an adjacency matrix, and combine it with node feature sequences to form graph input data;

[0021] The MixHop algorithm is used to perform first-order, second-order, and third-order convolutions on the graph input data to generate a multi-order neighborhood feature set.

[0022] By applying topological homology constraints to the multi-order neighborhood feature set, connectivity, loop and cavity features are preserved, resulting in a topology-preserving feature set.

[0023] A time-scale coupling relationship is established for the topology-preserving feature set, so that the first-order convolutional feature represents short-term degradation, the second-order convolutional feature represents medium-term degradation, and the third-order convolutional feature represents long-term degradation, thus generating a time-scale feature set.

[0024] By imposing energy consistency constraints on the time-scale feature set and introducing power flow energy, partial discharge energy density and temperature rise energy, a physical consistency feature set is generated.

[0025] Contrastive learning constraints are introduced into the physical consistency feature set to enhance order differences, and a heterogeneous convolution operator consisting of Laplacian operator, graph wavelet operator and random walk operator is integrated to generate a fused feature vector.

[0026] The fused feature vectors are input into the classification and regression modules, and the outputs include the equipment health status, anomaly score, and degradation stage. Based on the output results, maintenance priority ranking and operation and maintenance decision instructions are generated and written into the digital management system.

[0027] Optionally, the construction and processing of the device diagram structure specifically includes:

[0028] Each high-voltage device is designated as a node, forming a node set;

[0029] An edge set is established based on electrical connections and geographical adjacency, with each edge corresponding to a pair of nodes that are connected or spatially adjacent.

[0030] An adjacency matrix is ​​generated based on the set of nodes and the set of edges. When there is a connection between nodes in the adjacency matrix, the corresponding position is set to one, and when there is no connection between nodes, the corresponding position is set to zero.

[0031] The node feature sequence is organized into a node feature matrix, where each row of the node feature matrix corresponds to the feature vector of a node.

[0032] The adjacency matrix and the node feature matrix are combined to form the graph input data.

[0033] Optionally, the generation of the multi-order neighborhood feature set specifically includes:

[0034] The adjacency matrix is ​​normalized to generate a normalized adjacency matrix;

[0035] The first-order neighborhood convolution is calculated on the node feature matrix using the normalized adjacency matrix to generate a first-order convolution feature matrix. Each row of the first-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the first-order neighborhood.

[0036] The second-order neighborhood convolution is calculated on the node feature matrix by using the square of the normalized adjacency matrix to generate a second-order convolution feature matrix. Each row of the second-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the second-order neighborhood.

[0037] The cube of the normalized adjacency matrix is ​​used to calculate the third-order neighborhood convolution on the node feature matrix, generating a third-order convolution feature matrix. Each row of the third-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the third-order neighborhood.

[0038] The first-order convolutional feature matrix, the second-order convolutional feature matrix, and the third-order convolutional feature matrix are concatenated along the node dimension to form a multi-order neighborhood feature set.

[0039] Optionally, the generation of the topology-preserving feature set specifically includes:

[0040] Each node feature vector in the multi-order neighborhood feature set is mapped to point cloud data. Each point in the point cloud data corresponds to a node feature vector. The distance between points is determined by the similarity between node feature vectors.

[0041] A filtering function is constructed on the point cloud data. The filtering function takes the numerical value of the node feature vector as input and assigns a scalar value to each point. The scalar values ​​are arranged in ascending order and are used to control the generation process of the simplex.

[0042] Based on the distance between points and the scalar value of the filter function, simplexes are generated according to progressively increasing scale thresholds. Within the threshold range, two points are connected as an edge, three points are connected in pairs as a triangular face, four points are connected in pairs as a three-dimensional simplex, and so on, forming a sequence of simplexes.

[0043] In simple complex sequences, structural changes are tracked, and the generation and disappearance times of connected components, loops, and cavities are recorded.

[0044] The generation and disappearance times of connected components, loops, and cavities at different scales are represented as barcodes to form a persistent barcode set.

[0045] By using a persistent barcode set to constrain a multi-level neighborhood feature set, a topology-preserving feature set is output.

[0046] Optionally, the generation of the time-scale feature set specifically includes:

[0047] The topology-preserving feature set is divided into time windows according to time order, and each time window contains continuous node feature vectors;

[0048] Model the first-order convolutional features within each time window, calculate the average value, fluctuation amplitude, and rate of change of the first-order convolutional features within the time window, and combine the calculation results into a short-term degradation feature sequence.

[0049] Model the second-order convolutional features within each time window, calculate the temporal repetition of the second-order convolutional features within the time window, the similarity between features of different nodes, and the matrix features composed of the similarity of all nodes, and combine the calculation results into a mid-term degradation feature sequence.

[0050] The third-order convolutional features are modeled within each time window. The maximum and minimum value ranges, the offset of the long-term average value, and the overall fluctuation range of the third-order convolutional features within the time window are calculated, and the calculation results are combined into a long-term degradation feature sequence.

[0051] Establish coupling relationships between adjacent time windows, align short-term degradation feature sequences, medium-term degradation feature sequences, and long-term degradation feature sequences to form cross-scale correspondences;

[0052] By sequentially combining cross-scale correspondences along the time dimension, a time-scale feature set is generated.

[0053] Optionally, the generation of the physical consistency feature set specifically includes:

[0054] Extract current data, voltage data, partial discharge pulse data, and temperature data from the time-scale feature set by node and time window, and perform time alignment.

[0055] Within each time window, the current sequence and voltage sequence are multiplied one by one according to the corresponding time point to obtain the instantaneous power sequence. The instantaneous power sequence is then accumulated point by point within the time window to generate the power flow energy characteristic sequence.

[0056] Partial discharge pulses are detected within each time window. The product of the amplitude and duration of a single pulse is calculated to obtain the energy of a single pulse. Then, the energies of all pulses within the time window are accumulated one by one to generate a partial discharge energy density characteristic sequence.

[0057] Within each time window, the difference curve between the temperature sequence and the baseline temperature is calculated with reference to the baseline temperature. The difference curve is then accumulated point by point within the time window to generate a temperature rise energy characteristic sequence.

[0058] The power flow energy characteristic sequence, partial discharge energy density characteristic sequence, and temperature rise energy characteristic sequence are normalized respectively, and combined into an energy constraint matrix in the dimensions of nodes and time windows. The rows of the energy constraint matrix correspond to the combination of nodes and time windows, and the columns correspond to the three types of energy characteristics in turn.

[0059] The time-scale feature set is stacked into a time-scale feature matrix. A mapping matrix is ​​introduced to align the time-scale feature matrix with the energy constraint matrix. During the alignment process, the time-scale feature matrix is ​​weighted and adjusted according to the value of the energy constraint matrix.

[0060] The weighted and adjusted time-scale feature matrix is ​​restored to a set of physically consistent features according to nodes and time windows.

[0061] Optionally, the generation of the fused feature vector specifically includes:

[0062] First-order convolutional features, second-order convolutional features, and third-order convolutional features are extracted from the physical consistency feature set. Features from the same node in adjacent time windows are selected to form positive sample pairs, and features from different nodes in the same time window are selected to form negative sample pairs.

[0063] Normalization is performed on the first-order, second-order, and third-order convolutional features, and the normalized inner product is used as the feature similarity measure to obtain the positive sample pair similarity sequence and the negative sample pair similarity sequence.

[0064] To maintain the similarity of positive samples and distinguish the differences of negative samples across convolutional features of different orders, a contrastive loss function is constructed based on positive and negative sample similarity. This contrastive loss function is defined by the following formula:

[0065] ;

[0066] in, This indicates the convolution order, with values ​​of first, second, and third. Indicates the first The anchored sample and the positive sample at the Similarity in convolutional features of order 1 Indicates the first The anchored samples and negative samples at the _ Similarity in convolutional features of order 1 Indicates temperature parameter, Indicates batch size. The formula represents the number of negative samples. It is determined by exponentially amplifying the similarity of positive samples and using the exponential sum of the similarity of negative samples as the normalized denominator, thereby forming a contrast constraint based on probability distribution.

[0067] Weight coefficients are set for first-order convolutional features, second-order convolutional features, and third-order convolutional features respectively, and the total contrastive loss is obtained by weighted summation of the contrastive losses of each order;

[0068] The Laplacian convolution operator, graph wavelet convolution operator, and random walk convolution operator are applied to the physical consistency feature set to generate Laplacian convolution features, graph wavelet convolution features, and random walk convolution features, respectively.

[0069] The first-order, second-order, and third-order convolutional features are updated with the total contrastive loss as the optimization objective. The updated convolutional features are then concatenated with the Laplacian convolutional features, graph wavelet convolutional features, and random walk convolutional features along the feature dimension to generate a fused feature vector.

[0070] Optionally, the generation and processing of maintenance priority ranking and operation and maintenance decision instructions specifically include:

[0071] The fused feature vector is input into the classification module, which includes a fully connected layer and a probability output function. The probability output function performs an exponential operation on the output score of each category and normalizes it on the category dimension to obtain the probability distribution of the device belonging to each health status category.

[0072] The device's health status category is output based on a probability distribution.

[0073] The fused feature vector is input into the regression module, which includes a fully connected layer and a linear output layer. The linear output layer generates the regression prediction result by multiplying the matrix and the vector and adding the bias vector.

[0074] Anomaly scores are calculated based on regression prediction results. Anomaly scores indicate the degree of anomaly of the equipment during the monitoring period.

[0075] The degradation stage index is calculated based on the regression prediction results. The degradation stage index represents the degradation level of the equipment at different time scales.

[0076] The health status category, abnormal score, and degradation stage indicators are used as joint input and output to the operation and maintenance instruction generation module. A priority ranking function is established in the operation and maintenance instruction generation module, and maintenance priority ranking and operation and maintenance decision instructions are generated based on the joint output.

[0077] The maintenance priority ranking and operation and maintenance decision instructions are written into the digital management system.

[0078] The beneficial effects of this invention are:

[0079] The digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks proposed in this invention can make full use of the multimodal monitoring data of high-voltage equipment. By constructing a graph structure and using MixHop convolution to extract multi-order neighborhood features, it not only takes into account both short-distance and long-distance dependencies, but also maintains connectivity, loop and cavity features by combining topological homology constraints, thereby making up for the shortcomings of traditional methods in terms of insufficient structural feature representation.

[0080] In terms of time and physical dimensions, this invention establishes a time-scale coupling relationship, models convolutional features of different orders as short-term, medium-term and long-term degradation features respectively, and introduces power flow energy, partial discharge energy density and temperature rise energy to construct physical consistency constraints, so that the feature expression can simultaneously reflect the dynamic evolution and physical mechanism of the device, solving the problems of time-scale fragmentation and insufficient physical constraints in the prior art.

[0081] In terms of feature optimization and result application, this invention introduces contrastive learning constraints into the physically consistent feature set to distinguish the differences in features of different convolutional orders, and integrates heterogeneous operators such as Laplacian convolution, graph wavelet convolution, and random walk convolution to improve feature diversity and expressive power. Ultimately, this invention can output equipment health status, anomaly scores, and degradation stages in the classification and regression modules, and generate maintenance priority rankings and operation and maintenance instructions, achieving direct connection between diagnostic results and operation and maintenance decisions, thus improving the accuracy of diagnosis and the executability of management. Attached Figure Description

[0082] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0083] Figure 1 The flowchart shows a high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks proposed in this invention.

[0084] Figure 2 This is a schematic diagram illustrating the topological homology constraints of a high-voltage equipment digital management and intelligent diagnostic system based on graph convolutional networks proposed in this invention.

[0085] Figure 3 This is a schematic diagram of the time-scale modeling of a high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks proposed in this invention. Detailed Implementation

[0086] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0087] refer to Figure 1-3 A digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks, comprising:

[0088] The data acquisition module is used to collect and preprocess multimodal monitoring data to generate node feature sequences;

[0089] The graph construction module is used to build a set of nodes and a set of edges, generate an adjacency matrix, and combine it with the node feature sequence to form graph input data;

[0090] The graph convolution processing module is used to perform convolution on the graph input data using the MixHop algorithm to generate a multi-level neighborhood feature set;

[0091] The topology constraint module is used to apply topological homology constraints on multi-order neighborhood feature sets to generate topology-preserving feature sets;

[0092] The time-scale modeling module is used to divide the topology-preserving feature set into time windows to model convolutional features and generate time-scale feature sets.

[0093] The energy constraint module is used to calculate the power flow energy, partial discharge energy density, and temperature rise energy in the time-scale feature set and then weight and adjust them to generate a physical consistency feature set.

[0094] The contrastive learning and heterogeneous convolution module is used to construct sample pairs on a physically consistent feature set and calculate the contrastive loss, which is then combined with heterogeneous convolution to generate a fused feature vector.

[0095] The classification and regression module is used to input the fused feature vector into the classification network to output the health status category and input it into the regression network to output the abnormality score and the degradation stage index.

[0096] The operation and maintenance instruction generation module is used to generate maintenance priority ranking and operation and maintenance decision instructions based on health status category, anomaly score and degradation stage indicators, and write them into the digital management system.

[0097] This invention establishes a complete digital management and intelligent diagnostic process for high-voltage equipment by constructing modules for data acquisition, graph construction, graph convolution processing, topological constraints, time-scale modeling, energy constraints, contrastive learning and heterogeneous convolution, classification and regression, and operation and maintenance instruction generation. This achieves a closed-loop processing mechanism from multimodal monitoring data acquisition to intelligent diagnosis and operation and maintenance decision output. The system not only extracts multi-level neighborhood features from different dimensions but also maintains feature consistency layer by layer through topological, time-scale, and energy constraints. Finally, it combines contrastive learning and heterogeneous convolution to generate high-quality fused features, improving the accuracy of equipment health status diagnosis, anomaly score calculation, and degradation stage identification, thus providing a reliable basis for maintenance sequencing and operation and maintenance instructions.

[0098] In this embodiment, the modules are interconnected using the following method:

[0099] Collect and preprocess multimodal monitoring data from high-voltage equipment to generate node feature sequences;

[0100] Construct a device graph structure, form an adjacency matrix, and combine it with node feature sequences to form graph input data;

[0101] The MixHop algorithm is used to perform first-order, second-order, and third-order convolutions on the graph input data to generate a multi-order neighborhood feature set.

[0102] By applying topological homology constraints to the multi-order neighborhood feature set, connectivity, loop and cavity features are preserved, resulting in a topology-preserving feature set.

[0103] A time-scale coupling relationship is established for the topology-preserving feature set, so that the first-order convolutional feature represents short-term degradation, the second-order convolutional feature represents medium-term degradation, and the third-order convolutional feature represents long-term degradation, thus generating a time-scale feature set.

[0104] By imposing energy consistency constraints on the time-scale feature set and introducing power flow energy, partial discharge energy density and temperature rise energy, a physical consistency feature set is generated.

[0105] Contrastive learning constraints are introduced into the physical consistency feature set to enhance order differences, and a heterogeneous convolution operator consisting of Laplacian operator, graph wavelet operator and random walk operator is integrated to generate a fused feature vector.

[0106] The fused feature vectors are input into the classification and regression modules, and the outputs include the equipment health status, anomaly score, and degradation stage. Based on the output results, maintenance priority ranking and operation and maintenance decision instructions are generated and written into the digital management system.

[0107] This invention provides a comprehensive design for the collaboration between modules at the methodological level, forming an orderly connection throughout the entire process from data acquisition to digital management. By employing MixHop convolution on the graph input data, it can generate hierarchical feature sets of different neighborhoods. Combined with steps such as topological coherence, time-scale modeling, and energy constraints, it ensures the full preservation and fusion of multi-level feature information. This method organically combines complex multimodal monitoring data with a graph convolutional network structure, enabling it not only to identify short-term, medium-term, and long-term degradation characteristics of high-voltage equipment, but also to uniformly model health status, anomaly scores, and degradation stages. This provides high-quality input for the digital management system to generate accurate maintenance priorities and operation and maintenance decisions.

[0108] In this embodiment, the construction and processing of the device diagram structure specifically includes:

[0109] Each high-voltage device is designated as a node, forming a node set;

[0110] An edge set is established based on electrical connections and geographical adjacency, with each edge corresponding to a pair of nodes that are connected or spatially adjacent.

[0111] An adjacency matrix is ​​generated based on the set of nodes and the set of edges. When there is a connection between nodes in the adjacency matrix, the corresponding position is set to one, and when there is no connection between nodes, the corresponding position is set to zero.

[0112] The node feature sequence is organized into a node feature matrix, where each row of the node feature matrix corresponds to the feature vector of a node.

[0113] The adjacency matrix and the node feature matrix are combined to form the graph input data.

[0114] This invention effectively models the electrical connections and geographical adjacencies between high-voltage equipment by introducing node and edge sets in the graph construction step. This allows complex power grid equipment systems to be expressed in graph structure, further generating an adjacency matrix and combining it with the node feature matrix to form graph data that can be directly input into a graph convolutional network. This modeling method preserves both physical connection features and geographical topological features, laying a solid foundation for subsequent multi-level convolution and feature aggregation. It ensures that the extracted graph convolutional features reflect the real interaction relationships between equipment, thereby improving the physical rationality and reliability of the diagnosis.

[0115] In this embodiment, the generation of the multi-order neighborhood feature set specifically includes:

[0116] The adjacency matrix is ​​normalized to generate a normalized adjacency matrix;

[0117] The first-order neighborhood convolution is calculated on the node feature matrix using the normalized adjacency matrix to generate a first-order convolution feature matrix. Each row of the first-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the first-order neighborhood.

[0118] The second-order neighborhood convolution is calculated on the node feature matrix by using the square of the normalized adjacency matrix to generate a second-order convolution feature matrix. Each row of the second-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the second-order neighborhood.

[0119] The cube of the normalized adjacency matrix is ​​used to calculate the third-order neighborhood convolution on the node feature matrix, generating a third-order convolution feature matrix. Each row of the third-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the third-order neighborhood.

[0120] The first-order convolutional feature matrix, the second-order convolutional feature matrix, and the third-order convolutional feature matrix are concatenated along the node dimension to form a multi-order neighborhood feature set.

[0121] This invention employs the MixHop algorithm in graph convolution processing. Through first-, second-, and third-order operations on the normalized adjacency matrix, it calculates convolutional features within different neighborhood ranges and concatenates them along the node dimension to form a multi-order neighborhood feature set. This design can simultaneously capture information from both local and distant neighborhoods, avoiding the feature insufficiency problem that may result from single-order convolution. Furthermore, the multi-order feature fusion further enhances the richness of node representation, providing high-quality input for subsequent topological homology constraints and time-scale modeling, resulting in more accurate and comprehensive diagnostic results.

[0122] In this embodiment, the generation of the topology-preserving feature set specifically includes:

[0123] Each node feature vector in the multi-order neighborhood feature set is mapped to point cloud data. Each point in the point cloud data corresponds to a node feature vector. The distance between points is determined by the similarity between node feature vectors.

[0124] A filtering function is constructed on the point cloud data. The filtering function takes the numerical value of the node feature vector as input and assigns a scalar value to each point. The scalar values ​​are arranged in ascending order and are used to control the generation process of the simplex.

[0125] Based on the distance between points and the scalar value of the filter function, simplexes are generated according to progressively increasing scale thresholds. Within the threshold range, two points are connected as an edge, three points are connected in pairs as a triangular face, four points are connected in pairs as a three-dimensional simplex, and so on, forming a sequence of simplexes.

[0126] In simple complex sequences, structural changes are tracked, and the generation and disappearance times of connected components, loops, and cavities are recorded.

[0127] The generation and disappearance times of connected components, loops, and cavities at different scales are represented as barcodes to form a persistent barcode set.

[0128] By using a persistent barcode set to constrain a multi-level neighborhood feature set, a topology-preserving feature set is output.

[0129] This invention introduces topological homology analysis onto a multi-order neighborhood feature set. Through point cloud mapping, filter function construction, simplex generation, and persistent barcode extraction, it comprehensively characterizes the generation and disappearance processes of topological features such as connectivity, loops, and cavities. This approach preserves the topological properties of the graph structure at different scales and uses persistent homology results to constrain multi-order convolutional features. This retains topological invariance during deep learning, preventing the loss of global topological information due to stacked computations, resulting in a final feature set with stronger robustness and interpretability.

[0130] In this embodiment, the generation of the time-scale feature set specifically includes:

[0131] The topology-preserving feature set is divided into time windows according to time order, and each time window contains continuous node feature vectors;

[0132] Model the first-order convolutional features within each time window, calculate the average value, fluctuation amplitude, and rate of change of the first-order convolutional features within the time window, and combine the calculation results into a short-term degradation feature sequence.

[0133] Model the second-order convolutional features within each time window, calculate the temporal repetition of the second-order convolutional features within the time window, the similarity between features of different nodes, and the matrix features composed of the similarity of all nodes, and combine the calculation results into a mid-term degradation feature sequence.

[0134] The third-order convolutional features are modeled within each time window. The maximum and minimum value ranges, the offset of the long-term average value, and the overall fluctuation range of the third-order convolutional features within the time window are calculated, and the calculation results are combined into a long-term degradation feature sequence.

[0135] Establish coupling relationships between adjacent time windows, align short-term degradation feature sequences, medium-term degradation feature sequences, and long-term degradation feature sequences to form cross-scale correspondences;

[0136] By sequentially combining cross-scale correspondences along the time dimension, a time-scale feature set is generated.

[0137] This invention introduces temporal modeling based on topology-preserving feature sets, modeling convolutional features from different time windows as short-term, medium-term, and long-term degradation features, and forming a temporal feature set through cross-scale correspondence. This method not only captures rapid device fluctuations in the short term but also extracts degradation trends in the medium to long term, ultimately unifying information from different scales within the same time frame. By establishing coupling relationships between time windows, it can realistically reflect the dynamic evolution of device degradation processes, providing a more scientific temporal scale representation for subsequent energy consistency constraints and anomaly detection.

[0138] In this embodiment, the generation of the physical consistency feature set specifically includes:

[0139] Extract current data, voltage data, partial discharge pulse data, and temperature data from the time-scale feature set by node and time window, and perform time alignment.

[0140] Within each time window, the current sequence and voltage sequence are multiplied one by one according to the corresponding time point to obtain the instantaneous power sequence. The instantaneous power sequence is then accumulated point by point within the time window to generate the power flow energy characteristic sequence.

[0141] Partial discharge pulses are detected within each time window. The product of the amplitude and duration of a single pulse is calculated to obtain the energy of a single pulse. Then, the energies of all pulses within the time window are accumulated one by one to generate a partial discharge energy density characteristic sequence.

[0142] Within each time window, the difference curve between the temperature sequence and the baseline temperature is calculated with reference to the baseline temperature. The difference curve is then accumulated point by point within the time window to generate a temperature rise energy characteristic sequence.

[0143] The power flow energy characteristic sequence, partial discharge energy density characteristic sequence, and temperature rise energy characteristic sequence are normalized respectively, and combined into an energy constraint matrix in the dimensions of nodes and time windows. The rows of the energy constraint matrix correspond to the combination of nodes and time windows, and the columns correspond to the three types of energy characteristics in turn.

[0144] The time-scale feature set is stacked into a time-scale feature matrix. A mapping matrix is ​​introduced to align the time-scale feature matrix with the energy constraint matrix. During the alignment process, the time-scale feature matrix is ​​weighted and adjusted according to the value of the energy constraint matrix.

[0145] The weighted and adjusted time-scale feature matrix is ​​restored to a set of physically consistent features according to nodes and time windows.

[0146] This invention introduces energy consistency constraints into the time-scale feature set, transforming current, voltage, partial discharge, and temperature signals into three types of physical energy indicators: power flow energy, partial discharge energy density, and temperature rise energy. An energy constraint matrix is ​​constructed and aligned with the time-scale feature matrix to achieve weighted feature adjustment. This method ensures that the convolutional features are consistent with the actual energy distribution through physical constraints, enabling the feature representation to reflect the physical laws governing equipment operation. This guarantees that the diagnostic model not only relies on mathematical calculations but also possesses physical interpretability, providing more reliable results for health status identification and anomaly analysis.

[0147] In this embodiment, the generation of the fused feature vector specifically includes:

[0148] First-order convolutional features, second-order convolutional features, and third-order convolutional features are extracted from the physical consistency feature set. Features from the same node in adjacent time windows are selected to form positive sample pairs, and features from different nodes in the same time window are selected to form negative sample pairs.

[0149] Normalization is performed on the first-order, second-order, and third-order convolutional features, and the normalized inner product is used as the feature similarity measure to obtain the positive sample pair similarity sequence and the negative sample pair similarity sequence.

[0150] To maintain the similarity of positive samples and distinguish the differences of negative samples across convolutional features of different orders, a contrastive loss function is constructed based on positive and negative sample similarity. This contrastive loss function is defined by the following formula:

[0151] ;

[0152] in, This indicates the convolution order, with values ​​of first, second, and third. Indicates the first The anchored sample and the positive sample at the Similarity in convolutional features of order 1 Indicates the first The anchored samples and negative samples at the _ Similarity in convolutional features of order 1 Indicates temperature parameter, Indicates batch size. The formula represents the number of negative samples. It is determined by exponentially amplifying the similarity of positive samples and using the exponential sum of the similarity of negative samples as the normalized denominator, thereby forming a contrast constraint based on probability distribution.

[0153] Weight coefficients are set for first-order convolutional features, second-order convolutional features, and third-order convolutional features respectively, and the total contrastive loss is obtained by weighted summation of the contrastive losses of each order;

[0154] The Laplacian convolution operator, graph wavelet convolution operator, and random walk convolution operator are applied to the physical consistency feature set to generate Laplacian convolution features, graph wavelet convolution features, and random walk convolution features, respectively.

[0155] The first-order, second-order, and third-order convolutional features are updated with the total contrastive loss as the optimization objective. The updated convolutional features are then concatenated with the Laplacian convolutional features, graph wavelet convolutional features, and random walk convolutional features along the feature dimension to generate a fused feature vector.

[0156] This invention introduces contrastive learning into a physically consistent feature set, constructing a contrastive loss function using positive and negative sample pairs to enhance the discriminative ability between convolutional features of different orders. Simultaneously, it combines Laplacian, graph wavelet, and random walk operators for heterogeneous convolutional feature extraction, fusing the updated convolutional features with the heterogeneous features to generate a high-dimensional fused feature vector. This method optimizes convolutional features under contrastive constraints and enhances diversity through heterogeneous operators, ultimately forming a fused feature vector with both discriminative power and robustness. This provides better input for classification and regression modules, improving overall diagnostic performance.

[0157] In this embodiment, the generation and processing of maintenance priority ranking and operation and maintenance decision instructions specifically include:

[0158] The fused feature vector is input into the classification module, which includes a fully connected layer and a probability output function. The probability output function performs an exponential operation on the output score of each category and normalizes it on the category dimension to obtain the probability distribution of the device belonging to each health status category.

[0159] The device's health status category is output based on a probability distribution.

[0160] The fused feature vector is input into the regression module, which includes a fully connected layer and a linear output layer. The linear output layer generates the regression prediction result by multiplying the matrix and the vector and adding the bias vector.

[0161] Anomaly scores are calculated based on regression prediction results. Anomaly scores indicate the degree of anomaly of the equipment during the monitoring period.

[0162] The degradation stage index is calculated based on the regression prediction results. The degradation stage index represents the degradation level of the equipment at different time scales.

[0163] The health status category, abnormal score, and degradation stage indicators are used as joint input and output to the operation and maintenance instruction generation module. A priority ranking function is established in the operation and maintenance instruction generation module, and maintenance priority ranking and operation and maintenance decision instructions are generated based on the joint output.

[0164] The maintenance priority ranking and operation and maintenance decision instructions are written into the digital management system.

[0165] This invention utilizes a dual output mechanism of classification and regression modules to transform fused feature vectors into health status categories, anomaly scores, and degradation stage indicators. These are further processed in the maintenance instruction generation module to generate maintenance priority ranking and maintenance decision instructions. This design enables the direct conversion of diagnostic results into executable management decisions, achieving an automated closed loop from monitoring to diagnosis to maintenance. Through joint modeling of health status classification, anomaly score quantification, and degradation stage location, a more comprehensive reference can be provided for the maintenance of power equipment, making maintenance decisions more scientific and operable.

[0166] Example 1:

[0167] To verify the feasibility of this invention in practice, it was applied to the digital management and intelligent diagnosis of high-voltage power grid equipment. The monitored objects included high-voltage circuit breakers, disconnect switches, current transformers, and voltage transformers. The monitoring data sources covered current signals, voltage signals, partial discharge pulse signals, and equipment temperature signals. Traditional methods often rely on a single sensor or a simple statistical model, resulting in incomplete feature extraction and insufficient accuracy of diagnostic results. This invention effectively solves the problems of insufficient data correlation, missing topological features, and lack of physical consistency constraints by combining graph convolutional networks with topological coherence constraints.

[0168] In this scenario, the current, voltage, and temperature during equipment operation are first collected synchronously by multimodal sensors, and pulse signals are recorded by a partial discharge detector to obtain a continuous monitoring sequence. Then, noise and abnormal interference are removed through preprocessing steps, and the monitoring sequence is transformed into a node feature sequence. Based on the electrical connection relationship and spatial adjacency relationship between devices, a graph structure is established and an adjacency matrix is ​​generated to form graph input data containing node features and edge relationships.

[0169] In the process of processing the graph input data, the MixHop algorithm is used to perform convolution operations in the first-order, second-order, and third-order neighborhoods to capture multi-order dependency features of short and long distances, forming a multi-order neighborhood feature set. In order to avoid losing topological information in degradation modeling, this invention applies topological coherence constraints to the feature set. By constructing a filter function and a simplex, persistent barcodes of connected components, loops, and cavities are calculated to form a topology-preserving feature set. Then, the topology-preserving feature set is divided according to the time window to establish a cross-scale coupling relationship. The first-order convolution features correspond to short-term degradation signals, the second-order convolution features correspond to medium-term degradation trends, and the third-order convolution features correspond to long-term degradation patterns, thereby obtaining a time-scale feature set.

[0170] Based on the time-scale feature set, this invention further introduces power flow energy, partial discharge energy density, and temperature rise energy to construct an energy constraint matrix, which is then weighted and adjusted to obtain a physical consistency feature set. Positive and negative sample pairs are then constructed, and contrastive learning constraints are used to enhance the differential expression of different convolution orders. Combined with Laplacian convolution, graph wavelet convolution, and random walk convolution, a fused feature vector is generated. Finally, the fused feature vector is input into the classification and regression module to obtain the equipment's health status, anomaly score, and degradation stage. Based on the diagnostic results, maintenance priority ranking and operation and maintenance instructions are generated and written into the digital management system for direct use by operation and maintenance personnel.

[0171] To verify the effectiveness of this invention, a batch of high-voltage equipment was selected, and both the conventional method and the method of this invention were used for diagnosis. The detection accuracy, false alarm rate, diagnosis time, and ability to identify degradation trends were compared. The comparison results are shown in Table 1.

[0172] Table 1. Comparison of Diagnostic Performance under Different Methods

[0173]

[0174] As shown in Table 1, the present invention significantly outperforms existing methods in terms of diagnostic accuracy, degradation trend recognition capability, and false alarm rate. The traditional support vector machine has a long-term degradation recognition rate of only 65.8%, and the conventional graph convolutional network has a rate of 72.4%, while the present invention reaches 87.2%. It also improves the recognition of short-term and medium-term degradation by more than 10 percentage points compared with the traditional method. In terms of false alarm rate, the present invention reduces the false alarm rate to 4.1%, which is about two-thirds less than that of the traditional support vector machine. At the same time, the present invention performs well in terms of computational efficiency, with the average diagnosis time shortened to 2.9 seconds, which better meets the real-time requirements of power grid operation and maintenance.

[0175] As can be seen from the above embodiments, the present invention not only improves the accuracy of high-voltage equipment condition diagnosis, but also can characterize the degradation process in layers on a time scale, ensuring the consistency between the diagnosis results and the physical characteristics of the equipment, and can directly generate maintenance priorities and operation and maintenance instructions, thereby effectively solving the shortcomings of the prior art in terms of incomplete features, insufficient physical constraints, and disconnect between diagnosis results and operation and maintenance.

[0176] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital management and intelligent diagnostic system for high-voltage equipment based on graph convolutional networks, characterized in that, include: The data acquisition module is used to collect and preprocess multimodal monitoring data to generate node feature sequences; The graph construction module is used to build a set of nodes and edges, generate an adjacency matrix, and combine it with the node feature sequence to form graph input data. Specifically: Each high-voltage device is designated as a node, forming a node set; An edge set is established based on electrical connections and geographical adjacency, with each edge corresponding to a pair of nodes that are connected or spatially adjacent. An adjacency matrix is ​​generated based on the set of nodes and the set of edges. When there is a connection between nodes in the adjacency matrix, the corresponding position is set to one, and when there is no connection between nodes, the corresponding position is set to zero. The node feature sequence is organized into a node feature matrix, where each row of the node feature matrix corresponds to the feature vector of a node. The adjacency matrix and the node feature matrix are combined to form the graph input data; The graph convolution processing module is used to perform convolution on the graph input data using the MixHop algorithm to generate a multi-level neighborhood feature set; The topology constraint module is used to apply topological homology constraints on multi-order neighborhood feature sets to generate topology-preserving feature sets; The time-scale modeling module is used to divide the topology-preserving feature set into time windows to model convolutional features and generate time-scale feature sets. The energy constraint module is used to calculate the power flow energy, partial discharge energy density, and temperature rise energy within a time-scale feature set, and then weight and adjust them to generate a physically consistent feature set. Specifically: Extract current data, voltage data, partial discharge pulse data, and temperature data from the time-scale feature set by node and time window, and perform time alignment. Within each time window, the current sequence and voltage sequence are multiplied one by one according to the corresponding time point to obtain the instantaneous power sequence. The instantaneous power sequence is then accumulated point by point within the time window to generate the power flow energy characteristic sequence. Partial discharge pulses are detected within each time window. The product of the amplitude and duration of a single pulse is calculated to obtain the energy of a single pulse. Then, the energies of all pulses within the time window are accumulated one by one to generate a partial discharge energy density characteristic sequence. Within each time window, the difference curve between the temperature sequence and the baseline temperature is calculated with reference to the baseline temperature. The difference curve is then accumulated point by point within the time window to generate a temperature rise energy characteristic sequence. The power flow energy characteristic sequence, partial discharge energy density characteristic sequence, and temperature rise energy characteristic sequence are normalized respectively, and combined into an energy constraint matrix in the dimensions of nodes and time windows. The rows of the energy constraint matrix correspond to the combination of nodes and time windows, and the columns correspond to the three types of energy characteristics in turn. The time-scale feature set is stacked into a time-scale feature matrix. A mapping matrix is ​​introduced to align the time-scale feature matrix with the energy constraint matrix. During the alignment process, the time-scale feature matrix is ​​weighted and adjusted according to the value of the energy constraint matrix. The weighted and adjusted time-scale feature matrix is ​​restored to a set of physically consistent features according to nodes and time windows; The contrastive learning and heterogeneous convolution module is used to construct sample pairs on a physically consistent feature set and calculate the contrastive loss, which is then combined with heterogeneous convolution to generate a fused feature vector. The classification and regression module is used to input the fused feature vector into the classification network to output the health status category and input it into the regression network to output the abnormality score and the degradation stage index. The operation and maintenance instruction generation module is used to generate maintenance priority ranking and operation and maintenance decision instructions based on health status category, anomaly score and degradation stage indicators, and write them into the digital management system.

2. The high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks according to claim 1, characterized in that, The modules are connected in the following way: Collect and preprocess multimodal monitoring data from high-voltage equipment to generate node feature sequences; Construct a device graph structure, form an adjacency matrix, and combine it with node feature sequences to form graph input data; The MixHop algorithm is used to perform first-order, second-order, and third-order convolutions on the graph input data to generate a multi-order neighborhood feature set. By applying topological homology constraints to the multi-order neighborhood feature set, connectivity, loop and cavity features are preserved, resulting in a topology-preserving feature set. A time-scale coupling relationship is established for the topology-preserving feature set, so that the first-order convolutional feature represents short-term degradation, the second-order convolutional feature represents medium-term degradation, and the third-order convolutional feature represents long-term degradation, thus generating a time-scale feature set. By imposing energy consistency constraints on the time-scale feature set and introducing power flow energy, partial discharge energy density and temperature rise energy, a physical consistency feature set is generated. Contrastive learning constraints are introduced into the physical consistency feature set to enhance order differences, and a heterogeneous convolution operator consisting of Laplacian operator, graph wavelet operator and random walk operator is integrated to generate a fused feature vector. The fused feature vectors are input into the classification and regression modules, and the outputs include the equipment health status, anomaly score, and degradation stage. Based on the output results, maintenance priority ranking and operation and maintenance decision instructions are generated and written into the digital management system.

3. The high-voltage equipment digital management and intelligent diagnostic system based on graph convolutional networks according to claim 2, characterized in that, The generation of the multi-order neighborhood feature set specifically includes: The adjacency matrix is ​​normalized to generate a normalized adjacency matrix; The first-order neighborhood convolution is calculated on the node feature matrix using the normalized adjacency matrix to generate a first-order convolution feature matrix. Each row of the first-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the first-order neighborhood. The second-order neighborhood convolution is calculated on the node feature matrix by using the square of the normalized adjacency matrix to generate a second-order convolution feature matrix. Each row of the second-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the second-order neighborhood. The cube of the normalized adjacency matrix is ​​used to calculate the third-order neighborhood convolution on the node feature matrix, generating a third-order convolution feature matrix. Each row of the third-order convolution feature matrix corresponds to the feature vector of a node after aggregation in the third-order neighborhood. The first-order convolutional feature matrix, the second-order convolutional feature matrix, and the third-order convolutional feature matrix are concatenated along the node dimension to form a multi-order neighborhood feature set.

4. The high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks according to claim 2, characterized in that, The generation of the topology-preserving feature set specifically includes: Each node feature vector in the multi-order neighborhood feature set is mapped to point cloud data. Each point in the point cloud data corresponds to a node feature vector. The distance between points is determined by the similarity between node feature vectors. A filtering function is constructed on the point cloud data. The filtering function takes the numerical value of the node feature vector as input and assigns a scalar value to each point. The scalar values ​​are arranged in ascending order and are used to control the generation process of the simplex. Based on the distance between points and the scalar value of the filter function, simplexes are generated according to progressively increasing scale thresholds. Within the threshold range, two points are connected as an edge, three points are connected in pairs as a triangular face, four points are connected in pairs as a three-dimensional simplex, and so on, forming a sequence of simplexes. In simple complex sequences, structural changes are tracked, and the generation and disappearance times of connected components, loops, and cavities are recorded. The generation and disappearance times of connected components, loops, and cavities at different scales are represented as barcodes to form a persistent barcode set. By using a persistent barcode set to constrain a multi-level neighborhood feature set, a topology-preserving feature set is output.

5. The high-voltage equipment digital management and intelligent diagnostic system based on graph convolutional networks according to claim 2, characterized in that, The generation of the time-scale feature set specifically includes: The topology-preserving feature set is divided into time windows according to time order, and each time window contains continuous node feature vectors; Model the first-order convolutional features within each time window, calculate the average value, fluctuation amplitude, and rate of change of the first-order convolutional features within the time window, and combine the calculation results into a short-term degradation feature sequence. Model the second-order convolutional features within each time window, calculate the temporal repetition of the second-order convolutional features within the time window, the similarity between features of different nodes, and the matrix features composed of the similarity of all nodes, and combine the calculation results into a mid-term degradation feature sequence. The third-order convolutional features are modeled within each time window. The maximum and minimum value ranges, the offset of the long-term average value, and the overall fluctuation range of the third-order convolutional features within the time window are calculated, and the calculation results are combined into a long-term degradation feature sequence. Establish coupling relationships between adjacent time windows, align short-term degradation feature sequences, medium-term degradation feature sequences, and long-term degradation feature sequences to form cross-scale correspondences; By sequentially combining cross-scale correspondences along the time dimension, a time-scale feature set is generated.

6. The high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks according to claim 2, characterized in that, The generation of the fused feature vector specifically includes: First-order convolutional features, second-order convolutional features, and third-order convolutional features are extracted from the physical consistency feature set. Features from the same node in adjacent time windows are selected to form positive sample pairs, and features from different nodes in the same time window are selected to form negative sample pairs. Normalization is performed on the first-order, second-order, and third-order convolutional features, and the normalized inner product is used as the feature similarity measure to obtain the positive sample pair similarity sequence and the negative sample pair similarity sequence. To maintain the similarity of positive samples and distinguish the differences of negative samples on convolutional features of different orders, a contrastive loss function is constructed based on the similarity of positive samples and the similarity of negative samples. Weight coefficients are set for first-order convolutional features, second-order convolutional features, and third-order convolutional features respectively, and the total contrastive loss is obtained by weighted summation of the contrastive losses of each order; The Laplacian convolution operator, graph wavelet convolution operator, and random walk convolution operator are applied to the physical consistency feature set to generate Laplacian convolution features, graph wavelet convolution features, and random walk convolution features, respectively. The first-order, second-order, and third-order convolutional features are updated with the total contrastive loss as the optimization objective. The updated convolutional features are then concatenated with the Laplacian convolutional features, graph wavelet convolutional features, and random walk convolutional features along the feature dimension to generate a fused feature vector.

7. The high-voltage equipment digital management and intelligent diagnosis system based on graph convolutional networks according to claim 2, characterized in that, The generation and processing of maintenance priority ranking and operation and maintenance decision instructions specifically include: The fused feature vector is input into the classification module, which includes a fully connected layer and a probability output function. The probability output function performs an exponential operation on the output score of each category and normalizes it on the category dimension to obtain the probability distribution of the device belonging to each health status category. The device's health status category is output based on a probability distribution. The fused feature vector is input into the regression module, which includes a fully connected layer and a linear output layer. The linear output layer generates the regression prediction result by multiplying the matrix and the vector and adding the bias vector. Anomaly scores are calculated based on regression prediction results. Anomaly scores indicate the degree of anomaly of the equipment during the monitoring period. The degradation stage index is calculated based on the regression prediction results. The degradation stage index represents the degradation level of the equipment at different time scales. The health status category, abnormal score, and degradation stage indicators are used as joint input and output to the operation and maintenance instruction generation module. A priority ranking function is established in the operation and maintenance instruction generation module, and maintenance priority ranking and operation and maintenance decision instructions are generated based on the joint output. The maintenance priority ranking and operation and maintenance decision instructions are written into the digital management system.

Images