An intelligent high-voltage switch cabinet fault diagnosis method and system based on heterogeneous graph structure learning
By adopting a method based on heterogeneous graph structure learning, the problems of difficult data labeling and high graph structure dependence in high-voltage switchgear fault diagnosis are solved, achieving efficient and accurate fault type judgment and improving the reliability and diagnostic efficiency of power grid equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID ANHUI ELECTRIC POWER CO LTD ELECTRIC POWER SCI RES INST
- Filing Date
- 2022-12-19
- Publication Date
- 2026-05-15
AI Technical Summary
Existing fault diagnosis methods for high-voltage switchgear are insufficient to meet the demands of rapid power grid development for equipment reliability. In particular, they suffer from difficulties in data annotation, poor data quality, and high dependence on graph structure learning, leading to inaccurate fault type identification.
A heterogeneous graph structure learning approach is adopted. Through sensor data collection, cleaning and processing, and using unsupervised anomaly detection algorithms and expert knowledge annotation, a heterogeneous graph structure learning algorithm is constructed to automatically learn the interaction relationships of sensor networks of heterogeneous components in high-voltage switchgear. This breaks the isomorphism assumption, mines network structure information, and achieves accurate operational status assessment.
It significantly improves the accuracy and efficiency of fault diagnosis in high-voltage switchgear, enabling real-time and accurate determination of fault types, reducing the difficulty of data annotation and reliance on manually constructed diagram structures, and improving the reliability and accuracy of diagnosis.
Smart Images
Figure CN116244617B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a fault diagnosis method for intelligent high-voltage switchgear in the power industry. Background Technology
[0002] High-voltage switchgear is the most widely used and numerous type of switchgear in power construction, and it is widely used in power generation, transmission, and distribution. Currently, the demand for electricity from various industries is constantly increasing with my country's rapid economic development. As an important power equipment, high-voltage switchgear products are also constantly being iterated and improved, developing towards intelligence, high reliability, and low maintenance.
[0003] High-voltage switchgear plays a dual role in closing power lines and protecting system safety, ensuring the normal and reliable operation of power distribution networks. Its main characteristics are its widespread use, large quantity, and relatively high accident rate. my country has experienced numerous high-voltage switchgear flash explosions, causing complete equipment shutdowns and significant economic losses. Common high-voltage switchgear faults include failure to operate, maloperation, opening and closing faults, insulation faults, current-carrying faults, and external force faults. Therefore, a reliable high-voltage switchgear fault diagnosis system is needed. This system should be able to promptly detect changes in the switchgear's operating status based on sensor data, accurately determine the fault type, assist power experts in identifying repair directions, and conduct rapid and precise fault repairs. This prevents the paralysis of the entire power system and the resulting irreparable losses due to delayed repairs.
[0004] High-voltage switchgear fault diagnosis systems are crucial for power grid operation and monitoring. Traditional methods for detecting various switchgear faults primarily rely on live-line testing, using manual or robotic-assisted intelligent equipment, which is risky and costly. Traditional methods also rely on periodic data collection, resulting in difficult data storage, limited data volume, significant background interference variations, uncertain time intervals, and highly uneven data distribution. In recent years, with the significant improvement in computing power and the development of sensor technology, a series of neural network-based methods have emerged to address switchgear fault diagnosis. However, sequential neural network models struggle to capture the relationships between sensor networks of high-voltage switchgear components. Therefore, with the excellent performance of graph neural networks in modeling non-Euclidean spatial structures, it has become possible to use graph neural networks to model the relationships between sensor networks of switchgear components. However, the following two problems still exist in high-voltage switchgear fault diagnosis:
[0005] 1. Neural network-based methods are data-driven, and fault diagnosis of high-voltage switchgear requires a large amount of labeled data to train the model. The internal operating environment of high-voltage switchgear is more complex than that of ordinary switchgear, with more components and a variety of sensors. Manually labeling data requires considering the status of each sensor, which leads to difficulties in labeling and low efficiency.
[0006] 2. Existing graph neural network models rely excessively on given graph structure data for message passing, and are highly sensitive to data quality. Artificially constructed graph structure data also suffers from poor data quality and structural deficiencies. Graph structure learning methods are based on the assumption of graph isomorphism, which is difficult to apply, and cosine similarity measurement is not suitable for sensor networks of heterogeneous components in high-voltage switchgear.
[0007] In summary, existing fault diagnosis methods for high-voltage switchgear are insufficient to meet the reliability demands of the rapidly developing power grid. How to accurately detect data defects, reliably assess the operating status of high-voltage switchgear, and accurately determine the type of fault when it occurs are urgent problems to be solved. Summary of the Invention
[0008] The technical problem to be solved by this invention is how to reliably assess the working status of high-voltage switchgear and accurately determine the type of fault when a fault occurs.
[0009] This invention solves the above-mentioned technical problems through the following technical means: a fault diagnosis method for intelligent high-voltage switchgear based on heterogeneous graph structure learning, characterized by including the following steps:
[0010] Step 1: Collect, clean and process sensor data from intelligent high-voltage switchgear. The sensor data is heterogeneous, including temperature, local, mechanical characteristics and gas sensor data.
[0011] Step 2: Based on the data processed in Step 1, use an unsupervised anomaly detection algorithm to find outliers to distinguish between running states and abnormal states, including the following steps:
[0012] Step 22: Use the K-means algorithm to cluster the sensor data to separate normal values and outliers in the sensor data;
[0013] Step 24: Use the Isolation Forest algorithm to perform anomaly detection on the sensor data and find outliers in the data;
[0014] The union of the outlier data from steps 22 and 24 is taken as the final filtered outlier data.
[0015] Step 3: Construct a switchgear fault diagnosis database. Use crowdsourcing technology combined with expert knowledge to label fault types. Mark normal operating conditions as non-fault and abnormal conditions as fault types.
[0016] Step 4: Design a heterogeneous graph structure learning algorithm to find a representation space that meets the similarity metric of heterogeneous sensor data, automatically learn the interaction relationships of the sensor network of heterogeneous components in the intelligent high-voltage switchgear, and mine network structure information to achieve accurate operation status assessment of the high-voltage switchgear.
[0017] As a further optimized technical solution, the sensor data collection and cleaning process includes: deleting duplicate data collected from the same sensor; filling missing values in the data with the mean of the corresponding sensor data; and normalizing the data using a data maximum / minimum scaling method.
[0018] As a further optimized technical solution, the sensor data includes:
[0019] Temperature sensor data: contact temperature data, contact arm temperature data, and temperature data from multiple points on the inner wall;
[0020] Partial discharge sensor data: ultrasonic detection signal, TEV transient ground wave signal, UHF ultra-high frequency signal;
[0021] Mechanical characteristic data: opening and closing coil current, opening and closing speed;
[0022] Gas sensor data: gas content.
[0023] As a further optimized technical solution, the specific process of the K-means algorithm is as follows: Select clusters K=2, representing normal values and outliers respectively, randomly select the center of each cluster and denot it as z1 and z2, and then iterate the data of each type of sensor. In each iteration, calculate the distance of each sample x to the two cluster centers z1 and z2, and assign x to the cluster center with the closer distance. After traversing the data of each type of sensor, calculate the mean of the two cluster sets and update it as the center value. Repeat the above clustering algorithm until the cluster centers remain unchanged, stop the iteration, and return the clustering result.
[0024] The specific process of the isolated forest algorithm is as follows: Take Ψ data points from each type of sensor as subsamples and place them into the root node of an isolated tree, where each data vector h = {h1, h2, ... h} n},h iFor values in different dimensions, a dimension is randomly assigned. Within the current node's data range, a cut point p is randomly generated. Cut point p occurs between the maximum and minimum values of the specified dimension in the current node's data. Points less than p in the currently selected dimension are assigned to the left branch of the current node, and points greater than or equal to p are assigned to the right branch of the current node. The segmentation ends when there is only one data point in a leaf node. By repeating the above algorithm, t isolated trees are obtained. Then, the entire sensor database is traversed. For each sample x, the anomaly score s(x,Ψ) is calculated through the t isolated trees. The calculation process is as follows:
[0025]
[0026] Where Ψ is the number of samples, h(x) is the height of each tree, and c(Ψ) is the average path length for a given number of samples Ψ, used to standardize the path length h(x) of sample x. By averaging the scores of each tree, the final anomaly score is obtained, and the k sample points with the highest anomaly scores are taken as the final anomaly points.
[0027] As a further optimized technical solution, step 4 includes: reading data from the switchgear fault diagnosis database constructed in step 3, dividing it into training and validation sets, constructing a heterogeneous component sensor network, designing a heterogeneous graph structure learning algorithm and training the model on the training set, and automatically capturing the interaction relationships of the intelligent high-voltage switchgear component sensor network. This includes three parts: graph structure construction, graph structure modeling, and message passing.
[0028] The graph structure is constructed as follows: a heterogeneous component sensor network is constructed based on the training set data in the switchgear fault diagnosis database. The component sensors in the high-voltage switchgear are used as nodes in the heterogeneous component sensor network, and the edges in the heterogeneous component sensor network represent the interaction relationships between components.
[0029] Graph structure modeling utilizes a decoupled graph convolutional neural network G to obtain aggregated information from heterogeneous component sensor network node nodes, updates the decoupled embedding X, and predicts the relation matrix A. * ;
[0030] For message passing, a new graph convolutional neural network G* is constructed to perform message passing and information aggregation.
[0031] As a further optimized technical solution, the graph structure construction process is as follows: First, the node features are initialized using heterogeneous sensor data in the switchgear fault diagnosis database to obtain the coupled data feature matrix X'. The distance between each sensor is calculated using the cosine distance metric, i.e., vector dot product operation. Second, the number of neighboring nodes k is set, and all nodes in the heterogeneous group component sensor network of the high-voltage switchgear are traversed. For each group component sensor node, the k nearest nodes are selected as neighboring nodes, and the edges between group component sensors are constructed to obtain the initialized heterogeneous group component sensor network, i.e., the relation matrix A.
[0032] As a further optimized technical solution, the graph structure modeling process is as follows: The decoupled embedding is updated after aggregating and updating the information of heterogeneous component sensor network nodes by obtaining the information from a decoupled graph convolutional neural network G. H represents the decoupled sensor embedding of the component, and ReLU is the activation function. Let f be the Laplacian matrix of the graph, representing the interaction relationships between nodes in a heterogeneous component sensor network. Let D be the degree matrix of the graph minus the relation matrix A. Design a decoupling mapping function f. k Mapping the input coupled data feature matrix X' to different decoupling spaces yields the decoupled component sensor embedding Z. k =f k (X′), where the decoupling mapping function f k The representation space transformation is performed using a two-layer MLP, where X' is the input coupled data feature matrix X', and Z... k To decouple and embed the component sensor in the k-th decoupling space, traverse all nodes in the heterogeneous component sensor network, calculate the similarity between the node and its neighboring nodes in each decoupling space, and select the decoupling space Z with the highest similarity. i Finally, the decoupled sensor component was selected and embedded in H=Z. i Using a decoupled graph convolutional neural network G through the Laplacian matrix of the graph To facilitate message passing, each node in the heterogeneous component sensor network aggregates information from all its first-order neighbors, enabling information interaction. Information aggregation within the decoupled space breaks the isomorphism assumption, allowing for the acquisition of similarities between heterogeneous components and the aggregation of heterogeneous information. After obtaining the decoupled embedding X updated through message passing, the similarity relationship between nodes in the heterogeneous component sensor network is evaluated using a cosine similarity metric function. High similarity indicates high information correlation, thus establishing edges between two nodes. This process is repeated for all nodes in the heterogeneous component sensor network, updating the network's relational matrix A. * .
[0033] As a further optimized technical solution, the specific message passing process is as follows: a new graph convolutional neural network G* is constructed for message passing and information aggregation, where G* is a ReLU (G * (A * XW), where W is a linear layer, the initial embedding of a node is obtained through linear transformation, and the updated node embedding h∈R n+d Where n is the number of sensors in the component group, i.e., the number of nodes, and d is the dimension of the representation vector. Message passing aggregates information based on the edges in the heterogeneous component sensor network. During model training, the network structure information can be implicitly learned and input into the model pooling layer to obtain the overall state representation embedding of the high-voltage switchgear. Finally, it is input into the model p = sigmoid(MLP(mean(h))) for fault diagnosis, where sigmoid is the activation function, MLP is a linear layer, and the model predicts the probability of belonging to each category.
[0034] As a further optimization technique, the model's loss function is as follows:
[0035]
[0036] Where y i The actual label is marked. To predict the probability of being in the i-th category, where C is the fault category and m is the number of samples, the model is trained on the training set through backpropagation of loss and the optimal solution is obtained. By minimizing the loss, the model can automatically learn the structural information of the heterogeneous component sensor network related to fault diagnosis.
[0037] This invention also provides an intelligent high-voltage switchgear fault diagnosis system based on heterogeneous graph structure learning, comprising the following modules:
[0038] Data acquisition module: used to collect, clean and process sensor data of intelligent high voltage switchgear. The sensor data is heterogeneous data, including temperature, local, mechanical characteristics and gas sensor data.
[0039] Data processing module: Based on the data processed by the data acquisition module, it uses an unsupervised anomaly detection algorithm to find outliers to distinguish between running and abnormal states. This module includes the following units:
[0040] The K-means algorithm unit is used to cluster the sensor data using the K-means algorithm to separate normal values and outliers in the sensor data.
[0041] The isolated forest algorithm unit is used to perform anomaly detection on the sensor data using the isolated forest algorithm to find outliers in the data.
[0042] The union of outlier data from K-means algorithm units and isolated forest algorithm units is used as the final filtered outlier data.
[0043] Switchgear Fault Diagnosis Database Construction Module: This module is used to build a switchgear fault diagnosis database. It uses crowdsourcing technology combined with expert knowledge to label fault types. Normal operating conditions are labeled as non-fault, and abnormal conditions are labeled as fault types.
[0044] Heterogeneous graph structure learning algorithm module: Used to design heterogeneous graph structure learning algorithms, find representation spaces that meet the similarity measurement of heterogeneous sensor data, automatically learn the interaction relationships of sensor networks of heterogeneous components in intelligent high-voltage switchgear, and mine network structure information to achieve accurate operation status assessment of high-voltage switchgear.
[0045] The advantages of this invention are:
[0046] 1. The fault diagnosis method for intelligent high-voltage switchgear provided by this invention collects and cleans sensor data from the high-voltage switchgear by deploying sensors, and constructs a high-voltage switchgear operating status database including multiple sub-databases. Addressing the problems of difficult and inefficient data annotation in constructing the anomaly detection database, an unsupervised anomaly detection algorithm is designed, combining the k-means algorithm's suitability for dense clusters and isolated forests and its sensitivity to anomalies. This solves the problem of low annotation efficiency for large amounts of normal operation data of high-voltage switchgear, accurately perceiving anomalies in the sensor data distribution. Expert crowdsourcing technology is used to construct an anomaly diagnosis database, significantly improving annotation efficiency and ensuring the accuracy and reliability of data annotation. Finally, a high-voltage switchgear fault diagnosis database is constructed, providing a data foundation for automatic fault diagnosis and data support for the training of deep models.
[0047] 2. The fault diagnosis method for intelligent high-voltage switchgear provided by this invention, after constructing a high-voltage switchgear fault diagnosis database, addresses the problem that existing graph structure representation models overly rely on manually designed graph structure data. It designs a heterogeneous graph structure learning algorithm and a decoupled heterogeneous network representation method for component sensors, breaking the constraint of graph neural networks based on the isomorphism assumption and mining a representation space that conforms to the similarity metric of heterogeneous sensor data. Through graph structure construction, graph structure modeling, and message passing mechanisms, it automatically learns the interaction relationships between the sensor networks of high-voltage switchgear components, captures the key network and connection structure leading to switchgear faults, accurately perceives the network chain reaction caused by component faults, and significantly improves the accuracy of high-voltage switchgear fault diagnosis. Real-time and accurate fault diagnosis of high-voltage switchgear is achieved by training the model and deploying a high-voltage switchgear fault diagnosis device. Attached Figure Description
[0048] Figure 1This is a flowchart of the intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning according to an embodiment of the present invention;
[0049] Figure 2 This is a flowchart illustrating the discovery of outliers in data according to an embodiment of the present invention;
[0050] Figure 3 This is a schematic diagram illustrating the construction process of the fault diagnosis database in an embodiment of the present invention;
[0051] Figure 4 This is a schematic diagram illustrating the construction process of the fault diagnosis database in an embodiment of the present invention. Detailed Implementation
[0052] 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 in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.
[0053] Please see Figure 1 This embodiment discloses a fault diagnosis method for intelligent high-voltage switchgear based on heterogeneous graph structure learning, including the following steps:
[0054] Step 1: Data collection, cleaning, and processing;
[0055] In this step, sensor data is collected, and the collected data is cleaned and processed to eliminate data noise and improve data quality.
[0056] The data cleaning and processing of the sensor-collected data includes: deleting duplicate data collected by the same sensor; filling missing values in the data with the mean of the corresponding sensor data; and normalizing the data using a maximum-minimum scaling method to ensure that the data maintains a consistent measurement scale, given that sensor monitoring components differ in type and distribution. For example, UHF partial discharge sensors monitor real-time changes in telecommunications within high-voltage switchgear. The waveform data acquired in minutes is used as the sensor's initialization vector, and missing points are filled with the mean of the vector before data normalization.
[0057] The following describes the sensor data for high-voltage switchgear:
[0058] 1. Types of sensors:
[0059] Temperature sensor: The temperature sensor is fixed to the moving contact and contact arm with screws. In addition, high-sensitivity temperature sensors are fixed at three points on the inner wall of the handcart compartment to indirectly measure the temperature rise of the contacts.
[0060] Partial discharge sensors: Ultrasonic partial discharge sensors, TEV transient ground wave partial discharge sensors, and UHF ultra-high frequency partial discharge sensors are used to measure partial discharge signals in the switch cabinet;
[0061] Mechanical characteristic sensors: The opening and closing coil current monitoring device and the switch mechanical characteristic sensor are installed on the secondary circuit of the instrument room and the circuit breaker operating shaft. The opening and closing coil current is collected from the opening and closing coil current monitoring device, and the opening and closing speed is collected from the switch mechanical characteristic sensor.
[0062] Gas sensor: When insulation components inside the high-voltage switchgear are damaged due to partial discharge and abnormal heating, the insulation material will deteriorate and decompose due to high temperature. A gas sensor unit is used to detect the gas content in real time.
[0063] 2. Sensor Heterogeneous Data Characterization
[0064] Sensor defects are usually a process; therefore, when designing sensor data characterization, this invention considers the data change process by collecting sensor data several times (about ten times or more) as feature vectors.
[0065] The structural data of the temperature sensor are characterized as follows:
[0066] Contact temperature Contact arm temperature Temperature at point 1 on inner wall Temperature at two points on the inner wall Temperature at 3 points on the inner wall
[0067] The partial discharge sensor structure data is characterized as follows:
[0068] Ultrasonic detection signal TEV transient ground wave signal UHF signal
[0069] The mechanical properties and structural data are characterized as follows:
[0070] Opening and closing coil current (A) t1 A t2 …A tn ...A t1 A t2 …A tn ); Opening and closing speed (v) t1 ,v t2 …v tn ...vt1 ,v t2 …v tn ).
[0071] The structural data of the gas sensor are characterized as follows:
[0072] Gas content (Q) t1 Q t2 …Q tn ...Q t1 Q t2 …Q tn ).
[0073] Step 2: Outlier detection;
[0074] See Figure 2 and Figure 3 In this step, based on the data processed in step 1, an unsupervised anomaly detection algorithm is used to find outliers to distinguish between running states and abnormal states. This includes the following steps:
[0075] Step 22: Use the K-means algorithm to cluster the temperature, local, mechanical characteristics and gas sensor data respectively. By designing the clustering category to 2, normal values and outliers in the four types of sensor data are separated.
[0076] The specific process of the K-means algorithm is as follows: Select two clusters, K=2, representing normal values and outliers respectively. Randomly select the center of each cluster, denoted as z1 and z2. Then, iterate through the data from each sensor cluster. In each iteration, calculate the distance from each sample x to the two cluster centers z1 and z2, and assign x to the cluster with the closer center. After traversing the data from each sensor cluster, calculate the mean of the two cluster sets and update it with the center value. Repeat the above clustering algorithm until the cluster centers remain unchanged, then stop the iteration and return the clustering result.
[0077] Step 24: Use the Isolation Forest algorithm to detect anomalies in temperature, local, mechanical properties, and gas sensor data to identify outliers in the data.
[0078] The specific process of the isolated forest algorithm is as follows: Take Ψ data points from each type of sensor as subsamples and place them into the root node of an isolated tree, where each data vector h = {h1, h2, ... h...} n},h iLet p represent the values under different dimensions. Randomly specify a dimension and, within the current node's data range, randomly generate a cut point p. Cut point p occurs between the maximum and minimum values of the specified dimension in the current node's data. Assign points less than p in the currently selected dimension to the left branch of the current node, and assign points greater than or equal to p to the right branch. The segmentation ends when a leaf node contains only one data point. By repeating the above algorithm, t isolated trees are obtained. Then, the entire sensor database is traversed. For each input sample x, the anomaly score s(x,Ψ) is calculated using the t isolated trees. The calculation process is as follows:
[0079]
[0080] Where Ψ is the number of samples, h(x) is the height of each tree, and c(Ψ) is the average path length for a given number of samples Ψ, used to standardize the path length h(x) of sample x. The final anomaly score is obtained by averaging the scores of each tree, and the k sample points with the highest anomaly scores are taken as the final anomalies.
[0081] Based on the different focuses of the two algorithms in steps 22 and 24, and considering the characteristics of large data volume and sparse outliers in heterogeneous sensors, combined with the high efficiency and dense clustering of k-means and the sensitivity of isolated forests to outliers, the union of the outlier data in steps 22 and 24 is adopted as the final filtered outlier data.
[0082] As a preferred technical solution, in step 22, the K-means algorithm is used to cluster the temperature, local and mechanical sensor data respectively, separate normal values and outliers, select the class with more samples as normal values, and the other class as outliers.
[0083] As a preferred technical solution, in step 24, the above algorithm is repeated to obtain t isolated trees. Generally, the value of t is selected as 100, at which point the system overhead is small and the results tend to converge.
[0084] As a preferred technical solution, in step 24, for each sample x, the anomaly score is calculated through t isolated trees to obtain the anomaly point.
[0085] Step 3: Build a switchgear fault diagnosis database;
[0086] See Figure 3 In this step, crowdsourcing technology combined with expert knowledge is used to label fault types. Normal operating conditions are labeled as non-faulty, while abnormal conditions are labeled as fault types. The anomaly detection algorithms proposed in steps 22 and 24 can distinguish between operating and abnormal states, and use expert knowledge to label the fault types of abnormal states, thus constructing a switchgear fault diagnosis database.
[0087] Fault types include: insulation damage, mechanical faults, temperature rise faults, misoperation and arcing faults, as well as composite types of multiple faults.
[0088] The crowdsourcing technology typically involves outsourcing the task of labeling the operating status of switchgear to relevant practitioners in power companies on a voluntary basis, thereby utilizing a large amount of human resources to complete the data labeling task.
[0089] The integration of expert knowledge refers to the evaluation and labeling of sensor data by high-voltage switchgear research experts. Since a single abnormal sensor data point cannot determine the fault type, the expert evaluation comprehensively considers the status of all current sensor data points and, in conjunction with specific examples, determines the nature of the fault.
[0090] Step 4: Design a Heterophilic Graph Structure Learning (HGSL) algorithm to find a representation space that meets the similarity metric of heterogeneous sensor data, automatically learn the interaction relationships of the sensor network of heterogeneous components in the intelligent high-voltage switchgear, and mine network structure information to achieve accurate assessment of the operating status of the high-voltage switchgear.
[0091] See Figure 4 In this step, data from the switchgear fault diagnosis database constructed in step 3 is read, training and validation sets are divided, a heterogeneous group component sensor network is constructed, a heterogeneous graph structure learning algorithm is designed and the model is trained on the training set, and the interaction relationship of the intelligent high-voltage switchgear group component sensor network is automatically captured. This includes three parts: graph structure construction, graph structure modeling and message passing.
[0092] Graph structure construction: A heterogeneous component sensor network is constructed based on the training set data in the switchgear fault diagnosis database. Sensors from various components in the high-voltage switchgear are used as nodes in the heterogeneous component sensor network, and edges in the network represent the interactions between components. First, the node features are initialized using the heterogeneous sensor data from the switchgear fault diagnosis database to obtain the coupled data feature matrix. The distance between each sensor is calculated using the cosine distance metric, i.e., vector dot product operation. Then, the number of neighbor nodes is set to k, and all nodes in the heterogeneous component sensor network of the high-voltage switchgear are traversed. For each component sensor node, the k nearest nodes are selected as neighbor nodes, and the edges between the component sensors are constructed to obtain the initialized heterogeneous component sensor network, i.e., the relation matrix A.
[0093] As a preferred approach, the corresponding heterogeneous component sensor network nodes are initialized using the data values recorded in various sensor databases. Then, the distance between the nodes and each component sensor is calculated. If k = 10, the 10 closest component sensor nodes are selected as neighbor nodes, and the edges between the component sensor nodes are constructed to obtain the interaction relationship graph of the initialized heterogeneous component sensor network, which is the relationship matrix A.
[0094] Graph structure modeling utilizes a decoupled graph convolutional neural network G to obtain aggregated information from heterogeneous component sensor network node nodes, updates the decoupled embedding X, and predicts the relation matrix A. * The specific steps are as follows: The decoupled embedding is updated after aggregating and updating the information of the sensor network nodes of the heterogeneous component group using a decoupled graph convolutional neural network G. H represents the decoupled sensor embedding of the component, and ReLU is the activation function. Let be the Laplacian matrix of the graph, representing the interaction relationships between nodes in a heterogeneous component sensor network, and let D be the degree matrix of the graph minus the relation matrix A. Due to the heterogeneity of the component sensor network, existing cosine similarity methods are insufficient to measure the similarity between the embedding nodes of heterogeneous component sensor networks. Therefore, a decoupling mapping function f is designed. k Mapping the input coupled data feature matrix X' to different decoupling spaces yields the decoupled component sensor embedding Z. k =f k (X′), where the decoupling mapping function f k The representation space transformation is performed using a two-layer MLP, where X' is the input coupled data feature matrix X', and Z... k To decouple and embed the component-based sensor in the k-th decoupling space, traverse all nodes in the heterogeneous component-based sensor network, calculate the similarity between each node and its neighboring nodes in each decoupling space, and select the decoupling space Z with the highest similarity. i Finally, the decoupled sensor component was selected and embedded in H=Z. i Using a decoupled graph convolutional neural network G through the Laplacian matrix of the graph. Message passing is performed, with each node in the heterogeneous component sensor network aggregating information from all its first-order neighbors to achieve information interaction. Information aggregation in the decoupled space breaks the isomorphism assumption, obtaining the similarity between heterogeneous components and performing heterogeneous information aggregation. After obtaining the decoupled embedding X updated through message passing, the similarity relationship between each node in the heterogeneous component sensor network is evaluated using a cosine similarity metric function. High similarity indicates high information correlation, so an edge is constructed between the two nodes. This process is repeated for all nodes in the heterogeneous component sensor network to update the relation matrix A.* .
[0095] The nodes in the sensor network of each heterogeneous component group are based on the Laplace matrix. It transmits information with its neighboring nodes to obtain information about the sensor components of its neighbors, thereby discovering the interaction relationships between nodes in the heterogeneous sensor component network.
[0096] Message passing: Construct a new graph convolutional neural network G* for message passing and information aggregation. G* is a ReLU (Gi) network. * (A * XW), where W is a linear layer, the initial embedding of a node is obtained through linear transformation, and the updated node embedding h∈R n+d Where n is the number of sensors in the component group, i.e., the number of nodes, and d is the dimension of the representation vector. Message passing aggregates information based on the edges in the heterogeneous component sensor network, and the network structure information can be implicitly learned during model training. This is input into the model pooling layer to obtain the overall state representation embedding of the high-voltage switchgear, and finally input into the model p = sigmoid(MLP(mean(h))) for fault diagnosis, where sigmoid is the activation function, MLP is a linear layer, and the model predicts the probability of belonging to each category.
[0097] The model's loss function is as follows:
[0098]
[0099] Where y i The actual label is marked. To predict the probability of being in the i-th category, where C is the fault category and m is the number of samples, the model is trained on the training set through backpropagation of loss and the optimal solution is obtained. By minimizing the loss, the model can automatically learn the structural information of the heterogeneous component sensor network related to fault diagnosis.
[0100] This invention constructs a complex heterogeneous component sensor network based on data from temperature sensors, partial discharge sensors, mechanical characteristic sensors, and gas sensors. It considers the interaction relationships between components, such as high-voltage switchgear faults being caused by abnormalities in multiple components or abnormalities in one component causing abnormalities in other components. In step 1, when collecting sensor data for characterization, it considers time-series changes, places sensors, and reduces data collection costs.
[0101] This invention proposes an unsupervised anomaly detection algorithm. Addressing the challenges of data labeling and low efficiency in constructing anomaly detection databases, this algorithm combines the advantages of k-means, which is suitable for dense clusters and isolated forests and sensitive to anomalies, to solve the problem of low labeling efficiency for high-voltage switchgear normal operation data. It accurately detects anomalies in sensor data distribution and utilizes expert crowdsourcing technology to construct an anomaly diagnosis database, significantly improving labeling efficiency while ensuring the accuracy and reliability of data labeling, thus resolving the issues of limited data volume and uneven data distribution.
[0102] This invention proposes a heterogeneous graph structure learning model, including graph construction, graph structure modeling, and message passing. Graph structure modeling decouples coupled heterogeneous data, breaking the isomorphism assumption. It mines the similarity features of heterogeneous data in the decoupled space and automatically learns the interaction relationships between various heterogeneous component sensors in a heterogeneous component sensor network. This eliminates the dependence on manually constructed graph structure data and automatically captures the network structure. Message passing implicitly learns network structure information, thereby achieving accurate and reliable fault diagnosis. The fault diagnosis model is deployed to achieve real-time diagnosis when a fault occurs.
[0103] According to the present invention, a fault diagnosis method for high-voltage switchgear based on heterogeneous graph structure learning is proposed. Compared with traditional live detection and graph neural network methods, the present invention avoids the over-reliance on artificially constructed component sensor network structures of traditional methods. Based on the heterogeneity of high-voltage switchgear data, a heterogeneous graph structure learning method is designed to decouple and couple the data. It breaks the isomorphism assumption of graph neural networks and finds a decoupling space suitable for measuring the similarity of nodes in the heterogeneous component sensor network of high-voltage switchgear for message passing. The network structure is automatically learned according to the fault diagnosis task, eliminating erroneous connections and noise from affecting the overall representation of the switchgear's operating status. Compared with the fault diagnosis method based on graph neural networks (84.3%), the accuracy is improved by 2.3%.
[0104] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A fault diagnosis method for intelligent high-voltage switchgear based on heterogeneous graph structure learning, characterized in that: Includes the following steps: Step 1: Collect, clean and process sensor data from intelligent high-voltage switchgear. The sensor data is heterogeneous, including temperature, local, mechanical characteristics and gas sensor data. Step 2: Based on the data processed in Step 1, use an unsupervised anomaly detection algorithm to find outliers to distinguish between running states and abnormal states, including the following steps: Step 22: Use the K-means algorithm to cluster the sensor data to separate normal values and outliers in the sensor data; Step 24: Use the Isolation Forest algorithm to perform anomaly detection on the sensor data and find outliers in the data; The union of the outlier data from steps 22 and 24 is taken as the final filtered outlier data. Step 3: Construct a switchgear fault diagnosis database. Use crowdsourcing technology combined with expert knowledge to label fault types. Mark normal operating conditions as non-fault and abnormal conditions as fault types. Step 4: Design a heterogeneous graph structure learning algorithm to find a representation space that meets the similarity metric of heterogeneous sensor data. Automatically learn the interaction relationships between the sensor networks of heterogeneous components in the intelligent high-voltage switchgear, and mine network structure information to achieve accurate operational status assessment of the high-voltage switchgear. Specifically, this includes: Read the data from the switchgear fault diagnosis database constructed in step 3, divide it into training and validation sets, construct a heterogeneous component sensor network, design a heterogeneous graph structure learning algorithm, and train the model on the training set to automatically capture the interaction relationships of the intelligent high-voltage switchgear component sensor network. This includes three parts: graph structure construction, graph structure modeling, and message passing. The graph structure is constructed as follows: a heterogeneous component sensor network is constructed based on the training set data in the switchgear fault diagnosis database. The component sensors in the high-voltage switchgear are used as nodes in the heterogeneous component sensor network, and the edges in the heterogeneous component sensor network represent the interaction relationships between components. Graph structure modeling utilizes a decoupled graph convolutional neural network G to obtain the relationship matrix of the decoupled embedded X updated after aggregating information from heterogeneous component sensor network nodes, and then predicting it. ; Message passing, constructing a new graph convolutional neural network G Perform message passing and information aggregation.
2. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 1, characterized in that: The sensor data collection and cleaning process includes: deleting duplicate data collected from the same sensor; filling missing values in the data with the mean of the corresponding sensor data; and normalizing the data using a maximum and minimum value scaling method.
3. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 1, characterized in that: The sensor data includes: Temperature sensor data: contact temperature data, contact arm temperature data, and temperature data from multiple points on the inner wall; Partial discharge sensor data: ultrasonic detection signal, TEV transient ground wave signal, UHF ultra-high frequency signal; Mechanical characteristic data: opening and closing coil current, opening and closing speed; Gas sensor data: gas content.
4. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 1, characterized in that: The specific process of the K-means algorithm is as follows: Select two clusters K, representing normal values and outliers respectively. Randomly select the center of each cluster and denot it as z1 and z2. Then iterate through the data of each type of sensor. In each iteration, calculate the distance of each sample x to the two cluster centers z1 and z2, and assign x to the cluster center with the closer distance. After traversing the data of each type of sensor, calculate the mean of the two cluster sets and update it as the center value. Repeat the above clustering algorithm until the cluster centers remain unchanged, then stop the iteration and return the clustering result. The specific process of the isolated forest algorithm is as follows: Take Ψ data points from each type of sensor as subsamples and place them into the root node of an isolated tree. Each data vector... For values across different dimensions, a dimension is randomly assigned. Within the current node's data range, a cut point p is randomly generated. Cut point p occurs between the maximum and minimum values of the specified dimension in the current node's data. Points in the currently selected dimension less than p are assigned to the left branch of the current node, and points greater than or equal to p are assigned to the right branch. The segmentation ends when a leaf node contains only one data point. By repeating the above algorithm, t isolated trees are obtained. Subsequently, the entire sensor database is traversed, and for each sample x, an anomaly score is calculated using the t isolated trees. The calculation process is as follows: Where Ψ is the number of samples, h(x) is the height of each tree, and c(Ψ) is the average path length for a given number of samples Ψ, used to standardize the path length h(x) of sample x. By averaging the scores of each tree, the final anomaly score is obtained, and the k sample points with the highest anomaly scores are taken as the final anomaly points.
5. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 1, characterized in that: The graph structure construction process is as follows: First, the node features are initialized using heterogeneous sensor data from the switchgear fault diagnosis database to obtain the coupled data feature matrix X'. The distance between each sensor is calculated using the cosine distance metric, i.e., vector dot product operation. Second, the number of neighboring nodes k is set, and all nodes in the heterogeneous group component sensor network of the high-voltage switchgear are traversed. For each group component sensor node, the k nearest nodes are selected as neighboring nodes, and the edges between group component sensors are constructed to obtain the initialized heterogeneous group component sensor network, i.e., the relation matrix A.
6. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 5, characterized in that: The graph structure modeling process is as follows: Information from heterogeneous component sensor network nodes is aggregated and updated using a decoupled graph convolutional neural network G, resulting in the decoupled embedding X = H represents the decoupled sensor embedding of the component, and ReLU is the activation function. , Let be the Laplacian matrix of the graph, representing the interaction relationships between nodes in a heterogeneous component sensor network; let D be the degree matrix of the graph minus the relation matrix A; design a decoupling mapping function. Mapping the input coupled data feature matrix X' to different decoupling spaces yields the decoupled component sensor embedding. The decoupling mapping function The representation space is transformed by a two-layer MLP. 'X' is the input coupled data feature matrix. To decouple and embed the component sensor in the k-th decoupling space, traverse all nodes in the heterogeneous component sensor network, calculate the similarity between the node and its neighboring nodes in each decoupling space, and select the decoupling space with the highest similarity. Finally, the decoupled component sensor was selected and embedded in H= Using a decoupled graph convolutional neural network G through the Laplacian matrix of the graph To facilitate message passing, each node in the heterogeneous component sensor network aggregates information from all its first-order neighbors, enabling information interaction. Information aggregation within the decoupled space breaks the isomorphism assumption, acquiring similarities between heterogeneous components and aggregating heterogeneous information. After obtaining the decoupled embedding X updated through message passing, the similarity relationship between nodes in the heterogeneous component sensor network is evaluated using a cosine similarity metric function. High similarity indicates high information correlation, thus establishing edges between two nodes. This process is repeated for all nodes in the heterogeneous component sensor network, thereby updating the relation matrix of the heterogeneous component sensor network. .
7. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 6, characterized in that: The specific message passing process is as follows: Construct a new graph convolutional neural network G. G performs message passing and information aggregation. for W is a linear layer, which obtains the initial embedding of nodes through linear transformation, and the updated node embeddings... Where n is the number of sensors in the component group, i.e., the number of nodes, and d is the dimension of the representation vector. Message passing aggregates information based on the edges in the heterogeneous component sensor network. During model training, network structure information can be implicitly learned and input into the model pooling layer to obtain the overall state representation embedding of the high-voltage switchgear, which is finally input into the model. Fault diagnosis is performed, where sigmoid is the activation function, MLP is a linear layer, and the model predicts the probability of belonging to each category.
8. The intelligent high-voltage switchgear fault diagnosis method based on heterogeneous graph structure learning as described in claim 7, characterized in that: The model's loss function is as follows: in The actual label is marked. To predict the probability of being in the i-th category, where C is the fault category and m is the number of samples, the model is trained on the training set through backpropagation of loss and the optimal solution is obtained. By minimizing the loss, the model can automatically learn the structural information of the heterogeneous component sensor network related to fault diagnosis.
9. A fault diagnosis system for intelligent high-voltage switchgear based on heterogeneous graph structure learning, characterized in that: Includes the following modules: Data acquisition module: used to collect, clean and process sensor data of intelligent high voltage switchgear. The sensor data is heterogeneous data, including temperature, local, mechanical characteristics and gas sensor data. Data processing module: Based on the data processed by the data acquisition module, it uses an unsupervised anomaly detection algorithm to find outliers to distinguish between running and abnormal states. This module includes the following units: The K-means algorithm unit is used to cluster the sensor data using the K-means algorithm to separate normal values and outliers in the sensor data. The isolated forest algorithm unit is used to perform anomaly detection on the sensor data using the isolated forest algorithm to find outliers in the data. The union of outlier data from K-means algorithm units and isolated forest algorithm units is used as the final filtered outlier data. Switchgear Fault Diagnosis Database Construction Module: This module is used to build a switchgear fault diagnosis database. It uses crowdsourcing technology combined with expert knowledge to label fault types. Normal operating conditions are labeled as non-fault, and abnormal conditions are labeled as fault types. Heterogeneous graph structure learning algorithm module: Used to design heterogeneous graph structure learning algorithms, find representation spaces that conform to the similarity metric of heterogeneous sensor data, automatically learn the interaction relationships of sensor networks of heterogeneous components in intelligent high-voltage switchgear, and mine network structure information to achieve accurate operational status assessment of high-voltage switchgear. Specifically, it includes: The system reads data from the switchgear fault diagnosis database constructed in the switchgear fault diagnosis database construction module, divides it into training and validation sets, constructs a heterogeneous component sensor network, designs a heterogeneous graph structure learning algorithm, trains the model on the training set, and automatically captures the interaction relationships of the intelligent high-voltage switchgear component sensor network. This process includes three parts: graph structure construction, graph structure modeling, and message passing. The graph structure is constructed as follows: a heterogeneous component sensor network is constructed based on the training set data in the switchgear fault diagnosis database. The component sensors in the high-voltage switchgear are used as nodes in the heterogeneous component sensor network, and the edges in the heterogeneous component sensor network represent the interaction relationships between components. Graph structure modeling utilizes a decoupled graph convolutional neural network G to obtain the relationship matrix of the decoupled embedded X updated after aggregating information from heterogeneous component sensor network nodes, and then predicting it. ; Message passing, constructing a new graph convolutional neural network G Perform message passing and information aggregation.