Multi-dimensional data fault diagnosis system for high-voltage switchgear of transformer substation

Abstract

The invention discloses a substation high-voltage switchgear multi-dimensional data fault diagnosis system, which comprises a multi-dimensional data acquisition module, a data processing module and a diagnosis decision module, and is characterized in that the multi-dimensional data acquisition module is used for acquiring multi-dimensional data information of substation high-voltage switchgear; the data processing module is used for preprocessing the collected multi-dimensional data information, and the diagnosis decision module comprises a historical data-based fault diagnosis unit, a data mining-based fault diagnosis unit, a machine learning-based fault diagnosis unit and a high-voltage switchgear-based fault diagnosis unit. The multi-dimensional data information of the switchgear is collected, information of different sources is integrated and analyzed, and support is provided for comprehensive and accurate diagnosis results. And meanwhile, fault diagnosis is performed on information of different sources in a targeted manner by adopting multiple fault diagnosis modes, so that the accuracy, effectiveness and reliability of fault diagnosis are improved, and clear guidance and requirements are provided for research of a fault diagnosis technology.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for high-voltage switchgear in substations, and more particularly to a multi-dimensional data fault diagnosis system for high-voltage switchgear in substations. Background Technology

[0002] Substations contain numerous main and auxiliary equipment, especially high-voltage switchgear used for controlling, protecting, and disconnecting circuits in the power system. To ensure the safe and stable operation of the power system, there is a wealth of status sensing devices and management operations for these switchgear. However, the integration and utilization of various data resources related to switchgear fault states, such as text, numerical values, images, and simulations collected from remote signaling, online monitoring, testing, and inspections, are insufficient. Existing fault diagnosis methods suffer from fragmented data, and diagnostic results often require the joint participation of experts from various fields, involving the collection of data from multiple sources to identify faults. This lack of accuracy and precision makes them unsuitable for the current intelligent development needs of the power grid. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned problems by providing a multi-dimensional data fault diagnosis system for high-voltage switchgear in substations. This system collects multi-dimensional data from the switchgear, integrates and analyzes information from different sources, and provides support for generating comprehensive and accurate diagnostic results. Furthermore, it employs multiple fault diagnosis modes to target fault diagnoses based on information from different sources, improving the accuracy, effectiveness, and reliability of fault diagnosis. This provides clear guidance and requirements for research in fault diagnosis technology.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: According to one aspect of the present invention, a multi-dimensional data fault diagnosis system for high-voltage switchgear in substations is provided, comprising a multi-dimensional data acquisition module, a data processing module, and a diagnostic decision module; The multidimensional data acquisition module is used to acquire multidimensional data information from high-voltage switchgear in substations. The data processing module is used to preprocess the collected multidimensional data information; The diagnostic decision module includes a fault diagnosis unit based on historical data, a fault diagnosis unit based on data mining, a fault diagnosis unit based on machine learning, and a fault diagnosis unit based on high-voltage switchgear. The historical data-based fault diagnosis unit is used to diagnose faults based on the historical data of the equipment. The data mining-based fault diagnosis unit is used to diagnose hidden fault data. The machine learning-based fault diagnosis unit is used for fault diagnosis of high-dimensional, multi-source, and small-sample data. The fault diagnosis unit based on high-voltage switchgear utilizes multiple models to diagnose faults in equipment data.

[0005] Preferably, the multi-dimensional data acquisition module includes a body information acquisition unit, an anomaly information acquisition unit, an online data monitoring unit, and a live-line detection data acquisition unit; The body information acquisition unit is used to collect basic data on the operation of power equipment; The abnormal information collection unit is used to collect any information, data or events that are inconsistent with the normal state or behavior during the operation of the equipment. The online data monitoring unit is used to collect and transmit data to the data processing center in real time; The live-line monitoring data acquisition unit is used to collect equipment status information obtained through live-line detection technology when the electrical equipment is in operation.

[0006] Preferably, the data processing module includes a data fusion unit, a feature extraction unit, a feature fusion unit, and a fault data association rule analysis unit; The data fusion unit is used to fuse multidimensional data; The feature extraction unit is used to extract feature data that characterizes the operating status of the equipment; The feature fusion unit is used to fuse feature data; The fault data association rule analysis unit is used to generate data association rules.

[0007] Preferably, the historical data fault diagnosis unit includes a data cleaning subunit and a model building subunit. The data cleaning subunit processes the historical data using the KNN interpolation algorithm, and the model building subunit is used to build an XGBoost model and perform fault diagnosis using the XGBoost model.

[0008] Preferably, the data mining fault diagnosis unit includes a data preprocessing subunit and a model building subunit. The data preprocessing subunit processes the data using the isolated forest algorithm, and the model building subunit is used to build a K-Means model and perform fault diagnosis using the K-Means model.

[0009] Preferably, the machine learning fault diagnosis unit includes a data stream processing subunit and a model building subunit. The data stream processing subunit is used to process real-time data streams, and the model building subunit is used to build an XGBoost model and use the XGBoost model for fault diagnosis.

[0010] Preferably, the fault diagnosis unit based on high-voltage switchgear includes a gas pressure logic diagnosis model, an oil pump pressurization anomaly diagnosis model, an infrared thermal diagnosis model, a switching waveform diagnosis model, a disconnecting switch switching status diagnosis model, a mechanical characteristic diagnosis model, a local diagnostic model, and an X-ray detection and analysis model. The gas pressure logic diagnostic model is used for fault diagnosis of SF6 gas pressure; the oil pump pressurization anomaly diagnostic model is used for fault diagnosis of oil pressure in hydraulic energy storage mechanisms; the infrared thermal diagnostic model is used for fault diagnosis of equipment operating temperature; the circuit breaker opening and closing waveform diagnostic model is used for fault diagnosis of closing current time-domain waveform; the disconnector switch opening and closing status diagnostic model is used for detection of disconnector switch equipment targets; the mechanical characteristic diagnostic model is used for fault diagnosis of mechanical characteristic waveform data; the local diagnostic model is used for fault diagnosis of discharge signals; and the X-ray detection and analysis model is used for fault diagnosis of the internal structure of GIS and switchgear.

[0011] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention collects multidimensional data from switching equipment, integrates and analyzes information from different sources, and provides support for producing comprehensive and accurate diagnostic results. Simultaneously, it employs multiple fault diagnosis modes to perform targeted fault diagnosis based on information from different sources, improving the accuracy, effectiveness, and reliability of fault diagnosis, and providing clear guidance and requirements for research in fault diagnosis technology. Attached Figure Description

[0012] Figure 1 This is a functional structure diagram of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0014] Please see Figure 1 This invention provides a multi-dimensional data fault diagnosis system for high-voltage switchgear in substations, and the technical solution is as follows: The multi-dimensional data fault diagnosis system for high-voltage switchgear in substations includes a multi-dimensional data acquisition module, a data processing module, and a diagnostic decision module.

[0015] The multi-dimensional data acquisition module includes a body information acquisition unit, an anomaly information acquisition unit, an online data monitoring unit, and a live-line detection data acquisition unit. The body information acquisition unit collects basic data on the operation of the power equipment. The anomaly information acquisition unit collects any information, data, or events that deviate from the normal state or behavior during equipment operation. The online data monitoring unit collects and transmits data to the data processing center in real time using sensors, data acquisition devices, and other equipment. The live-line monitoring data acquisition unit collects equipment status information obtained through live-line detection technology while the electrical equipment is in operation. All of the above data is collected through various sensors. The multi-dimensional data of high-voltage switchgear includes key indicators such as electrical quantities (e.g., voltage, current, power), mechanical characteristics (e.g., contact stroke, speed, vibration), temperature (e.g., contact temperature, busbar temperature), and insulation performance (e.g., insulation resistance, dielectric loss tangent, partial discharge).

[0016] The entity information acquisition unit includes the identification information and performance boundaries of the high-voltage switchgear. The identification information includes the equipment model, manufacturer, manufacturing date, and unique serial number. The technical specifications within the performance boundaries include rated voltage, rated current, short-time withstand current, and mechanical life, clearly defining the equipment's operational capabilities under normal and extreme conditions. It also includes information recorded by the operation and maintenance (O&M) system, including equipment commissioning information, maintenance history, family defect data, and major modification logs. This forms a dynamic health record spanning the equipment's service life. These parameters collectively constitute the benchmark for equipment selection and matching, condition degradation analysis, and decommissioning assessment, supporting standardized management and intelligent application of multi-dimensional data.

[0017] In the anomaly information acquisition unit, anomalies are categorized according to their type. These types are classified based on failure mechanisms into standardized categories such as mechanical system malfunctions, electrical performance degradation, thermodynamic imbalances, and environmentally related faults. The system is aligned with industry-standard coding systems, supporting cross-system data interaction and historical case matching. Anomaly time information is recorded using millisecond-level timestamps. Combined with equipment operating cycles, operation frequency, and environmental parameters, the system analyzes the suddenness, periodicity, or gradual nature of faults, providing a time-series analysis basis for fault mode identification and risk trend prediction. Anomaly locations are precisely pinpointed using the equipment's three-dimensional coordinate system and hierarchical topology, clarifying the physical level and spatial distribution of the fault. This is synchronized with the design parameters and maintenance records of the equipment components, forming a multi-granularity positioning capability from the macroscopic whole machine to the microscopic components. Through spatiotemporal feature fusion and multi-source data cross-validation, the system constructs a dynamic model for fault evolution paths, enabling root cause tracing and optimization of handling strategies. This drives the transformation of maintenance response from passive handling to proactive prevention, effectively improving equipment reliability management and maintenance resource allocation efficiency.

[0018] In the online data monitoring unit, the online monitoring system captures the real-time operating status of high-voltage switchgear through a multi-dimensional sensor network, providing holistic data support for fault early warning and health assessment. Monitoring data includes: Current and voltage data: Real-time acquisition of current and voltage data from high-voltage switchgear. High-precision current and voltage sensors ensure the accuracy and reliability of the data. This data can be used to assess the operating status of the equipment and promptly detect any abnormalities.

[0019] SF6 gas pressure data: By accessing real-time SF6 gas pressure data and integrating multi-dimensional features such as ambient temperature, historical data, and phase data, continuous tracking and accurate evaluation of the equipment's sealing status can be achieved.

[0020] Oil pump pressure anomaly data: Real-time acquisition of oil pump operating data, such as pressure and start-up frequency, combined with analysis of the hydraulic system's working status. Through high-precision pressure sensors and frequency counters, pressure changes and start-up counts of the oil pump during operation can be accurately recorded, allowing for timely detection of anomalies and the issuance of alarms.

[0021] Opening and closing waveform data: Dynamically capture the current and voltage waveform characteristics during switch operation, such as pre-breakdown and arcing time, and diagnose mechanical and electrical linkage faults through waveform distortion analysis.

[0022] Disconnector switch opening and closing status data: The position sensor provides real-time feedback on the contact position status, and the motor current curve is used to diagnose opening and closing jams. The number of operations is recorded simultaneously to assess the mechanical life.

[0023] Mechanical characteristic data: Continuously monitor motion parameters such as contact stroke, speed, and vibration, construct dynamic characteristic curves for opening and closing, identify wear of transmission components or fatigue of springs, and support quantitative assessment of mechanical health.

[0024] In the live-line detection data acquisition unit, the live-line detection data consists of equipment status data collected by specialized instruments and sensors during the operation of electrical equipment. This data can be used to assess the operating status and health of the equipment. Specifically, it includes the following data: Infrared thermal data: Based on non-contact detection technology using infrared thermal imagers, it captures the surface temperature distribution characteristics of equipment to identify defects such as overheating of contacts and abnormal internal heating. It is suitable for periodic temperature inspection and hotspot location analysis of live equipment.

[0025] Partial discharge data: High-frequency current, ultrasonic or ultra-high frequency sensors are used to detect partial discharge signals inside the equipment. The degree of insulation degradation is determined by time-frequency domain feature analysis. It is mainly used for insulation status diagnosis of key components such as cable terminals and GIS equipment.

[0026] X-ray data: Utilizing penetrating X-ray imaging technology, it allows for the visualization of internal mechanical structures without disassembling the equipment, detecting contact wear, component deformation, or assembly abnormalities. It is suitable for specialized in-depth inspections of the mechanical condition of switchgear.

[0027] The above information and data were collected through various methods, as detailed below: Sensor Data Acquisition: Multiple types of sensors are deployed in key areas of the high-voltage switchgear to monitor operating status in real time. Temperature sensors track temperature rise at contacts and connection points, vibration sensors capture mechanical impact signals, and gas sensors detect the state of the insulating medium. Sensors transmit data back in real time via wired industrial networks or low-power wireless communication. Combined with edge computing nodes for local filtering and anomaly detection, this ensures data timeliness and reliability, providing high-precision raw data for subsequent analysis.

[0028] Remote monitoring system data acquisition: A cloud-edge collaborative remote monitoring architecture is constructed to achieve comprehensive equipment status awareness. Data preprocessing and real-time alarms are completed at the edge, while historical data is centrally stored and intelligent analysis models are run in the cloud. The system integrates power grid operating parameters and environmental data, dynamically analyzes equipment health status through a streaming computing engine, supports remote control command issuance and risk warnings, forming a closed-loop management system of "monitoring-analysis-response".

[0029] Manual inspection and data collection: As an effective supplement to automated monitoring, manual inspection focuses on verifying complex anomalies and hidden defects. Portable equipment such as infrared thermal imagers and ultrasonic detectors are used, combined with visual inspection to record equipment appearance anomalies and mechanical loosening. Inspection results are digitally archived through terminals and cross-validated with automated data to optimize the diagnostic rule base. Inspection cycles are dynamically adjusted based on equipment health status to improve resource utilization and ensure timely closed-loop handling of potential hazards.

[0030] Sensor networks, remote systems, and manual inspections form a multi-level data complementarity, with automated monitoring achieving high-frequency continuous coverage, manual inspections providing in-depth verification, and remote systems coordinating analysis and decision-making.

[0031] The data processing module includes a data fusion unit, a feature extraction unit, a feature fusion unit, and a fault data association rule analysis unit. In the data fusion unit, a synchronous acquisition and standardized processing flow for multi-source data is first established. A unified time synchronization mechanism coordinates the data acquisition rhythms of different sensors, eliminating timing discrepancies and ensuring the temporal consistency of heterogeneous data such as electrical quantities, mechanical quantities, and environmental parameters. Simultaneously, diverse raw data, including switch states, vibration signals, and temperature changes, undergo format conversion and cleaning to filter noise interference and outliers, forming a standardized basic dataset that provides complete and reliable data support for subsequent analysis.

[0032] In the feature extraction unit, core feature parameters that characterize the operating status of the equipment are extracted.

[0033] The feature fusion unit focuses on multi-dimensional feature extraction and correlation analysis. It extracts key parameters reflecting equipment operating status from electrical data, identifies dynamic indicators characterizing mechanism motion from mechanical data, and captures changing trends affecting equipment performance from environmental data. By establishing a cross-dimensional feature correlation model, it uncovers the inherent connections between different data, transforming isolated information such as electrical anomalies, mechanical vibrations, and temperature fluctuations into comprehensive judgment criteria. This enhances the ability of feature parameters to characterize potential faults and addresses the limitations of a single data source for diagnosis.

[0034] The fault data association rule analysis unit combines time series analysis and correlation mining techniques to analyze the potential correlation patterns between parameters in historical data, constructing a multi-dimensional feature matrix to provide highly reliable input data for subsequent diagnostic models. Specifically, fault data association rule analysis includes the following steps: 1. Data Preprocessing: Data quality is improved through cleaning and discretization, providing structured input for subsequent analysis. Data cleaning employs noise filtering, missing value repair, and duplicate data removal. Statistical methods and physical models are combined to verify data rationality, ensuring complete and reliable information. Low-quality data is flagged using a credibility score, triggering manual review. Data discretization converts continuous parameters into discrete intervals, classifying them based on business thresholds and distribution characteristics to adapt to association rule mining needs while enhancing feature interpretability. The discretization strategy is dynamically adjusted to adapt to equipment aging and environmental changes, optimizing model generalization capabilities. This process, by improving data consistency, reducing redundant dimensions, and enhancing business adaptability, builds a high-credibility data foundation for fault diagnosis models, supporting accurate analysis and rule generation.

[0035] 2. Association Rule Mining: In association rule mining, the FP-Growth algorithm is used to efficiently mine strong association rules in multidimensional data by constructing a frequent pattern tree (FP-Tree). This method is suitable for the analysis of high-dimensional, real-time monitoring data of power equipment. The following is its specific implementation process in fault diagnosis: (1) Construction of transaction database: Time window segmentation: A 5-minute sliding window is used to divide the real-time data stream to ensure that each window contains equipment status characteristics, such as oil pump pressurization frequency, contact speed deviation and other parameters.

[0036] Feature discretization: Convert continuous parameters into discrete labels, such as temperature being divided into low temperature, normal, and high temperature levels, and pressure being divided into stable, fluctuating, and decreasing levels.

[0037] Transaction item generation: Combine the discretized features into transaction items. For example, combine the oil pump pressurization frequency of more than 3 times per hour and the contact speed deviation of more than 15% into one transaction.

[0038] (2) FP-Tree Construction and Compression Frequent item filtering: Statistically analyze the support of individual items, filter out rare items with support below 0.1%, such as electromagnetic interference signals, and retain high-frequency features.

[0039] Tree structure construction: Sort the itemsets in descending order of support, construct an FP-Tree, and store the item name, count, and node links in each node.

[0040] Tree compression: Merge duplicate paths, such as oil pump failure to contact wear and oil pump failure to partial discharge, to enhance shared parent nodes and reduce redundant storage.

[0041] (3) Conditional pattern base mining Suffix path extraction: For frequent items such as SF6 pressure anomaly, extract all prefix paths in the FP-Tree to generate conditional pattern bases.

[0042] Recursive subtree construction: Recursively mine multi-layer associations from conditional pattern bases, such as extending from SF6 pressure anomalies to combinations of contact velocity decrease and infrared thermographic anomalies.

[0043] (4) Generation of association rules Confidence calculation: Calculate the confidence of rules based on frequent itemsets, such as the probability of C occurring when A and B occur simultaneously, and retain strong rules with a confidence level higher than 85%.

[0044] Rule filtering: Combine the physical characteristics of the device to eliminate false rules. For example, the correlation between high ambient temperature and enhanced partial discharge needs to be verified to see if it is sensor interference.

[0045] Rule classification: Rules are classified into levels based on support and confidence. For example, a level 1 rule must have a support of more than 5% and a confidence of more than 90%.

[0046] (5) Real-time diagnosis and update Incremental update: The FP-Tree is dynamically updated every 15 minutes to adapt to gradual changes in equipment status, such as slow pressure decay caused by aging seals.

[0047] Knowledge graph integration: Import rules into a graph database to build a triplet reasoning network of fault modes, feature combinations and handling measures.

[0048] Fault matching: Real-time matching of operating data with the rule base, triggering early warnings and generating maintenance suggestions, such as checking the transmission linkage when the tripping time is extended or the vibration energy entropy exceeds the limit.

[0049] 3. Rule Evaluation and Screening: Evaluation metrics are used to assess the effectiveness of association rules using indicators such as lift. A lift greater than 1 indicates a positive correlation between two data items; the higher the lift, the stronger the correlation. Based on the evaluation metrics, meaningful and valuable association rules are selected. Redundant and invalid rules are removed, retaining only those that provide important information for fault diagnosis.

[0050] 4. Rule Base Integration and Application: Construct a knowledge graph, transforming rules into fault mode-feature combination-disposal suggestion triples, and import them into the Neo4j graph database to support semantic reasoning. A real-time matching engine is deployed, using a lightweight rule engine to compare device status data with the rule base in real time, triggering multi-level alerts. An incremental learning mechanism updates the FP-Tree incrementally every 24 hours, incorporating the latest fault cases and continuously optimizing the rule base coverage.

[0051] Fault data association rule analysis is an important technique in data mining, used to discover the relationships between different data items in a dataset. In high-voltage switchgear fault data, there may be underlying connections between different fault characteristics and types. Association rule analysis can extract valuable correlation information from large amounts of historical fault data, providing a basis for fault diagnosis. When performing correlation analysis, the dynamic correlation of equipment operating parameters is quantitatively analyzed to uncover the statistical relationship between fault characteristics and potential causes. Based on statistical methods, the correlation strength of multi-source parameters (such as electrical, mechanical, and environmental data) is calculated, and core parameter combinations strongly correlated with faults are selected to construct a multi-dimensional feature matrix. This supports the feature input of intelligent diagnostic models, improving the accuracy of fault root cause localization. When performing trend analysis, the long-term evolution patterns of key parameters are analyzed by combining equipment lifecycle data. Time-series prediction algorithms capture performance degradation trends, and dynamic threshold settings identify abnormal behaviors deviating from baselines. Aging rates are quantified, and potential fault windows are predicted, driving the transformation of maintenance strategies from post-event handling to pre-event prevention. When performing fault tree analysis, the fault cause chain is decomposed based on a logical reasoning model, and critical failure paths are located through probability calculations. By dynamically updating fault logic nodes based on real-time monitoring data, maintenance personnel can quickly identify high-probability root causes and optimize maintenance paths and resource scheduling. The fault tree and diagnostic rule base work together to form a closed-loop reasoning from symptoms to root causes, improving the efficiency of handling complex faults.

[0052] The fault diagnosis unit based on historical data includes a fault diagnosis method based on historical data, the specific method including the following steps: 1. Data Preprocessing and Feature Engineering Data cleaning: Box plots are used to detect and remove outliers caused by sensor drift, such as abnormal data points with pressure exceeding 5 MPa. For missing values, KNN interpolation is used to impute data segments with a missing rate of less than 5% to ensure data continuity. For fields with high missing rates, selective removal is performed after manual review based on equipment operation logs.

[0053] Feature construction: Extract time-domain features of the data, including current harmonic distortion rate, vibration signal mean and variance. Calculate the frequency domain dominant frequency amplitude using Fast Fourier Transform, and extract vibration energy entropy using wavelet packet decomposition. Discrete chemical condition labels, dividing ambient temperature and humidity into three ranges: low temperature, normal temperature, and high temperature; and classifying equipment operating years into three categories: 0-5 years, 5-10 years, and over 10 years.

[0054] 2. Model Training and Optimization Parameter configuration: Initialize XGBoost model parameters, setting the objective function to binary logistic regression, the number of trees to 500, the maximum depth to 6 layers, the row sampling rate to 0.8, the column sampling rate to 0.7, the learning rate to 0.05, the L2 regularization coefficient to 1.5, and the AUC value as the evaluation metric. To address the issue of a low proportion of faulty samples, the class weight parameter `scale_pos_weight` is set to 20 to improve the minority class identification capability.

[0055] Hyperparameter optimization: A Bayesian optimization algorithm was used to search for the optimal parameter combination, focusing on adjusting the maximum depth of 4-8 layers and the learning rate in the range of 0.01-0.1. The model performance was evaluated using 5-fold cross-validation, and the parameter combination with the highest AUC value and the lowest risk of overfitting was selected.

[0056] 3. Model Interpretation and Decision Support Feature Importance Visualization: The built-in feature importance assessment interface of XGBoost is called to output the top 10 key features sorted by information gain, such as vibration energy, tripping time deviation, and SF6 leakage rate. A bar chart is generated to visually demonstrate the contribution of each feature to fault prediction, guiding maintenance personnel to focus on core monitoring parameters.

[0057] SHAP Value Analysis: The SHAP interpretation tool is used to analyze individual prediction results and calculate the marginal impact of each feature on the probability of a single failure. High-impact features are displayed through a force-directed graph; for example, prolonged tripping time increases the probability of failure, while normal ambient temperature reduces risk. A list of key feature contributions is included with the diagnostic report to aid in root cause localization.

[0058] XGBoost has become the preferred algorithm for historical data-driven fault diagnosis due to its high accuracy, feature importance analysis capabilities, and advantages in processing heterogeneous data.

[0059] The data mining-based fault diagnosis unit includes a data mining-based fault diagnosis method, which specifically includes the following steps: 1. Data Preprocessing and Feature Engineering Data cleaning: The Isolation Forest algorithm is used to detect and remove outliers, such as sudden changes in pressure or temperature data caused by sensor drift or electromagnetic interference. For missing values, if the missing rate of a certain feature exceeds 10%, that feature is directly removed; the remaining missing values ​​are imputed using the mean of historical data from similar devices to ensure data integrity.

[0060] Feature standardization: Z-score standardization is performed on multi-dimensional parameters to convert data of different dimensions such as pressure, vibration amplitude, and temperature into a standard distribution with a mean of 0 and a variance of 1, thereby eliminating the influence of dimensional differences on clustering results.

[0061] Feature dimensionality reduction: Principal component analysis is used to compress high-dimensional features to 10-dimensional principal components, retaining 95% of the original information. For example, the original 50-dimensional mechanical vibration spectrum, current harmonics, and other features are converted into low-dimensional orthogonal vectors, significantly improving the efficiency of subsequent clustering.

[0062] 2. Implementation of cluster analysis Optimal K-value selection: The optimal number of clusters is determined by combining the silhouette coefficient and the elbow rule. The silhouette coefficient measures the similarity of a sample to its own cluster and other clusters, ranging from -1 to 1; a value closer to 1 indicates better clustering. The elbow rule selects the K-value at the inflection point by plotting the sum of squared errors (SSE) curves for different K-values. For example, when K=5, the silhouette coefficient reaches its peak of 0.65, and the SSE decrease trend slows significantly, indicating the optimal number of clusters.

[0063] Clustering execution: The K-Means++ algorithm is used to initialize cluster centers, avoiding the local optima problem caused by traditional random initialization. A maximum of 300 iterations is set, with a tolerance error of 1e-4, and the cluster labels for each sample are output. During the algorithm iteration process, the cluster center positions are dynamically updated until convergence.

[0064] Visual verification: The high-dimensional data was mapped to a two-dimensional space using the t-SNE algorithm, and a scatter plot was drawn to observe the cluster separation effect. The normal operating condition data was concentrated in a specific area, and the mechanical fault and insulation fault clusters showed obvious spatial separation, verifying the rationality of the clustering.

[0065] 3. Extraction of association rules Cluster feature analysis: Calculate characteristic statistics for each cluster to identify key differential features. For example, the mean contact velocity of mechanically faulty clusters is 30%-50% lower than that of normal clusters, and the standard deviation of vibration energy entropy is significantly increased.

[0066] Rule generation: The Apriori algorithm is used to mine intra-cluster feature associations, with a minimum support of 0.5% and a minimum confidence of 75%. A rule is generated such as "When the vibration energy entropy is greater than 2.5 and the daily SF6 pressure drop rate exceeds 0.05 MPa, the probability of seal failure reaches 88%".

[0067] Rule optimization: Filter out false rules by incorporating the physical mechanisms of the equipment. For example, the statistical correlation between ambient temperature and partial discharge pulse count needs to be verified using the thermal expansion coefficient of the insulating material to eliminate rules with no actual causal relationship.

[0068] 4. Diagnostic model construction and validation Rule base construction: Valid rules are encoded into IF-THEN structures and arranged in descending order of confidence level. For example, high-confidence rules are triggered first, while low-confidence rules serve as auxiliary references.

[0069] Online diagnostics: Real-time data, after standardization and dimensionality reduction, is input into a K-Means model to predict its cluster, triggering the corresponding cluster's association rule base to output diagnostic conclusions. For example, when new data is classified into the mechanical fault cluster, the system automatically checks whether the contact speed and vibration characteristics match preset rules.

[0070] Performance validation: Randomly select 20% of historical data to validate the model, calculate the F1-score to evaluate the diagnostic accuracy, with a target value exceeding 85%. Statistically analyze the false alarm rate under normal operating conditions, with a target of keeping it below 5%.

[0071] Utilizing data mining techniques, the multi-dimensional data fault diagnosis platform for high-voltage switchgear employs the K-Means clustering algorithm to extract valuable association rules from large amounts of fault data. These rules help identify fault characteristics, improving the accuracy and efficiency of fault diagnosis.

[0072] The Means algorithm is an unsupervised clustering algorithm whose core objective is to divide data samples into K mutually exclusive clusters, maximizing the similarity of samples within the same cluster and maximizing the difference between different clusters. The algorithm iteratively optimizes the cluster center positions, minimizing the sum of squared Euclidean distances from each sample to its respective cluster center. Its advantages include high computational efficiency and ease of implementation, making it suitable for the rapid discovery of hidden fault patterns in massive amounts of power equipment data, especially in scenarios lacking prior fault labels.

[0073] The machine learning-based fault diagnosis unit includes a machine learning-based fault diagnosis method, which includes the following steps: 1. Real-time data stream processing Data Acquisition and Caching: Multiple sensor signals, including parameters such as pressure, vibration, current, and temperature, are acquired in real time via industrial gateways or edge computing nodes. The sampling frequency is set from 1Hz to 10kHz based on the signal characteristics. A circular buffer is used to temporarily store data to ensure temporal continuity. The window length is typically set to 1 second to accommodate real-time feature calculation requirements.

[0074] Streaming Feature Extraction: Parallel computation of time-domain and frequency-domain features: Time-domain features include the mean, variance, and waveform factor within a sliding window, used to capture the overall trend of the signal; frequency-domain features extract the main frequency amplitude through Fast Fourier Transform and calculate energy entropy using wavelet packet decomposition algorithm to reveal transient anomalies in the signal. The event correlation module synchronizes the timestamps of equipment operation commands to ensure precise alignment between mechanical actions and electrical signals.

[0075] 2. Lightweight model deployment Model conversion and compression: The trained XGBoost model is converted to the ONNX open format, and the model weights are compressed from FP32 precision to FP16 using quantization technology, reducing the model size to 150MB, which is compatible with embedded devices such as NVIDIA Jetson Nano or Huawei Atlas 200DK.

[0076] Inference engine acceleration: Integrates the TensorRT acceleration library, enables multi-threaded parallel computing, and keeps the time for a single inference operation within 10 milliseconds. Memory usage is optimized through memory pool management technology to ensure stable operation within a 200MB memory limit, while also supporting multi-task concurrent processing.

[0077] 3. Online reasoning and early warning decision-making Real-time feature input and standardization: Input feature vectors need to undergo Z-score standardization to eliminate dimensional differences. For example, contact speed deviation is converted into a multiple of standard deviation, and SF6 pressure drop rate is normalized as a percentage of equipment rating.

[0078] Dynamic multi-level early warning mechanism: The threshold is dynamically adjusted according to the output probability: when the probability of early warning is 60%-80%, abnormal events are recorded and trend analysis is initiated; when the probability of abnormality is 80%-95%, a work order is generated and pushed to the operation and maintenance system; when the probability of emergency exceeds 95%, the equipment protection program is immediately triggered, such as cutting off the power or starting the redundant system.

[0079] 4. Incremental learning and hot model updates Incremental data caching and labeling: Stores the original data fragments that trigger alerts, including data from a 30-second window before and after the event. After on-site verification by operations and maintenance personnel, the actual fault type is labeled, forming a labeled incremental dataset, which is then stored in the time-series database.

[0080] Online parameter fine-tuning: Incremental training is performed during the low-load period in the early morning each day, limiting the number of new trees to no more than 50, and the learning rate is adjusted to 0.02 to prevent overfitting. Historical knowledge is preserved through elastic weight solidification technology to ensure compatibility between old and new data patterns.

[0081] Model hot-switch validation: The old and new models are run in parallel, and the consistency of prediction results is compared within 24 hours. When the new model improves the F1-score on the validation set by more than 2% and the false positive rate does not increase significantly, the switch to the new model is seamless, with a service interruption time of less than 1 second.

[0082] 5. Real-time monitoring and closed-loop optimization Core performance metric tracking: Continuously monitor metrics such as inference latency (P99), fault detection rate, and false alarm rate. Set thresholds to trigger alarms, for example, triggering manual intervention when the fault detection rate is below 85% for two consecutive hours.

[0083] Data distribution drift detection: Calculate the KL divergence or JS distance between the features of real-time data and training data to detect sensor drift or changes in operating conditions. When the KL divergence exceeds 0.1, the data distribution calibration process is automatically triggered to recalculate the standardized parameters.

[0084] Automatic retraining mechanism: If performance metrics continue to deteriorate for three consecutive days, the system automatically initiates full-data retraining. Training based on 1 million samples is completed within 2 hours, and the updated model is validated through A / B testing.

[0085] In high-voltage switchgear fault diagnosis scenarios, the XGBoost algorithm is also used. This algorithm demonstrates significantly superior accuracy, efficiency, and interpretability compared to traditional methods when processing high-dimensional, multi-source, and small-sample data.

[0086] The fault diagnosis unit for high-voltage switchgear includes a gas pressure logic diagnosis model, an oil pump pressurization anomaly diagnosis model, an infrared thermal diagnosis model, a switching waveform diagnosis model, a disconnector switching status diagnosis model, a mechanical characteristic diagnosis model, a local authority diagnosis model, and an X-ray detection and analysis model. Specifically: Gas Pressure Logic Diagnostic Model: By constructing an intelligent diagnostic model for SF6 gas pressure, this model integrates real-time gas pressure monitoring data, fuses ambient temperature changes, historical pressure curves, and interphase pressure correlation characteristics, and establishes a multiple linear regression calculation framework to achieve dynamic analysis of gas production rate and relative gas production rate, possessing the ability to capture micro-pressure fluctuations. The model's built-in logic reasoning module, based on historical equipment operation data and environmental parameter evolution patterns, constructs a temperature compensation correction algorithm to achieve accurate quantification of gas leakage rate (accuracy 0.02 MPa / month), supporting gas pressure trend prediction. The system can provide early warnings of abnormal pressure evolution paths, effectively improving the sensitivity of gas status monitoring and the timeliness of operation and maintenance decisions.

[0087] Oil Pump Pressure Anomaly Diagnosis Model: Hydraulic energy storage mechanisms operate under high oil pressure for extended periods. Affected by factors such as oil particle contamination, deteriorated sealing performance, and fluctuating ambient temperature, intermittent and frequent pressure surges are common, leading to accelerated wear of sealing components and fatigue damage to critical parts. By analyzing pressure surge event signals triggered by the online monitoring system and remote signaling data on motor start / stop status, a multi-dimensional time series analysis model is constructed to quantitatively assess the hourly, daily, and weekly pressure surge frequency distribution characteristics and the evolution trend of single pressure surge duration. Integrating static attributes such as equipment manufacturing parameters, mechanism type maps, operating environment characteristics, and service life, a multi-dimensional state assessment system based on machine learning is established to accurately identify hidden defects such as internal leakage in the hydraulic circuit and decreased oil pump efficiency. This enables proactive early warning of mechanism health status and optimization of maintenance strategies, effectively preventing the risk of circuit breaker lockout due to system pressure loss. When an alarm is triggered, on-site personnel must continuously monitor the mechanism's status information, focusing on any increasing trend in alarm information. When the number of pressure surges triggers an alarm, an immediate on-site inspection should be conducted to check for leaks. When the pressure test duration alarm is triggered, in addition to checking whether the mechanism is leaking, it is also necessary to check whether the motor is burnt out.

[0088] Infrared thermal diagnostic model: Based on GIS equipment 3D parametric modeling technology, a high-precision simulation model library of multi-physics field coupling for typical structures is constructed. It integrates dynamic coupling analysis methods of electromagnetic field, fluid field, and temperature field, and simultaneously fuses SF6 gas component charged detection data to achieve cross-scale mapping of insulation performance and thermodynamic behavior. Through a multi-source data correlation analysis framework, the nonlinear coupling relationship between the operating environment (temperature and humidity, solar radiation intensity), load current fluctuations, and equipment thermal response is quantified. A characteristic fingerprint library of the shell-contact heat conduction path is established, and an inversion model based on the shell's infrared temperature distribution is constructed to achieve non-invasive rapid diagnosis of contact status. Combining gas decomposition product chromatographic detection and X-ray digital imaging technology with machine learning and deep learning algorithms, the temperature status assessment of GIS disconnect switches and busbars is achieved.

[0089] The circuit breaker waveform diagnostic model analyzes the recorded waveform files and performs time-domain waveform analysis of the closing current based on a visual detection algorithm model. It determines the current during each opening and closing operation of the circuit breaker group, calculates the cumulative contact erosion, and combines various advanced detection technologies such as the circuit breaker's own electrical life parameters, dynamic resistance test data, X-rays, and radiated electromagnetic waves. This allows for accurate determination of whether there are abnormal switching phase angles, abnormal closing resistance operation, reignition, and remaining electrical life calculations during the opening and closing process. Based on this, a diagnostic report is generated, clearly and intuitively displaying the diagnostic results, thus forming the circuit breaker waveform diagnostic model.

[0090] Diagnostic model for disconnector switch opening and closing status: Data is collected through drone aerial photography from a fixed angle along a fixed flight path or by pre-installed cameras at fixed locations. The model requires fine-tuning for different models. An improved deep learning target detection algorithm based on YOLOx is used to detect targets on the disconnector switch equipment, allowing for a preliminary assessment of the disconnector switch's opening and closing status. Simultaneously, a large number of disconnector switch images are collected using deep learning-based image segmentation methods, and the status of the disconnector switches is manually labeled. A disconnector switch segmentation algorithm model is trained based on the Mask R-CNN algorithm. The segmented images are input into the disconnector switch segmentation algorithm model to calculate the angle between the disconnector switch arm edge line and the baseline. The results are then combined with the results from the disconnector switch geometric model.

[0091] Mechanical characteristic diagnostic model: The mechanical characteristic diagnostic model requires the installation of corresponding mechanical characteristic sensors. Based on the action stage, it performs feature analysis on the waveform, determines the equipment fault type, and establishes a database. Utilizing image recognition algorithms and self-supervised learning capabilities, it constructs a training sample set based on the current and stroke curve data of the opening and closing coils, annotates feature quantities such as extreme values ​​and slopes, and forms a mechanical characteristic diagnostic model after model training. It calculates and analyzes the collected mechanical characteristic waveform data, analyzes the fault causes based on the time stage, waveform characteristics, and other patterns, and assists maintenance personnel in troubleshooting potential hazards such as mechanism jamming and coil faults, thereby reducing the risk of circuit breaker failure.

[0092] Partial Discharge Diagnostic Model: Through multiple fault simulation tests of full-scale prototypes at different voltage levels, the partial discharge diagnostic model establishes a sample library of partial discharge signals, including typical discharge types and different interference signals. Based on the large-scale model's image recognition capabilities, PRPD image features are fused with PD signal features. A corresponding partial discharge spectrum with grayscale features is generated based on the number of discharges. Using the YOLO image recognition and analysis algorithm, and after multiple rounds of iterative optimization, it possesses a certain degree of robustness and accuracy. It can accurately identify discharge types such as floating discharge, tip discharge, and air gap discharge, while eliminating common interference signals such as radar signals, mobile phone signals, and fluorescent lamps. This forms the partial discharge diagnostic model, effectively improving the timeliness and accuracy of defect partial discharge signal identification, reducing false alarms and missed alarms, and assisting on-site maintenance personnel in defect assessment.

[0093] X-ray Detection and Analysis Model: This model utilizes deep learning image recognition algorithms and X-ray detection and analysis technology to achieve non-destructive testing and accurate analysis of the internal structures of GIS and switchgear. It employs convolutional neural networks to achieve dynamic resolution feature extraction from X-ray images. Deep feature extraction, dimensionality reduction, and classification are performed on X-ray images of arbitrary resolution. Dynamic resolution technology and attention mechanisms are introduced to achieve feature extraction under a unified visual encoder. Object detection algorithms such as Faster R-CNN and YOLO are used to accurately detect the location and category of defects in X-ray images. A typical X-ray defect anomaly sample library is constructed. Images of internal and external fault defects in GIS and switchgear from various manufacturers are collected to provide an anomaly sample library for model training. The model is trained using a large number of samples to improve its generalization ability. The model is continuously tested and optimized until a satisfactory recognition rate is achieved.

[0094] Typical Fault Cases and Solutions Mechanical failure: 1. Incomplete Opening / Closing: Faulty operating mechanisms, wear of transmission components, and poor lubrication can all lead to abnormal or stuck opening / closing actions, thus affecting the stable switching and continuous operation of the power system. This problem is addressed by collecting mechanical characteristic parameters such as opening / closing time, speed, and stroke of high-voltage switches under normal and abnormal opening / closing conditions. Simultaneously, data on the material properties, dimensional tolerances, wear values, and lubrication status of transmission components (connecting rods, crank arms, pins, etc.) are integrated, along with operational log information such as the number of switch operations, control voltage, and frequency of actions. A multi-dimensional feature dataset is constructed to train a mechanical fault diagnosis model, enabling accurate identification and root cause analysis of incomplete opening / closing faults.

[0095] 2. Operating Mechanism Failure: As the core actuator controlling the opening and closing of the switch, the operating mechanism exhibits diverse failure modes. Typical defects include core movement stagnation in electromagnetic operating mechanisms, overheating and burnout of the closing winding, and fatigue deformation and stress fracture of the energy storage spring in spring operating mechanisms. These faults directly lead to inaccurate opening and closing timing or operational failure, posing a potential risk to the stable operation of the power grid. By collecting key parameters of mainstream operating mechanisms such as electromagnetic and spring types under normal and fault conditions: the electrical dimension includes the closing coil current waveform, operating voltage fluctuation, and circuit impedance characteristics; the mechanical dimension includes spring compression displacement, core movement trajectory accuracy, and energy storage status feedback. Simultaneously integrating full lifecycle data such as equipment manufacturer, model specifications, years of operation, and historical maintenance records, a multi-source heterogeneous training dataset is constructed to drive a deep neural network model to achieve accurate identification and fault location of operating mechanism defects.

[0096] 3. Transmission Component Failures: Mechanical defects such as accumulated wear, plastic deformation, or structural fracture in core transmission components like connecting rods, crank arms, and pins can lead to inaccurate or jammed opening and closing actions, resulting in operational timing deviations and decreased mechanical energy transmission efficiency, significantly deteriorating the dynamic performance of high-voltage switchgear. By acquiring multi-dimensional state data of transmission components under different mechanical defect conditions—dynamically monitoring high-frequency vibration spectra, noise characteristic frequency band energy distribution, and nonlinear fluctuation characteristics of operating force curves during opening and closing operations—and simultaneously integrating full lifecycle parameters such as component material properties, machining accuracy indicators, cumulative operating time, and typical load conditions, a health status assessment model for the transmission system is constructed based on vibration signal time-frequency domain analysis, acoustic emission feature extraction, and multi-source data fusion technology. This enables quantitative detection and fault tracing of early-stage problems such as mechanical wear and latent cracks.

[0097] Electrical fault: 1. Contact Overheating: Under long-term high-current operation, if the dynamic contact resistance of the contacts of a high-voltage switch increases abnormally, it will trigger the accumulation of Joule heating, leading to oxidation of the contact surface material and deterioration of the crystal structure. This type of thermal damage not only significantly shortens the mechanical life of the contacts but may also cause contact ablation pits and contact surface morphology distortion, seriously threatening the continuity of power supply. A multi-condition dataset covering high-voltage switchgear in substations of multiple voltage levels is constructed, including operating parameters under various conditions such as normal operation, slight contact overheating, and severe overheating. The electrical dimension covers features such as the effective value of load current, transient voltage fluctuations, and power factor shift; the thermodynamic dimension integrates key indicators such as contact temperature rise gradient and dynamic contact resistance. Simultaneously, full lifecycle data such as equipment model specifications, cumulative service time, typical load spectrum, and maintenance history are integrated to form a multi-dimensional parameter vector. A state assessment model based on the gradient boosting algorithm is trained, and through feature contribution analysis and heat conduction path modeling, the thermal state of the contacts is graded for assessment and degradation trend prediction.

[0098] 2. Insulation Breakdown: Long-term electrothermal stress or environmental humidity penetration causes molecular chain breakage and dielectric constant degradation in insulating materials. When combined with transient overvoltage impacts, the insulation strength will rapidly decay to the breakdown threshold, triggering interphase arcing and regional power outages. A dataset covering the entire lifecycle of high-voltage switchgear, encompassing various insulating media (ceramics, epoxy resin, etc.), is constructed. This dataset collects insulation state parameters under different operating conditions, including normal operation and after overvoltage impacts, such as time-varying insulation resistance curves, dielectric loss spectrum characteristics, and partial discharge pulse distribution. It also synchronously correlates external variables such as environmental temperature and humidity fluctuations and pollutant concentration gradients, integrating maintenance data such as equipment insulation aging assessment reports and preventative test historical records. A deep learning model based on a spatiotemporal attention mechanism is trained. Through dielectric response feature decoupling and multi-physics coupling analysis, quantitative assessment of microscopic defects in insulating materials and prediction of breakdown risk probabilities across multiple time scales are achieved.

[0099] 3. Overheating of Current-Carrying Conductors: Defects such as loose connection structures, insufficient effective current-carrying area, or oxidation at the contact interface in current-carrying conductors can lead to a nonlinear increase in contact resistance in the electrical contact area, exacerbating the Joule heating effect. If this heat accumulation process is not suppressed in time, it will cause phase transformation degradation of the conductor material and carbonization of the insulating medium, significantly increasing the risk of electrical fires in substations. A multi-condition dataset covering current-carrying conductors such as busbars and cables is constructed: dynamic parameters such as the effective value of current and temperature field distribution are collected under steady-state and transient operating modes; design attributes such as conductor material conductivity, geometric cross-sectional area parameters, connection structure type, and mechanical preload parameters are integrated simultaneously, as well as environmental thermodynamic indicators such as convective heat transfer coefficient and radiative heat dissipation efficiency of the installation area. Based on inverse modeling of heat conduction and coupled analysis of electro-thermal-mechanical multiphysics fields, a hybrid diagnostic model integrating physical mechanisms and data-driven approaches is trained to locate the overheating heat source of the conductor, calculate the contact degradation rate, and predict the critical failure time.

[0100] Gas malfunction: 1. SF6 Gas Leakage: SF6 gas-insulated switches may experience electrical sealing interface failure due to factors such as creep relaxation of seals, molecular chain breakage, insufficient flange surface machining precision, or microscopic defects in the sealing groove, leading to SF6 gas leakage. This leakage will cause deterioration of the gas medium strength and attenuation of dynamic arc extinguishing capability, significantly increasing the risk of discharge breakdown and reignition. A gas status database covering multiple models of GIS equipment is constructed, collecting multi-dimensional parameters under normal operation and maintenance and different leakage levels: physical quantities such as gas density gradient, pressure decay rate, and temperature field distribution; simultaneously integrating key attributes such as leak source location coordinates, micro-leakage diffusion paths, sealing material weather resistance parameters, and cumulative service time, and correlating external variables such as installation environment dew point temperature and atmospheric pressure fluctuations. Based on acoustic-infrared joint detection and multi-physics field coupling analysis, a multi-sensor data fusion diagnostic model is trained to achieve early identification of micro-leakage, precise location of leak sources, and dynamic assessment of remaining sealing life, supporting preventative maintenance decision-making.

[0101] 2. Excessive Moisture in SF6 Gas: Moisture intrusion into the SF6 gas insulating medium due to leakage in the sealing system or defects in the gas replenishment process leads to the formation of hydrolytic corrosion products such as sulfur-containing fluorine oxides under the high-temperature catalysis of an electric arc. This causes electrochemical corrosion of internal metal components and the formation of tracking marks on the insulator surface, resulting in simultaneous deterioration of both the strength of the gas medium and the mechanical integrity of the equipment. A multi-condition SF6 gas state database is constructed, integrating insulation state parameters under different humidity conditions: time-varying curves of gas insulation resistance, partial discharge phase distribution, and chromatographic analysis data of decomposition products (such as SO2 and HF concentrations). This database is synchronously linked to maintenance logs, adsorbent replacement cycles, and time-series records of absolute environmental humidity. A hybrid diagnostic model based on multi-source data fusion is trained. Through cross-parameter correlation analysis of gas trace moisture content and decomposition product concentration, the model achieves source tracing of moisture intrusion, quantitative assessment of insulation degradation rate, and critical corrosion risk early warning, supporting gas humidity control and component replacement decisions.

[0102] Other faults: 1. Secondary Circuit Faults: Defects in the secondary circuit, such as abnormal contact impedance, insulation breakdown, or logic element failure, can lead to inaccurate electrical control timing and protection logic failure, resulting in switch malfunctions or failures, severely weakening the reliability of power system operation commands. This system collects electrical parameters of the secondary circuit under normal operation and different types of faults (short circuit, open circuit, poor contact, etc.), such as voltage, current, resistance, and the operating status and control logic of components like relays and contactors. It also integrates full lifecycle data, including secondary circuit wiring topology and process specifications, cable aging assessment reports, and historical malfunction event records. Based on circuit topology analysis and transient response modeling, a hybrid diagnostic model integrating physical feature extraction and a multi-level inference engine is trained to locate poor contact, identify hidden short circuits, and trace back control logic anomalies, guiding rapid fault isolation and redundant circuit reconfiguration.

[0103] 2. Faults Caused by Environmental Factors: Harsh environmental conditions such as high temperature, high humidity, and dust storms can negatively impact the normal operation of switchgear. High temperatures worsen heat dissipation, leading to overheating and damage to electrical components; dust accumulation on component surfaces can reduce insulation performance, causing short-circuit faults. We collect electrical performance data, insulation surface condition data (degree of contamination, moisture level, etc.), and equipment protection measures (such as the use of dust covers and moisture-proof heaters) of high-voltage switchgear under different environmental conditions (high temperature, high humidity, dust storms, salt spray, etc.). We also record trends in environmental parameters and equipment operating time. Using these samples, we train a fault prediction model based on environmental monitoring and equipment condition assessment, enabling us to proactively develop measures to address the impact of environmental factors and ensure the safe operation of the equipment.

[0104] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-dimensional data fault diagnosis system for high-voltage switchgear in substations, characterized in that, It includes a multi-dimensional data acquisition module, a data processing module, and a diagnostic decision-making module; The multidimensional data acquisition module is used to acquire multidimensional data information from high-voltage switchgear in substations. The data processing module is used to preprocess the collected multidimensional data information; The diagnostic decision module includes a fault diagnosis unit based on historical data, a fault diagnosis unit based on data mining, a fault diagnosis unit based on machine learning, and a fault diagnosis unit based on high-voltage switchgear. The historical data-based fault diagnosis unit is used to diagnose faults based on the historical data of the equipment. The data mining-based fault diagnosis unit is used to diagnose hidden fault data. The machine learning-based fault diagnosis unit is used for fault diagnosis of high-dimensional, multi-source, and small-sample data. The fault diagnosis unit based on high-voltage switchgear utilizes multiple models to diagnose faults in equipment data.

2. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The multidimensional data acquisition module includes a body information acquisition unit, an anomaly information acquisition unit, an online data monitoring unit, and a live-line detection data acquisition unit; The body information acquisition unit is used to collect basic data on the operation of power equipment; The abnormal information collection unit is used to collect any information, data or events that are inconsistent with the normal state or behavior during the operation of the equipment. The online data monitoring unit is used to collect and transmit data to the data processing center in real time; The live-line monitoring data acquisition unit is used to collect equipment status information obtained through live-line detection technology when the electrical equipment is in operation.

3. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The data processing module includes a data fusion unit, a feature extraction unit, a feature fusion unit, and a fault data association rule analysis unit; The data fusion unit is used to fuse multidimensional data; The feature extraction unit is used to extract feature data that characterizes the operating status of the equipment; The feature fusion unit is used to fuse feature data; The fault data association rule analysis unit is used to generate data association rules.

4. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The historical data fault diagnosis unit includes a data cleaning subunit and a model building subunit. The data cleaning subunit processes historical data using the KNN interpolation algorithm, and the model building subunit is used to build an XGBoost model and perform fault diagnosis using the XGBoost model.

5. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The data mining fault diagnosis unit includes a data preprocessing subunit and a model building subunit. The data preprocessing subunit processes the data using the isolated forest algorithm, and the model building subunit is used to build a K-Means model and perform fault diagnosis using the K-Means model.

6. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The machine learning fault diagnosis unit includes a data stream processing subunit and a model building subunit. The data stream processing subunit is used to process real-time data streams, and the model building subunit is used to build an XGBoost model and use the XGBoost model for fault diagnosis.

7. The multi-dimensional data fault diagnosis system for substation high-voltage switchgear according to claim 1, characterized in that: The fault diagnosis unit based on high-voltage switchgear includes a gas pressure logic diagnosis model, an oil pump pressure abnormality diagnosis model, an infrared thermal diagnosis model, a switching waveform diagnosis model, a disconnecting switch switching status diagnosis model, a mechanical characteristic diagnosis model, a local diagnostic model, and an X-ray detection and analysis model. The gas pressure logic diagnostic model is used for fault diagnosis of SF6 gas pressure; the oil pump pressurization anomaly diagnostic model is used for fault diagnosis of oil pressure in hydraulic energy storage mechanisms; the infrared thermal diagnostic model is used for fault diagnosis of equipment operating temperature; the circuit breaker opening and closing waveform diagnostic model is used for fault diagnosis of closing current time-domain waveform; the disconnector switch opening and closing status diagnostic model is used for detection of disconnector switch equipment targets; the mechanical characteristic diagnostic model is used for fault diagnosis of mechanical characteristic waveform data; the local diagnostic model is used for fault diagnosis of discharge signals; and the X-ray detection and analysis model is used for fault diagnosis of the internal structure of GIS and switchgear.

Images