A Multi-Source Data Fusion and AI Adaptive Weighting System for Multi-Dimensional Fault Diagnosis of High-Voltage Switches

Abstract

This invention discloses a multi-source data fusion and AI adaptive weighting system for multi-dimensional fault diagnosis of high-voltage switches, belonging to the field of intelligent operation and maintenance and fault diagnosis technology for power equipment. Addressing the problems of traditional diagnostic data being singular, model weights being fixed, and the difficulty in timely identification of latent faults, this invention constructs a multi-source fusion architecture based on temperature rise, vibration, partial discharge, opening and closing stroke, and current waveform. The system dynamically adjusts diagnostic parameters through an AI adaptive weighting model, combined with time-series degradation trend extrapolation, to achieve equipment health assessment, early fault warning, and tiered operation and maintenance, thereby improving the ability to identify latent faults in high-voltage switches and the level of predictive maintenance.

Description

Technical Field

[0001] This invention relates to the fields of substation high-voltage switch status monitoring, power artificial intelligence fault diagnosis, and predictive maintenance of power equipment, specifically a multi-dimensional fault diagnosis system for high-voltage switches based on multi-source data fusion and AI adaptive weighting. Background Technology

[0002] High-voltage circuit breakers, disconnect switches, and GIS switchgear in substations are core equipment for power grid on / off control. Deterioration of equipment insulation, contact erosion, mechanical jamming, partial discharge, and excessive temperature rise can easily lead to major power accidents such as short circuits, equipment explosions, and station-wide power outages.

[0003] Existing traditional high-voltage switch diagnostic technologies have several inherent shortcomings: 1. The monitoring method is limited to a single electrical parameter, the circuit current, which cannot capture early latent defects without instantaneous electrical changes, such as vibration, partial discharge, and slow temperature rise, resulting in a high probability of missed detection. 2. The diagnostic model uses fixed-weight logic and does not distinguish between the equipment's years of operation, voltage level, and on-site temperature, humidity, and pollution environment. The accuracy of the model is greatly reduced in the case of old equipment and high-humidity and pollution scenarios, resulting in frequent false alarms. 3. It only provides passive alarms after a fault exceeds the limit, lacks trend analysis of long-term operating sequence data of the equipment, cannot predict potential deterioration faults in advance, and can only perform emergency repairs after the fact, frequently causing unplanned power outages; 4. Relying on manual periodic infrared, partial discharge, and mechanical characteristic inspections is difficult to support the construction of unmanned substations due to the wide distribution of substations, the large number of equipment, the large blind spots in inspections, and the high cost of manual operation and maintenance.

[0004] Existing multi-parameter monitoring solutions simply overlay multiple types of data, failing to achieve an integrated architecture that integrates heterogeneous feature fusion, adaptive weight adjustment based on operating conditions, and time-series degradation prediction. This fails to systematically address industry pain points such as early defect identification, cross-device adaptation, and early fault prediction, leaving a significant technological gap. Summary of the Invention

[0005] Purpose of the invention Overcoming the shortcomings of existing high-voltage switch fault monitoring, such as single dimension, poor adaptability of diagnostic models, only post-event alarms, and high operation and maintenance costs, this paper provides an integrated diagnostic system that integrates multi-source heterogeneous data fusion, AI dynamic adaptive weight discrimination, and time-series degradation trend prediction. This system improves the accuracy of early hidden fault identification, is compatible with high-voltage switches of all voltage levels and service life, enables predictive condition maintenance, and supports the implementation of unmanned intelligent operation and maintenance in substations.

[0006] Technical solution This invention constructs a five-layer serial diagnostic architecture: a multi-dimensional heterogeneous sensor acquisition unit, a data preprocessing and fusion unit, an AI adaptive weight diagnostic unit, a time-series degradation trend prediction unit, and an operation and maintenance strategy output unit. The core includes three original technologies: a multi-dimensional heterogeneous sensor data joint feature extraction mechanism, an equipment operating condition adaptive weight AI diagnostic model, and a time-series data equipment degradation trend prediction algorithm. The three core modules work together to complete the entire process of automated diagnosis, including data acquisition, feature fusion, intelligent scoring, trend prediction, and hierarchical operation and maintenance output.

[0007] Core Innovation Description 1. Joint feature extraction from multi-dimensional heterogeneous sensor data It integrates five completely heterogeneous monitoring signals: temperature rise, mechanical vibration, partial discharge, opening and closing stroke, and current waveform. It unifies noise reduction and normalizes to eliminate dimensional differences, constructs a fault feature map for high-voltage switches, and establishes a correlation mapping between multiple features and typical defects, thus solving the problem that single parameter monitoring cannot identify hidden faults.

[0008] 2. Equipment Operating Condition Adaptive Weighted AI Diagnostic Model Abandoning the industry's common fixed-weight diagnostic logic, the AI ​​model introduces equipment operating years, voltage level, ambient temperature and humidity, and historical defect records as weight correction factors. The AI ​​model automatically adjusts the weights of the five monitoring dimensions in real time. For old and highly polluted equipment, the weights of insulation and mechanical dimensions are strengthened, which greatly reduces the misjudgment and missed judgment rates of equipment in different scenarios.

[0009] 3. Algorithm for predicting the degradation trend of time-series data equipment Long-term storage device health time-series data is used to fit a health degradation curve. The rate of degradation is quantified by the slope of the curve's decline. Potential problems such as contact wear, insulation aging, and mechanical jamming can be predicted 1 to 3 months in advance. This upgrades the traditional post-fault alarm to a pre-fault predictive warning, reducing the risk of sudden power outages.

[0010] Implementation process S1. Install multiple types of sensors to synchronously collect raw data on high-voltage switch temperature rise, vibration, partial discharge, opening and closing stroke, and current waveform; S2. The preprocessing unit filters and reduces noise from multi-source heterogeneous data, normalizes dimensions, and performs joint extraction of multi-dimensional fault features to generate standardized feature vectors. The S3 AI adaptive weight diagnosis unit retrieves equipment files, environmental and historical defect parameters to dynamically update the weights of each dimension and calculates the overall health score of the equipment. S4. The timing degradation simulation unit fits the decay curve based on long-term health time-series data to predict medium- and long-term failure risks. S5. The system automatically pushes hierarchical operation and maintenance suggestions for inspection, live-line retesting, and power outage maintenance based on the predicted warning level.

[0011] Beneficial effects 1. Joint diagnosis of five types of heterogeneous data comprehensively covers electrical, mechanical, and insulation defects, significantly improving the accuracy of early hidden fault identification; 2. The adaptive dynamic weight model is compatible with all high-voltage switchgear, including newly commissioned and old equipment, different voltage levels, and different environmental conditions, making it highly versatile. 3. Timing degradation simulation enables early fault prediction, shifting from reactive emergency repairs to predictive condition-based maintenance, significantly reducing the risk of unplanned power outages in the power grid; 4. Automated online monitoring replaces large-scale manual inspections, eliminates blind spots in inspections, reduces manpower maintenance costs, and adapts to the construction needs of unmanned smart substations. Attached Figure Description

[0012] Figure 1. Overall five-layer system architecture block diagram of the present invention Figure 2. Flowchart of joint feature extraction processing for multi-source heterogeneous data Figure 3. Flowchart of AI Adaptive Weight Dynamic Update Diagnostic Logic Figure 4. Flowchart for predicting the degradation trend of time series data and providing graded early warning.

Claims

1. A multi-source data fusion and AI adaptive weighting system for multi-dimensional fault diagnosis of high-voltage switches, characterized in that, It includes a multi-dimensional heterogeneous sensing acquisition unit, a data preprocessing and fusion unit, an AI adaptive weight diagnosis unit, a time-series degradation trend inference unit, and an operation and maintenance strategy output unit; The multi-dimensional heterogeneous sensing and acquisition unit is deployed on the body of various high-voltage switchgear in the substation to synchronously collect five types of raw monitoring data: equipment temperature rise, mechanical vibration, partial discharge, opening and closing stroke, and circuit current waveform. The data preprocessing and fusion unit is connected to the multi-dimensional heterogeneous sensing and acquisition unit to complete multi-source heterogeneous data filtering and noise reduction, dimension normalization, joint fault feature extraction, and generate a standardized fusion feature set. The AI ​​adaptive weight diagnosis unit has a built-in equipment condition adaptive weight calculation model. It receives a standardized fusion feature set and dynamically updates the weights of each monitoring dimension in combination with the equipment's operating years, rated voltage level, on-site environmental temperature and humidity, and historical defect records, and outputs a comprehensive health score of the equipment in real time. The time-series degradation trend inference unit accesses long-term continuous health time-series data, fits the equipment health decay curve, infers the degradation progress of insulation, contacts, and mechanical structure, and generates medium- and long-term potential fault prediction results. The operation and maintenance strategy output unit pushes corresponding maintenance, inspection, and power outage troubleshooting suggestions according to the health score and the time-series degradation prediction results. The system relies on three core logic layers: multi-source heterogeneous feature fusion, dynamic adaptive weight AI discrimination, and time-series degradation prediction, to achieve the identification of latent defects in high-voltage switches and early fault warning.

2. The system according to claim 1, characterized in that, The multi-dimensional heterogeneous sensing joint feature extraction process includes: S1. Independently extract dimensional features from the temperature rise time-domain curve, vibration spectrum signal, partial discharge pulse sequence, opening and closing stroke displacement curve, and current harmonic waveform; S2. Eliminate the dimensional differences of the five types of sensor data and uniformly map them to the 0-1 feature value range; S3. Establish a fault feature map for high-voltage switches and associate different feature combinations with typical fault labels such as contact wear, insulation aging, mechanical jamming, and poor contact. S4. Output a unified format fused feature vector and send it to the AI ​​diagnostic unit for processing.

3. The system according to claim 1, characterized in that, The dynamic weight update rule of the AI ​​adaptive weight diagnostic model is as follows: Retrieve basic equipment file parameters: years of operation, rated voltage, equipment model; Read real-time environmental parameters: ambient temperature, ambient humidity, and cabinet dirt level; Load historical defect records: historical partial discharge exceeding limits, temperature rise exceeding limits, and mechanical failure logs; Based on the above multi-dimensional operating condition parameters, the diagnostic weights of the five monitoring dimensions of temperature rise, vibration, partial discharge, stroke, and current are iteratively updated. For old equipment and highly polluted environments, the weights of insulation and mechanical characteristics are automatically increased, while for newly commissioned equipment, the monitoring weights of current and temperature rise are emphasized.

4. The system according to claim 1, characterized in that, The execution logic of the time-series data device degradation trend prediction algorithm is as follows: Store continuous time-series health scores on a daily / weekly / monthly basis; A piecewise linear fitting + exponential decay model was used to construct the equipment health degradation curve; By setting multiple degradation thresholds and predicting the time of failure based on the slope of the curve, early warnings of potential problems such as output contact wear, insulation aging, and mechanical jamming can be provided 1 to 3 months in advance. Synchronous output identifies the causes of accelerated degradation, distinguishing between three types of degradation causes: equipment aging, environmental interference, and excessive operation frequency.

5. The system according to claim 1, characterized in that, The comprehensive health assessment system for the equipment is divided into four scoring ranges: healthy, alert, abnormal, and serious malfunction. Each level is matched with a corresponding threshold judgment standard, and a unique comprehensive health score is obtained by combining adaptive weighted calculation.

6. The system according to claim 1, characterized in that, The defect classification and early warning strategy includes a three-level push mechanism: Level 1 Warning: Health level has slightly decreased, and the deterioration trend is slowing down. A routine suggestion to increase on-site inspections will be sent. Level 2 warning: The degradation curve is declining rapidly, indicating the presence of hidden defects. Short-term live-line testing and infrared retesting maintenance tasks are pushed out. Level 3 warning: When the fault is approaching the critical point, a power outage for maintenance and component replacement is pushed out within a specified period.

7. The system according to claim 1, characterized in that, The complete diagnostic implementation chain of the system is as follows: multi-dimensional sensors synchronously collect five types of operating data of the equipment → heterogeneous data noise reduction and normalization and joint extraction of fault features → AI adaptive weight model calculates the real-time health score of the equipment → time series model fits the health decay curve to infer the medium and long-term deterioration trend → automatically outputs graded operation and maintenance disposal suggestions according to the warning level.

Images