A high-voltage switch cabinet fault diagnosis system

Abstract

The application discloses a high-voltage switch cabinet fault diagnosis system, and relates to the technical field of power equipment monitoring and fault diagnosis.The system is a comprehensive diagnosis system relying on hardware collection and software analysis.The system comprises a fault information collection module, a data preprocessing unit, a multi-dimension feature extraction module, a fusion diagnosis engine and a diagnosis result output module.The application has the following innovative points: for the first time, five types of sensors, i.e.temperature sensors, partial discharge sensors, arc light sensors, electrical parameter sensors and vibration sensors, are integrated;the arc light sensors adopt a triangular redundant monitoring network;the vibration sensors compensate for the blank of mechanical fault monitoring;the system covers all types of faults of a switch cabinet, such as thermal faults, insulation faults, arc faults and mechanical faults, and greatly reduces the false negative rate.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring and fault diagnosis technology, specifically a fault diagnosis system for high-voltage switchgear. Background Technology

[0002] High-voltage switchgear is the core equipment for power distribution, control and protection in power systems. Its operational reliability is directly related to the safety and stability of the power grid. With the advancement of smart grid construction, higher requirements are placed on the condition-based maintenance and fault prediction of high-voltage switchgear. The internal structure of high-voltage switchgear is complex, involving multiple functional modules such as insulation, conductivity and mechanics. Faults are diverse, including insulation deterioration, poor contact, mechanical jamming, and arc discharge. These faults are characterized by strong concealment, high suddenness and great harm.

[0003] Existing switchgear fault diagnosis methods have many shortcomings: First, some methods rely solely on monitoring a single physical quantity, resulting in a high rate of false negatives and failing to fully reflect the health status of the equipment. Second, existing multi-sensor fusion methods only integrate some sensors such as temperature, partial discharge, and electrical parameters, and do not integrate arc light and vibration sensors. Furthermore, the monitoring of arc light does not form a redundant network, resulting in insufficient ability to identify mechanical faults. Third, the multi-source data fusion layer lacks an adaptive weight allocation mechanism for signals at different time scales, resulting in poor fusion performance and insufficient diagnostic response sensitivity when sudden faults and gradual faults coexist or transform. Fourth, most diagnostic models are static models, which cannot adapt to equipment aging and sensor drift throughout the entire life cycle of the switchgear, and the accuracy of the models decreases with operating time. Fifth, feature extraction only focuses on basic statistical features and fails to explore advanced features in the frequency domain and time-frequency domain, resulting in weak ability to identify early and subtle faults. Sixth, the early warning mechanism is based on a single threshold, lacks a risk-based tiered early warning system, and does not provide targeted operation and maintenance suggestions in conjunction with the failure evolution trend.

[0004] In summary, there is an urgent need for a high-voltage switchgear fault diagnosis system and method that can fully integrate multiple types of sensors, extract multi-dimensional high-level features, achieve adaptive feature fusion, support full lifecycle model iteration, and accurately diagnose all types of faults, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] To address the problems of single-dimensional fault diagnosis, poor multi-source data fusion, lack of model self-adaptation, low accuracy of early fault identification, and incomplete fault type coverage in existing high-voltage switchgear fault diagnosis technologies, this invention provides a high-voltage switchgear fault diagnosis system and method. Through multi-sensor integration, multi-algorithm joint preprocessing, multi-dimensional advanced feature extraction, multi-scale spatiotemporal attention fusion, and dynamic online model updating, it achieves accurate, real-time, and full lifecycle fault diagnosis of high-voltage switchgear.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A fault diagnosis system for high-voltage switchgear is based on a comprehensive diagnostic system that combines hardware acquisition and software analysis. The system includes a fault information acquisition module, a data preprocessing unit, a multi-dimensional feature extraction module, a fusion diagnostic engine, and a diagnostic result output module.

[0007] The fault information acquisition module integrates five types of sensors: temperature, partial discharge, arc light, electrical parameters, and vibration, to achieve comprehensive perception of the switchgear's operating status. Among them, the temperature sensor adopts the surface acoustic wave passive wireless type to achieve accurate temperature measurement of key heat-generating parts. Arc light sensors form a triangular redundant monitoring network to improve the reliability of arc light signal capture; vibration sensors are deployed on the circuit breaker operating mechanism to compensate for the shortcomings of existing technologies in monitoring mechanical faults.

[0008] The data preprocessing unit employs a multi-algorithm joint denoising and time axis adaptive alignment mechanism. It designs dedicated denoising algorithms for the noise characteristics of different types of signals, while simultaneously unifying the time axis of heterogeneous data, eliminating data acquisition errors, and laying the foundation for feature extraction.

[0009] The multi-dimensional feature extraction module adds advanced features in the frequency domain and time-frequency domain on the basis of basic statistical features. It constructs a unique feature sub-vector for each type of sensor signal, forming a rich multi-dimensional fusion feature vector set, which improves the ability to identify early and subtle faults.

[0010] The fusion diagnostic engine is the core of this invention. It integrates a multi-scale spatiotemporal attention network, an adaptive weight allocation mechanism, and a dynamic drift compensation and online update module. The multi-scale spatiotemporal attention network enables the effective extraction of features at different time scales. The adaptive weight allocation mechanism dynamically adjusts feature weights according to fault modes to solve the problem of heterogeneous data fusion. The dynamic drift compensation and online update module enables full lifecycle iteration of the model to avoid the decline in model accuracy caused by equipment aging.

[0011] The diagnostic results output module not only outputs the fault type, location, and severity, but also adds fault evolution trends and hierarchical operation and maintenance suggestions, providing operation and maintenance personnel with accurate and comprehensive decision-making basis. At the same time, it triggers a hierarchical early warning / alarm mechanism based on risk scoring to achieve timely fault response.

[0012] This invention also provides a fault diagnosis method based on the above-mentioned diagnostic model. Through six steps, namely fault information collection, multi-algorithm preprocessing, multi-dimensional feature extraction, multi-scale spatiotemporal attention fusion diagnosis, dynamic drift compensation, and diagnostic result output, it realizes the whole process diagnosis of high-voltage switchgear faults. The steps are closely connected, the operation is convenient, the diagnostic accuracy is high, and the real-time performance is strong.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Multi-sensor integrated + redundant monitoring: For the first time, five types of sensors, namely temperature, partial discharge, arc light, electrical parameters and vibration, are integrated into one unit. The arc light sensor adopts a triangular redundant monitoring network, and the vibration sensor fills the gap in mechanical fault monitoring. It covers all types of faults such as thermal faults, insulation faults, arc faults and mechanical faults in switchgear, which greatly reduces the false alarm rate. 2. Multi-algorithm joint preprocessing + adaptive time axis alignment: Dedicated denoising algorithms are designed for the noise characteristics of different signals. At the same time, a unified time axis database is built to achieve accurate alignment of high-frequency and low-frequency data, effectively eliminating environmental noise and data acquisition errors, and improving data quality. 3. Multi-dimensional advanced feature extraction: Based on basic statistical features, advanced features in the frequency domain and time-frequency domain, such as Mel frequency cepstral coefficients and vibration waveform envelope, are added to construct a multi-dimensional fused feature vector set, which can effectively identify early and subtle faults and improve the sensitivity of fault diagnosis. 4. Multi-scale spatiotemporal attention fusion + dynamic weight allocation: A multi-scale spatiotemporal attention network is used to extract fast and slow features at different time scales. The dynamic weight allocation based on the attention mechanism can adaptively adjust the feature weights according to the fault mode, which solves the problem of the difficulty in effectively fusing heterogeneous data and significantly improves the recognition accuracy of complex faults. 5. Dynamic model drift compensation + online update: A dynamic drift compensation and online update module is designed. The model is iterated online through incremental learning of high-value difficult examples, which enables the model to have the adaptive capability throughout its entire life cycle and solves the problem of model accuracy decline caused by equipment aging and sensor drift. 6. Risk-based tiered early warning + full-dimensional diagnostic report: Outputs a full-dimensional diagnostic report including fault type, location, severity, and evolution trend, and implements tiered early warning based on risk scores. It also provides targeted operation and maintenance suggestions, realizing a closed loop from fault diagnosis to operation and maintenance decision-making, and improving the operation and maintenance efficiency of the power system. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall architecture of a high-voltage switchgear fault diagnosis system according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of the multi-dimensional feature extraction module in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the Multi-Scale Spatiotemporal Attention Network (MSTAN) in Embodiment 1 of the present invention; Figure 4 This is a flowchart illustrating the dynamic drift compensation and online update module in Embodiment 1 of the present invention. Figure 5 This is a flowchart of the steps of the high-voltage switchgear fault diagnosis method described in Embodiment 2 of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments. This embodiment mainly focuses on the comprehensive status monitoring of 10kV and 35kV high-voltage switchgear.

[0016] Example 1: A fault diagnosis system for high-voltage switchgear The high-voltage switchgear fault diagnosis system of this embodiment relies on a comprehensive diagnostic system including a fault information acquisition module, a data preprocessing unit, a multi-dimensional feature extraction module, a fusion diagnostic engine, and a diagnostic result output module. The specific configurations of each module are as follows: 1. Fault Information Acquisition Module (1) Temperature sensor: Surface acoustic wave (SAW) passive wireless temperature sensor is selected. One sensor is installed on each phase of the three key heat-generating parts of the switch cabinet: the plum blossom contact, the busbar lap surface, and the cable terminal head, for a total of 9 measuring points. The working frequency is 433MHz, the temperature range is -25℃ to 150℃, and the measurement accuracy is ±0.5℃. The temperature data is transmitted to the aggregation unit in the cabinet wirelessly. (2) Partial discharge sensor: A combination of ultra-high frequency (UHF) sensor, ultrasonic sensor and ground wave (TEV) sensor is adopted. The UHF sensor is installed near the cable compartment through the wall bushing at the back of the switch cabinet instrument room. The effective detection frequency range is 300MHz to 1500MHz to capture the discharge electromagnetic waves inside the insulation. The ultrasonic sensor is built into the insulation component. The ground wave sensor is magnetically attached to the outside of the metal cabinet wall of the switch cabinet to detect the transient voltage signal to ground generated by the discharge. (3) Arc sensor: An integrated ultraviolet / visible light sensor is selected and installed on the top inner wall of the busbar compartment, circuit breaker truck compartment and cable compartment respectively to form a triangular redundant monitoring network. It is sensitive to the specific wavelength ultraviolet light generated by the electric arc, with a response time of less than 1ms and a detection spectrum range of 185nm-750nm. (4) Electrical parameter acquisition unit: It is connected to the secondary side of the open-type Hall current sensor and voltage transformer to collect three-phase current, three-phase voltage, active power, reactive power and power factor. The sampling rate is set to 256 points per cycle to meet the requirements of harmonic analysis. (5) Vibration sensor: An accelerometer is selected and deployed at the circuit breaker operating mechanism to collect the vibration waveform during the opening and closing process. The frequency range is 0-10kHz and the sampling frequency is 20kHz.

[0017] 2. Data Preprocessing Unit (1) Joint denoising using multiple algorithms: Outliers exceeding physical limits were removed from the temperature data, and a moving average filter with a length of 5 was used for smoothing. The partial discharge signal is denoised by wavelet packet transform with 5-level decomposition of Daubechies4 wavelet basis, which separates white noise and periodic narrowband interference. Morphological filtering is applied to the vibration signal to preserve the mechanical vibration characteristics; median filtering is applied to the electrical parameters to remove periodic interference and spike noise. A pulse recognition algorithm is used to eliminate natural light interference in the arc light signal; (2) Adaptive alignment of time axis: Construct a unified time axis database, statistically aggregate high-frequency data such as partial discharge, arc light, and vibration within a 1-second time window (take the maximum value and average value), so that it is aligned with low-frequency data such as temperature and electrical parameters in the time dimension to form a unified time slice sample; (3) Normalization: Min-Max normalization is applied to the aligned heterogeneous data to scale the data to the [0,1] interval, eliminating dimensional differences and data acquisition errors.

[0018] 3. Multi-dimensional feature extraction module Features were extracted from the preprocessed standardized data, five types of feature sub-vectors were constructed, and then fused to form a multi-dimensional fused feature vector set: ; (1) Temperature feature vector ( ): Temperature amplitude, temperature rise rate, three-phase temperature difference, historical trend characteristics, absolute temperature rise, and three-phase imbalance; (2) Partial discharge characteristic sub-vectors ( ): Discharge amplitude, pulse repetition rate, discharge phase distribution (PRPD) spectral characteristics, skewness, kurtosis statistical characteristics, Mel frequency cepstral coefficients (MFCC) frequency domain characteristics, maximum discharge quantity, average discharge quantity, and number of discharges; (3) Arc light feature sub-vectors ( ): Light intensity amplitude, light pulse rise time, spectral energy characteristics, and light intensity threshold exceeding the limit; (4) Electrical parameter feature sub-vectors ( ): Current harmonic distortion rate, voltage sag amplitude, zero-sequence current component characteristics, current overload factor, total harmonic distortion rate (THDu); (5) Vibration characteristic sub-vectors ( ): Vibration waveform envelope, number of peaks, action time, vibration time-domain kurtosis, margin factor, and peak frequency of the main frequency during the opening and closing process.

[0019] 4. Fusion Diagnostic Engine The fusion diagnostic engine includes a multi-scale spatiotemporal attention network (MSTAN), an adaptive weight allocation mechanism, and a dynamic drift compensation and online update module. (1) Multi-scale Spatiotemporal Attention Network (MSTAN): Input layer: Multi-scale convolutional layer; the slow channel uses a one-dimensional convolutional layer with kernel size=64 to process temperature and electrical RMS value sequences and extract trend features. The fast channel uses a deep convolutional network with a kernel size of 3 to process the original waveforms of partial discharge, arcing, and vibration, and extract transient features; Attention Layer: Channel Attention Mechanism (CAM) layer, which learns the importance weight vector of each feature channel through a fully connected neural network. The formula is expressed as: ,in, The feature tensor after attention weighting, and Same dimensions The input is the original feature tensor. It is the Sigmoid activation function. This is the global average pooling function. It is a global max pooling function that automatically suppresses background noise channels and amplifies fault feature channels. Feature fusion layer and LSTM layer: The weighted slow features and fast features are concatenated and input into the Long Short-Term Memory (LSTM) network to capture the temporal evolution of the fault, such as the temporal chain of "weak partial discharge first, followed by a slow temperature rise"; Output layer: Softmax classifier, which outputs the probability distribution of fault categories and the confidence value of the diagnostic results. Fault types include poor contact, insulation aging, mechanical jamming, floating potential discharge, and arc short circuit. (2) Adaptive weight allocation mechanism: The neural network layer based on the attention mechanism calculates the weights based on the signal-to-noise ratio of each feature vector, the correlation with the fault category, and the real-time reliability coefficient of the sensor. The real-time reliability coefficient of the sensor is calculated by fusing three indicators: consistency, stability, and offset detection. This mechanism can dynamically adjust the weights according to the fault mode. For example, it increases the weight of partial discharge characteristics when insulation is aging, increases the weight of temperature characteristics when there is poor contact, increases the weight of vibration characteristics when there is mechanical jamming, and sets the arcing characteristics as the highest priority and gives them "veto power" when there is arcing short circuit. (3) Dynamic drift compensation and online update module: Confidence monitoring unit: Real-time monitoring of the confidence level of diagnostic results, with a set confidence threshold of 0.6; Sample screening unit: Marks samples with a confidence level that is in the critical range of 0.4-0.6 for a continuous preset number of diagnostic cycles as high-value difficult case samples, as well as misjudged samples and missed samples that have been acquired and verified by external terminals. Incremental learning unit: Maintains a dynamic experience pool, adds high-value difficult examples to the experience pool and removes the oldest examples, uses transfer learning techniques to fix the parameters of the bottom convolutional layers of the model, and only fine-tunes the training of the high-level fully connected layers and attention layers. Parameter update unit: After training is completed, the updated weight parameters are sent to the edge diagnostic terminal through OTA (Over-the-Air) technology to realize online iteration of the model.

[0020] 5. Diagnostic Result Output Module Based on the inference results of the fusion diagnostic engine, a comprehensive fault diagnosis report is output, including: (1) Operating status: Normal / Warning / Fault; (2) Fault types: poor contact, insulation aging, mechanical jamming, floating potential discharge, arc short circuit, etc., and give the fault probability; (3) Specific location of the fault: such as A-phase busbar, B-phase circuit breaker contacts, C-phase cable termination, etc.; (4) Severity level of the fault: general, severe, critical; (5) Fault evolution trend: Based on the time series analysis of the LSTM layer, the fault development speed is predicted; (6) Operation and maintenance suggestions: Provide specific suggestions for different severity levels. The general level is infrared retesting / vibration spectrum analysis, the severe level is generating and issuing power outage maintenance control work orders, and the critical level is immediately sending linkage trip control signals. Simultaneously, a tiered early warning / alarm mechanism based on risk scoring is triggered. The risk score is calculated by weighting fused feature values, feature change rates, and the average reliability coefficient of the sensors, as shown in the formula: ,in, To fuse feature values, For characteristic rate of change, The real-time reliability coefficient of the sensor; No warning is issued when the risk score is <0.2; a warning is issued (SMS / email) when the risk score is 0.2 ≤ risk score <0.5; a warning is issued (monitoring platform pop-up) when the risk score is 0.5 ≤ risk score <0.8; and an emergency warning is issued (local audible and visual alarm + remote emergency notification + tripping) when the risk score is ≥0.8.

[0021] Example 2: Fault Diagnosis Method for High-Voltage Switchgear The high-voltage switchgear fault diagnosis method in this embodiment is based on the diagnosis model of Embodiment 1 and includes the following steps: S1. Fault Information Acquisition: Through five types of sensors (temperature, partial discharge, arc light, electrical parameters, and vibration) in the fault information acquisition module, heterogeneous data on the operating status of key parts such as switchgear contacts, busbar lap surfaces, and circuit breaker operating mechanisms are collected simultaneously. The data collected by the sensors is transmitted to the cabinet aggregation unit via wireless (433MHz) / wired (Modbus / TCP / IP) methods, and then uploaded to the data preprocessing unit. The timestamp accuracy is ≤1ms. S2. Data Preprocessing: The data preprocessing unit performs multi-algorithm joint denoising, time axis adaptive alignment, normalization and dimension elimination on the collected raw data. First, outliers and interference signals of various types of data are removed. Dedicated denoising algorithms are used for different signal characteristics. Then, high-frequency data is statistically aggregated and aligned with the time axis of low-frequency data. Finally, the dimension difference is eliminated by Min-Max normalization to obtain standardized heterogeneous data. S3. Multi-dimensional feature extraction: The multi-dimensional feature extraction module extracts basic statistical features and high-level features in the frequency domain and time-frequency domain from the standardized heterogeneous data, and constructs five types of feature sub-vectors: temperature, partial discharge, arc light, electrical parameters, and vibration. The five types of sub-vectors are fused to form a multi-dimensional feature vector set, which is used as the input of the fusion diagnostic engine. S4. Multi-scale Spatiotemporal Attention Fusion Diagnosis: The multi-dimensional feature vector set is input into the Multi-scale Spatiotemporal Attention Network (MSTAN). First, the slow trend features and fast transient features are extracted by the multi-scale convolutional layer of the input layer. Then, the channel attention mechanism of the attention layer dynamically allocates the weights of each feature channel to suppress noise and amplify fault features. Subsequently, the feature fusion layer concatenates the weighted fast and slow features and inputs them into the LSTM layer to capture the temporal evolution of the fault. Finally, the softmax classifier of the output layer outputs the probability distribution of the fault type and the diagnostic confidence. S5. Dynamic Drift Compensation: The dynamic drift compensation and online update module monitors the diagnostic confidence of step S4 in real time. If the confidence is in the critical zone of 0.4-0.6 for a continuous preset number of diagnostic cycles, or if manual inspection confirms that the model has misjudged / missed, the original data in that time period is marked as high-value difficult case samples, added to the dynamic experience pool and the oldest sample is removed. The high-level parameters of the model are fine-tuned using transfer learning technology, and then the updated parameters are sent to the edge diagnostic terminal through OTA technology to realize the online iteration of the model. S6. Diagnostic Result Output: Based on the integrated diagnostic results from step S4, the diagnostic result output module outputs a comprehensive diagnostic report that includes fault type, specific location, severity level, evolution trend, and maintenance recommendations. At the same time, it calculates a risk score according to the risk scoring formula and triggers the corresponding graded early warning / alarm mechanism to achieve a closed loop between fault diagnosis and maintenance response.

[0022] Fault Diagnosis Cases Surface contamination of the B-phase busbar insulators in a 10kV switchgear caused surface discharge, which eventually developed into an arc short circuit. The diagnostic process of the model in this invention is as follows: 1. Initial stage (general level): The UHF sensor detects intermittent discharge pulses with small amplitudes. The Mel frequency cepstral coefficient detects the frequency domain characteristics of the discharge in the tiny air gap inside the insulation. Temperature, vibration, and electrical parameter data are normal. The model's fast-channel CNN extracts partial discharge pulse features, the attention layer assigns high weights to partial discharge features, and the LSTM layer analyzes the time series to find an increase in pulse density; Output diagnostic report: Phase B insulation abnormality, suspected surface discharge, fault probability 80%, severity moderate, risk score 0.35, warning issued; Maintenance recommendation: Infrared thermometry verification + partial discharge spectrum analysis.

[0023] 2. Intermediate (Severe): Discharge intensifies, TEV signal strengthens, dielectric loss heat causes the temperature of phase B busbar to slowly rise above that of phases A / C (temperature difference reaches 3℃), and the three-phase imbalance exceeds the normal range; The slow-channel CNN of the model captures the temperature difference trend, the attention layer highlights both partial discharge features and temperature difference features, and the LSTM layer identifies the temporal correlation pattern of "partial discharge causing overheating". Output diagnostic report: Phase B insulator has severe surface discharge accompanied by overheating, with a fault probability of 90%, a severe severity level, a risk score of 0.65, and a warning is issued; Maintenance recommendation: Schedule a power outage for maintenance within 24 hours.

[0024] 3. Final stage (critical level): An arc short circuit occurs, the arc sensor outputs a high level instantaneously, the light intensity amplitude far exceeds the threshold, and the current transformer detects the short circuit current; The model's arc feature changes instantaneously. The attention layer sets the arc feature as the highest priority and gives it "veto power," bypassing other slow features to achieve millisecond-level judgment. Diagnostic report output: Arc short circuit in phase B inside the cabinet, fault probability 99%, severity critical, risk score 0.92, emergency warning issued and tripping triggered, maintenance recommendation: immediately cut off power and replace insulators.

[0025] In this case, the model of the present invention identifies insulation abnormalities in the early stage of a fault, enabling early warning; accurately judges the fault development trend in the middle stage; and achieves millisecond-level fault diagnosis and tripping in the final stage, effectively avoiding equipment damage and power grid accidents.

Claims

1. A fault diagnosis system for high-voltage switchgear, characterized in that, The integrated diagnostic system based on hardware acquisition and software analysis includes a fault information acquisition module, a data preprocessing unit, a multi-dimensional feature extraction module, a fusion diagnostic engine, and a diagnostic result output module. The fault information acquisition module integrates a temperature sensor, a partial discharge sensor, an arc sensor, an electrical parameter acquisition unit, and a vibration sensor. The temperature sensor is a surface acoustic wave passive wireless temperature sensor, with one installed in each phase at three key heat-generating locations: the switchgear's perforated contact, the busbar lap joint, and the cable terminal. The arc sensor is an integrated sensor that combines ultraviolet and visible light monitoring functions, forming a triangular redundant monitoring network in the busbar compartment, circuit breaker compartment, and cable compartment. The vibration sensor is deployed at the circuit breaker operating mechanism to collect the opening and closing vibration waveforms. The data preprocessing unit employs a multi-algorithm joint denoising and time axis adaptive alignment mechanism. For partial discharge signals, wavelet packet transform with soft thresholding is used for denoising with Daubechies4 wavelet basis 5-level decomposition. For temperature data, a moving average filter with a length of 5 is used. For vibration signals, morphological filtering is used. A unified time axis database is constructed, and high-frequency data is statistically aggregated within a 1-second time window and aligned with temperature data. At the same time, heterogeneous data is normalized and dimensionless. The multi-dimensional feature extraction module constructs exclusive feature sub-vectors for temperature, partial discharge, arc light, electrical parameters, and vibration signals respectively. It also adds Mel frequency cepstral coefficient frequency domain features to the partial discharge signal and extracts the waveform envelope, number of peaks, and action time features of the vibration signal to form a multi-dimensional fused feature vector set. The fusion diagnostic engine includes a fusion model based on a multi-scale spatiotemporal attention network (MSTAN), an adaptive weight allocation mechanism, and a dynamic drift compensation and online update module. The adaptive weight allocation mechanism is a neural network layer based on the attention mechanism, which can dynamically adjust the weights of each feature according to the fault mode. It increases the weight of partial discharge features when insulation is aging, increases the weight of temperature features when there is poor contact, and increases the weight of vibration features when there is mechanical jamming. The diagnostic result output module outputs a diagnostic report that includes the fault type, specific location of the fault, severity level of the fault, and fault evolution trend. The fault types include poor contact, insulation aging, mechanical jamming, floating potential discharge, and arc short circuit.

2. The high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The features extracted by the multi-dimensional feature extraction module are specifically: Based on the temperature data, extract temperature amplitude, temperature rise rate, three-phase temperature difference, and historical trend characteristics; For partial discharge signals, the following features are extracted: discharge amplitude, pulse repetition rate, discharge phase distribution spectrum characteristics, skewness, steepness statistical characteristics, and Mel frequency cepstral coefficients (MFCC) frequency domain characteristics. For arc light signals, extract light intensity amplitude, light pulse rise time, spectral energy characteristics, and light intensity threshold exceeding indicators; Based on the electrical parameters, the characteristics of current harmonic distortion rate, voltage sag amplitude, and zero-sequence current component are extracted; For vibration signals, the envelope of the vibration waveform, the number of peaks, the action time, and the vibration time-domain kurtosis and margin factor characteristics are extracted during the opening and closing process.

3. The high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The Multi-Scale Spatiotemporal Attention Network (MSTAN) comprises an input layer, an attention layer, a feature fusion layer, an LSTM layer, and an output layer connected in sequence. The input layer is a multi-scale convolutional layer. The slow channel uses a one-dimensional convolutional layer with a kernel size of 64 to process the temperature and electrical effective value sequences and extract trend features. The fast channel uses a deep convolutional network with a kernel size of 3 to process the original waveforms of partial discharge, arcing, and vibration and extract transient features. The attention layer is a channel attention mechanism (CAM) layer, which learns the importance weight vector of each feature channel through a fully connected neural network, automatically suppressing background noise channels and amplifying fault feature channels; The feature fusion layer concatenates the weighted slow and fast features and inputs them into the LSTM layer to capture the temporal evolution of the fault. The output layer uses a Softmax classifier to output the probability distribution of fault categories, and the output results include confidence values.

4. The high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The dynamic drift compensation and online update module includes a confidence monitoring unit, a sample screening unit, an incremental learning unit, and a parameter update unit; The confidence monitoring unit monitors the confidence level of the diagnostic results in real time, and sets the confidence level threshold to 0.

6. The sample screening unit marks samples with a confidence level that are continuously within the critical range of 0.4-0.6 for a preset number of diagnostic cycles as high-value difficult case samples, as well as misjudged samples and missed samples that are verified and confirmed by external terminals. The incremental learning unit maintains a dynamic experience pool, adds high-value difficult sample to the experience pool and removes the oldest sample, and uses transfer learning technology to fix the parameters of the bottom convolutional layer of the model, while only fine-tuning the training of the high-level fully connected layer and attention layer. The parameter update unit uses OTA technology to send the trained weight parameters to the edge diagnostic terminal, enabling online iterative updates of the model.

5. A high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The partial discharge sensor is a combination of an ultra-high frequency sensor, an ultrasonic sensor, and a ground wave sensor. The ultra-high frequency sensor is installed near the cable compartment through-wall bushing at the back of the switchgear instrument room, with a detection frequency range of 300MHz to 1500MHz. The ultrasonic sensor is built into an insulating component, and the ground wave sensor is magnetically attached to the outside of the switchgear metal cabinet wall to detect the transient voltage signal to ground generated by the discharge.

6. The high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The electrical parameter acquisition unit is connected to the secondary side of the voltage transformer via an open-type Hall current sensor to collect three-phase current, three-phase voltage, active power, reactive power and power factor. The sampling rate is set to 256 points per cycle to meet the requirements of harmonic analysis.

7. The high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The adaptive weight allocation mechanism calculates weights based on the signal-to-noise ratio of each feature vector, its correlation with the fault category, and the sensor's real-time reliability coefficient. The sensor's real-time reliability coefficient is obtained by fusing three indicators—consistency, stability, and offset detection—through a preset weighted evaluation model.

8. A high-voltage switchgear fault diagnosis system according to claim 1, characterized in that, The fault severity level of the diagnostic result output module is divided into three levels: general, severe, and critical. Corresponding maintenance suggestions are output for different levels. The general level is infrared retesting / spectrum analysis, the severe level is generating and issuing a power outage maintenance control work order, and the critical level is immediately sending a linkage trip control signal.

9. A fault diagnosis method based on a high-voltage switchgear fault diagnosis system according to any one of claims 1-8, characterized in that, Includes the following steps: S1. Fault Information Acquisition: Through temperature, partial discharge, arc light, electrical parameters, and vibration sensors of the fault information acquisition module, heterogeneous data of the switchgear operating status are collected synchronously and transmitted to the data preprocessing unit via wireless / wired means. S2. Data preprocessing: The collected raw data is subjected to multi-algorithm joint denoising, time axis adaptive alignment, normalization and dimension elimination to obtain standardized heterogeneous data. S3. Multi-dimensional feature extraction: Extract specific features from the standardized heterogeneous data, construct feature sub-vectors for temperature, partial discharge, arc light, electrical parameters, and vibration, and fuse them to form a multi-dimensional feature vector set; S4. Multi-scale spatiotemporal attention fusion diagnosis: Input the multi-dimensional feature vector set into the multi-scale spatiotemporal attention network, extract fast and slow features through multi-scale convolutional layers, dynamically allocate feature weights through attention layers, capture fault temporal patterns through LSTM layers, and output fault type probability distribution and confidence level through Softmax classifier. S5. Dynamic Drift Compensation: The dynamic drift compensation and online update module monitors the diagnostic confidence, screens high-value difficult case samples, incrementally learns and fine-tunes the model parameters and updates them online. S6. Diagnostic Result Output: Based on the fusion diagnostic results, output a diagnostic report including fault type, specific location, severity level, evolution trend and operation and maintenance suggestions, and trigger the corresponding early warning / alarm mechanism.

10. The fault diagnosis method for high-voltage switchgear according to claim 9, characterized in that, The early warning / alarm mechanism described in step S6 is a graded mechanism based on risk score. The risk score is calculated by weighting the fused feature value, feature change rate and sensor average reliability coefficient. No warning is issued when the risk score is <0.2, a warning is issued when the risk score is 0.2≤risk score<0.5, a warning is issued when the risk score is 0.5≤risk score<0.8, and an emergency warning is issued and the circuit breaker is tripped when the risk score is ≥0.8.

Images