Online detection method and system for switchgear based on multi-sensor fusion

Abstract

The application discloses a complete switchgear online detection method and system based on multi-sensor fusion, relates to the technical field of equipment state detection, and comprises the following steps: collecting temperature distribution, local discharge signals, mechanical vibration waveforms and operating current data of the switchgear, respectively extracting key fault features based on preprocessed multi-source data, generating a high-dimensional feature matrix, constructing a graph convolution-long short-term memory hybrid network model as a fault diagnosis model, and obtaining a defect detection result according to space-time features; constructing a device degradation index based on defect features to calculate a device health index, probabilistically predicting the remaining service life in combination with a Wiener degradation model, and triggering a warning signal when a defect or a life is detected to be lower than a threshold value. The application solves the problems of great limitation of single sensor data and insufficient fault diagnosis precision in operation state monitoring of the complete switchgear, and further realizes fault evaluation and remaining life prediction in combination with an implicit defect evolution law.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring technology, and more specifically, to an online monitoring method and system for complete sets of switchgear based on multi-sensor fusion. Background Technology

[0002] With the rapid development of smart grids, switchgear, as a key component of the power system, directly impacts the safety and stability of the power grid through its operational reliability. However, during long-term operation, switchgear is susceptible to the combined effects of electrical, thermal, and mechanical factors, leading to the gradual accumulation of latent defects such as insulation degradation, mechanical jamming, and increased contact resistance. These defects can ultimately trigger sudden failures, causing severe economic losses. Therefore, achieving real-time monitoring of switchgear operating status, early fault warning, and remaining life assessment is of great significance for improving the level of intelligent operation and maintenance of the power grid.

[0003] Traditional switchgear condition monitoring mainly relies on single sensors such as infrared thermography or partial discharge detection, which has limitations such as data uniformity, difficulty in detecting latent defects, and low accuracy in lifetime assessment. In recent years, multi-sensor fusion technology has provided a new solution to these problems. By integrating data from multiple sensors such as infrared, vibration, current, partial discharge, and gas sensors, a more comprehensive equipment condition characterization system can be constructed. However, multi-source data suffers from heterogeneity, temporal asynchrony, and feature redundancy, making it difficult for traditional fusion methods such as Kalman filtering to effectively uncover deep fault correlation features. Furthermore, existing research focuses primarily on fault diagnosis, with insufficient exploration of latent defect evolution mechanisms and remaining lifetime prediction. The successful application of deep learning technology in image recognition, temporal prediction, and other fields has provided new methods for multi-sensor data fusion and fault assessment. Therefore, there is an urgent need to research a complete online detection method for switchgear based on multi-sensor fusion and deep learning, which can efficiently fuse multi-source data and accurately identify latent defects. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes an online detection method and system for complete sets of switchgear based on multi-sensor fusion. By efficiently fusing multi-source data, accurately identifying latent defects, and establishing a reliable lifespan prediction model, it provides technical support for the intelligent operation and maintenance of power equipment.

[0005] The first aspect of this invention provides an online detection method for complete switchgear based on multi-sensor fusion, comprising the following steps:

[0006] The temperature distribution, partial discharge signal, mechanical vibration waveform and operating current data of the switchgear are collected in real time by infrared thermal imaging sensor, partial discharge sensor, vibration sensor and current sensor, and the multi-source data is preprocessed.

[0007] Key fault features are extracted from preprocessed multi-source data to generate a high-dimensional feature matrix. A graph convolution-long short-term memory hybrid network model is constructed as a fault diagnosis model, and the graph structure is used to model the topological relationship between multiple sensors.

[0008] The high-dimensional feature matrix is ​​imported into the fault diagnosis model to extract spatial correlation features and time-dependent features, and the defect detection results are obtained based on the spatiotemporal features.

[0009] Based on defect characteristics, an equipment degradation index is constructed to calculate the equipment health index. Combined with the Wiener degradation model, the equipment performance degradation trend is characterized, and the remaining service life is predicted probabilistically. When a defect is detected or the service life is below the threshold, an early warning signal is triggered.

[0010] In this solution, the temperature distribution, partial discharge signal, mechanical vibration waveform, and operating current data of the switchgear are collected in real time, and the multi-source data are preprocessed, including:

[0011] Infrared thermal image data of the switchgear is collected to generate a temperature matrix, the temperature value of each pixel is labeled, the temperature distribution of the switchgear is obtained, and the temperature distribution is synchronized with the collected partial discharge signal, mechanical vibration waveform and operating current data in time.

[0012] The multi-source signals acquired in real time are cleaned, and data features are extracted from the cleaned multi-source signals. Based on the temperature distribution, hot spot temperature rise and regional temperature gradient are extracted as temperature features. Based on the partial discharge signal, the discharge amplitude, repetition rate and phase distribution are statistically analyzed as partial discharge features. Based on the mechanical vibration waveform, the kurtosis index and energy ratio are calculated as vibration features. Based on the operating current data, the dynamic resistance change rate and odd harmonic growth trend are analyzed as current features.

[0013] The temperature characteristics, partial discharge characteristics, vibration characteristics, and current characteristics are normalized into a standardized matrix. The neighborhood radius and minimum number of samples for DBSCAN clustering are set. Based on the neighborhood radius and minimum number of samples, clustering is performed to automatically identify core sample clusters, boundary samples, and noise points.

[0014] Abnormal data is identified and marked in the boundary samples and noise points. It is checked whether the abnormal markings appear synchronously in multiple sensors. Data verification is performed based on the inspection results to generate a labeled dataset containing normal data and abnormal data to be diagnosed.

[0015] In this scheme, key fault features are extracted from preprocessed multi-source data to generate a high-dimensional feature matrix, including:

[0016] Multi-dimensional features are extracted from preprocessed normal data and abnormal data to be diagnosed to form an initial feature pool, which includes time-domain features, frequency-domain features, time-frequency joint features and statistical features.

[0017] The mRMR algorithm is used to calculate the importance of each feature in the initial feature pool. The fault sensitivity factor is set by calculating the KL divergence of the feature in the abnormal data sample to be diagnosed. The improved correlation between the feature and the fault category is defined according to the fault sensitivity factor and mutual information. A candidate feature set is obtained, and the feature with the highest improved correlation is selected as the selected feature.

[0018] For the candidate feature set, the dynamic redundancy penalty term is adjusted according to the source sensor type of the candidate features, and the improved redundancy of the selected features is defined based on the dynamic redundancy penalty term and mutual information.

[0019] The feature importance score is calculated based on the maximum relevance and minimum redundancy strategy. Features that meet the preset criteria are selected based on the importance score. An adversarial feature selection mechanism is introduced to train the discriminator and attempt to distinguish the feature distribution of normal data from abnormal data. Through adversarial training, the features that best represent the fault but are not easily disturbed by noise are selected.

[0020] Output the optimal feature subset, and use an adaptive weighting strategy to assign higher weights to the optimal features corresponding to the abnormal data to be diagnosed, thereby constructing a high-dimensional feature matrix and obtaining the weighted high-dimensional feature matrix.

[0021] In this scheme, a graph convolutional-long short-term memory hybrid network model is constructed as a fault diagnosis model. The graph structure is used to model the topological relationships between multiple sensors, including:

[0022] Each sensor is treated as a node, and the node connection relationship is defined based on physical correlation to construct a sensor graph. The input features and initial weights of each node are obtained based on the weighted high-dimensional feature matrix, and the adjacency matrix corresponding to the sensor graph is obtained.

[0023] A graph convolutional-long short-term memory hybrid network model is constructed as a fault diagnosis model. The adjacency matrix is ​​imported into the fault diagnosis model, and graph convolutional layers are used to independently perform graph convolution at each time step to obtain spatial correlation features.

[0024] The spatial correlation features output by the graph convolutional layer are split according to the node dimension. Each node is independently input into the LSTM network, and the temporal dependency features are output. The final state of each node is obtained based on the spatial correlation features and the temporal dependency features.

[0025] An attention mechanism is introduced to calculate the importance weight of nodes in combination with the initial weights. The final state of the nodes is then weighted and fused using the importance weights to obtain global features. These global features are then input into a fully connected layer to obtain the probability distribution of fault categories.

[0026] If the probability of a fault category consistently exceeds the initial threshold but does not reach the fault threshold, it is marked as a latent fault defect.

[0027] In this solution, an equipment health index is calculated by constructing equipment degradation indicators based on defect characteristics, including:

[0028] The n sensor nodes that contribute the most to the current fault category are identified by the importance weight in the fault diagnosis model and labeled. The high-dimensional features corresponding to the labeled sensor nodes are used as the characterization features of the fault category.

[0029] The 95th percentile of the data from the first three months of equipment operation is used as a reference value. Each characteristic is dynamically normalized, and a degradation index is constructed by combining the importance weight and the normalized characteristic value. The larger the degradation index, the further the switching equipment deviates from the normal state.

[0030] The Sigmoid function is used to convert degradation indicators into equipment health indices, and a time decay factor is introduced to correct for lifespan decay, generating dynamically corrected equipment health indices.

[0031] In this scheme, the Wiener degradation model is used to characterize the device performance degradation trend, and a probabilistic prediction of the remaining service life is made, including:

[0032] According to equipment health index As degenerate trajectory observations, a Wiener procedure is defined based on these observations, expressed as follows:

[0033]

[0034] in The linear drift coefficient is... The diffusion coefficient is... This is standard Brownian motion;

[0035] Maximum likelihood estimation is used to obtain the linear drift coefficient characterizing the average degradation rate and the diffusion coefficient characterizing the intensity of random fluctuations based on historical equipment lifecycle data.

[0036] Predefined failure threshold The first crossing time is calculated based on degradation trajectory observations and failure thresholds, and the probability density function of the remaining service life is obtained. , represented as:

[0037]

[0038] The probability density function is used to calculate the future time based on the current device health index. Probabilistic prediction of the remaining useful life within the period.

[0039] In this solution, an early warning signal is triggered when a defect is detected or the lifespan is below a threshold, including:

[0040] Different levels of early warning are set according to the severity of defects and the urgency of remaining service life. The equipment health index is obtained based on real-time multi-source data, and the probability distribution of remaining service life is updated daily based on the Wiener degradation model.

[0041] The system acquires the equipment type and operating environment of the complete set of switchgear, dynamically adjusts the adaptive failure threshold based on the equipment type and operating environment according to historical detection data, outputs early warning signals based on the adaptive failure threshold, and performs visualization processing and multi-channel push.

[0042] The second aspect of the present invention provides an online detection system for complete sets of switchgear based on multi-sensor fusion. The system includes a multi-source data acquisition module, a data preprocessing and fusion module, a fault diagnosis module, a health assessment and life prediction module, and a visualization and early warning module.

[0043] The multi-source data acquisition module uses infrared thermal imaging sensors, partial discharge sensors, vibration sensors, and current sensors to collect real-time data on the temperature distribution, partial discharge signals, mechanical vibration waveforms, and operating current of the switching equipment.

[0044] The data preprocessing and fusion module preprocesses the multi-source data, extracts key fault features based on the preprocessed multi-source data, and generates a high-dimensional feature matrix.

[0045] The fault diagnosis module constructs a graph convolutional-long short-term memory hybrid network model as the fault diagnosis model. It uses graph structure to model the topological relationship between multiple sensors, integrates spatiotemporal features to perform multi-dimensional joint representation of the state of switching equipment, and identifies faults in switching equipment.

[0046] The health assessment and life prediction module constructs equipment degradation indicators based on defect characteristics to calculate the equipment health index, combines the Wiener degradation model to characterize the equipment performance decline trend, and makes a probabilistic prediction of the remaining service life.

[0047] The visualization and early warning module visualizes multi-sensor data, health index, and lifespan prediction trends. When a defect is detected or the lifespan is below a threshold, an early warning signal is triggered, and early warning visualization and push notifications are provided.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] This invention constructs a comprehensive equipment condition monitoring system by integrating multi-source heterogeneous data such as infrared thermography, partial discharge, vibration, and current. The method innovatively employs an improved mRMR algorithm for feature selection, effectively solving the problems of multimodal data redundancy and complementarity, and significantly improving the characterization capability of key fault features. By modeling the topological relationships between sensors using graph convolutional networks and combining this with long short-term memory networks to mine temporal evolution patterns, deep fusion of spatial and temporal features is achieved, resulting in a qualitative leap in the detection sensitivity of latent defects. The introduced attention mechanism further optimizes feature weight allocation, ensuring that the system can accurately capture potential faults such as early insulation degradation and mechanical wear.

[0050] In terms of health status assessment, equipment degradation indicators are constructed based on defect characteristics, and probabilistic prediction of remaining service life is achieved by combining them with the Wiener process model. The introduction of transfer learning strategy effectively solves the model generalization problem under small sample data, enabling newly commissioned equipment to quickly obtain accurate service life assessments. The multi-level early warning mechanism, based on the dynamic changes of health index and remaining service life, realizes a graded response from attention alerts to emergency shutdowns, providing a scientific basis for operation and maintenance decisions. This invention can effectively identify latent faults that are difficult to detect by traditional methods, significantly reducing the risk of unplanned power outages. At the same time, it optimizes the allocation of maintenance resources through accurate service life prediction, thereby improving the overall life cycle management level of switchgear. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0052] Figure 1 A flowchart of an online detection method for complete switchgear based on multi-sensor fusion is shown;

[0053] Figure 2 The flowchart for extracting key fault features and generating a high-dimensional feature matrix is ​​shown.

[0054] Figure 3 The flowchart for constructing a fault diagnosis model for fault identification is shown;

[0055] Figure 4 A block diagram of an online detection system for complete switchgear based on multi-sensor fusion is shown. Detailed Implementation

[0056] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0058] Figure 1 A flowchart of an online detection method for complete switchgear based on multi-sensor fusion is shown.

[0059] like Figure 1 As shown, this embodiment provides an online detection method for complete switchgear based on multi-sensor fusion, including:

[0060] S102 uses infrared thermal imaging sensors, partial discharge sensors, vibration sensors, and current sensors to collect temperature distribution, partial discharge signals, mechanical vibration waveforms, and operating current data of the switchgear in real time, and preprocesses the multi-source data.

[0061] S104: Based on the preprocessed multi-source data, key fault features are extracted to generate a high-dimensional feature matrix. A graph convolution-long short-term memory hybrid network model is constructed as a fault diagnosis model, and the graph structure is used to model the topological relationship between multiple sensors.

[0062] S106, import the high-dimensional feature matrix into the fault diagnosis model, extract spatial correlation features and time-dependent features, and obtain the defect detection results based on the spatiotemporal features;

[0063] S108: Based on defect characteristics, construct equipment degradation indicators to calculate equipment health index, combine Wiener degradation model to characterize equipment performance degradation trend, and make probabilistic predictions of remaining service life. When a defect is detected or the service life is below the threshold, trigger an early warning signal.

[0064] It should be noted that infrared thermal imagers are used to scan key parts of the switchgear, such as circuit breaker contacts and busbar connections, to acquire temperature distribution images. Infrared thermal image data of the switchgear is collected to generate a temperature matrix, and the temperature value of each pixel is labeled to obtain the temperature distribution of the switchgear. Partial discharge pulse signals are captured using high-frequency current transformers or ultra-high-frequency sensors, and parameters such as discharge amplitude, phase, and frequency are recorded, along with the original waveform data, to acquire partial discharge signals. Accelerometers are installed on the operating mechanism or support structure of the switchgear to monitor vibration acceleration waveforms, acquiring triaxial vibration signals (X / Y / Z directions), and extracting time-domain and frequency-domain parameters as mechanical vibration data. The main circuit current is measured using Rogowski coils or Hall effect sensors, and the effective current value, harmonic content, and transient waveform are recorded to acquire operating current data. The temperature distribution is synchronized with the acquired partial discharge signals, mechanical vibration waveforms, and operating current data. Using the current signal as a reference, interpolation is used to align the asynchronous sampling data from other sensors.

[0065] The real-time acquired multi-source signals undergo data cleaning, such as removing abnormal temperature points, eliminating white noise, retaining real discharge pulses, and filtering out equipment background noise. Data features are extracted from the cleaned multi-source signals: temperature features include hotspot temperature rise and regional temperature gradient; partial discharge features include discharge amplitude, repetition rate, and phase distribution; vibration features include kurtosis index and energy proportion calculated from mechanical vibration waveforms; and current features include dynamic resistance change rate and odd harmonic growth trend analyzed from operating current data. These temperature, partial discharge, vibration, and current features are normalized into a standardized matrix. The neighborhood radius and minimum sample number for DBSCAN clustering are set. Based on these neighborhood radius and minimum sample number, clustering automatically identifies core sample clusters, boundary samples, and noise points. Core sample clusters correspond to high-density normal operating conditions, boundary samples correspond to transitional states, and noise points correspond to isolated abnormal data. Abnormal data is identified and marked in the boundary samples and noise points according to the abnormal marking rules. For example, samples with hot spot temperature rise > 15K and regional temperature gradient > 8K / cm in the noise points are regarded as temperature abnormal points. Check whether the abnormal markings appear synchronously in multiple sensors. For example, partial discharge abnormality is accompanied by temperature rise. Data verification is performed based on the inspection results to generate a labeled dataset containing normal data and abnormal data to be diagnosed.

[0066] Figure 2 The flowchart for extracting key fault features and generating a high-dimensional feature matrix is ​​shown.

[0067] According to an embodiment of the present invention, key fault features are extracted from preprocessed multi-source data to generate a high-dimensional feature matrix. This includes: extracting multi-dimensional features from preprocessed normal data and abnormal data to be diagnosed to form an initial feature pool. The initial feature pool includes time-domain features (mean, variance, peak value, waveform factor, etc.), frequency-domain features (FFT spectral energy, wavelet packet decomposition coefficients, harmonic component proportion, etc.), time-frequency joint features (short-time Fourier transform energy distribution, etc.), and statistical features (skewness, kurtosis, entropy, etc.). The mRMR algorithm is used to calculate the importance of each feature in the initial feature pool. A fault sensitivity factor is set by calculating the KL divergence of the features in the abnormal data samples to be diagnosed, enhancing the ability to screen for latent defect features. Based on the fault sensitivity factor and mutual information, the improved correlation between features and fault categories is defined, a candidate feature set is obtained, and the feature with the highest improved correlation is selected as the selected feature. Represented as:

[0068]

[0069] in For the i-th feature, For fault category variables, As a fault-sensitive factor, This is the correlation moderating coefficient. Mutual information function;

[0070] For the candidate feature set, the dynamic redundancy penalty term is adjusted according to the source sensor type of the candidate features, and the improved redundancy of the candidate features and selected features is defined based on the dynamic redundancy penalty term and mutual information. , represented as:

[0071]

[0072] in This is the redundancy adjustment coefficient. Size of the selected feature subset For dynamic redundancy penalty terms, For the j-th candidate feature, Selected features;

[0073] Feature importance scores are calculated based on the maximum relevance and minimum redundancy strategies. Features that meet the preset criteria are selected based on the importance scores. An adversarial feature selection mechanism is introduced to train a discriminator to attempt to distinguish the feature distributions of normal and abnormal data. Through adversarial training, the features that best characterize the fault but are not easily affected by noise are selected. The optimal feature subset is output, and an adaptive weighting strategy is used to assign higher weights to the optimal features corresponding to the abnormal data to be diagnosed, forming a high-dimensional feature matrix. The weighted high-dimensional feature matrix is ​​then obtained.

[0074] Figure 3 The flowchart for constructing a fault diagnosis model for fault identification is shown.

[0075] According to an embodiment of the present invention, a graph convolutional-long short-term memory hybrid network model is constructed as a fault diagnosis model. This model utilizes a graph structure to model the topological relationships between multiple sensors, including: treating each sensor as a node; defining node connections based on physical correlations; constructing coupling edges based on the electromechanical correlations of vibration and current sensors; constructing time-dependent edges based on the thermal accumulation effects of partial discharge and temperature sensors; constructing weakly correlated edges based on the indirect influences of current and infrared sensors; constructing a sensor graph; obtaining the input features and initial weights of each node based on a weighted high-dimensional feature matrix; and obtaining the adjacency matrix corresponding to the sensor graph. The graph convolutional-long short-term memory hybrid network model is then constructed as a fault diagnosis model. The adjacency matrix is ​​imported into the fault diagnosis model. Graph convolutional layers are used to independently perform graph convolution at each time step to obtain spatial correlation features. The spatial correlation features output by the graph convolutional layers are split by node dimension, and each node is independently input into an LSTM network to output temporal dependency features. The final state of each node is obtained based on the spatial correlation features and temporal dependency features. An attention mechanism is introduced to calculate node importance weights in conjunction with initial weights. The final states of nodes are weighted and fused using these importance weights to obtain global features. These global features are input into a fully connected layer to obtain the fault category probability distribution. If a fault category probability consistently exceeds the initial threshold but does not reach the fault threshold, it is marked as a latent fault defect.

[0076] It should be noted that the n sensor nodes with the highest contribution to the current fault category are identified through importance weighting in the fault diagnosis model. The high-dimensional features corresponding to these labeled sensor nodes are used as the characterization features of the fault category. For example, the importance weighting focuses on features such as partial discharge pulse repetition rate, cumulative discharge amplitude, and phase distribution asymmetry. The 95th percentile of the data from the first three months of equipment operation is used as a reference value. Each characterization feature is dynamically normalized, and a degradation index is constructed by combining the importance weights and the normalized feature values. , Let the importance weight of the i-th node be . For normalized eigenvalues, The total number of nodes represents the total number of nodes. The larger the degradation index, the further the switching equipment deviates from its normal state.

[0077] The Sigmoid function is used to convert degradation indicators into equipment health indices. , As the threshold, k is the curve steepness exponent. A time decay factor is introduced to correct for lifespan decay, generating a dynamically corrected equipment health index. , This refers to the aging rate.

[0078] According to equipment health index As degenerate trajectory observations, a Wiener procedure is defined based on these observations, expressed as follows:

[0079]

[0080] in The linear drift coefficient is... The diffusion coefficient is... This is standard Brownian motion;

[0081] Maximum likelihood estimation is used to obtain the linear drift coefficient representing the average degradation rate and the diffusion coefficient representing the intensity of random fluctuations based on historical equipment lifecycle data. For new equipment data, sequential Monte Carlo (SMC) is used to dynamically update the linear drift coefficient and the diffusion coefficient.

[0082] Predefined failure threshold The first crossing time is calculated based on degradation trajectory observations and failure thresholds, and the probability density function of the remaining service life is obtained. , represented as:

[0083]

[0084] The probability density function is used to calculate the future time based on the current device health index. Probabilistic prediction of the remaining useful life within the period.

[0085] Using similar switching equipment with abundant historical data as the source domain and new equipment with limited data as the target domain, transfer learning is employed to train fault diagnosis and Wiener degradation models with the objective of minimizing the maximum mean difference. Parameters are dynamically adjusted to adapt to individual equipment differences, and transfer learning ensures that the model maintains high accuracy even on new equipment with scarce data.

[0086] It should be noted that the severity of defects is set based on the comparison between the health index and preset failure thresholds and warning thresholds. The urgency of the remaining service life is assessed based on the number of days remaining in the service life. Trigger conditions for different levels of warnings are preset. The equipment health index is obtained based on real-time multi-source data, and the probability distribution of the remaining service life is updated daily based on the Wiener degradation model. The equipment type and operating environment of the complete set of switchgear are obtained. Based on historical detection data, the adaptive failure threshold is dynamically adjusted according to the equipment type and operating environment. For example, the failure threshold of outdoor GIS equipment is higher than that of indoor switchgear. Based on the adaptive failure threshold, warning signals are output, visualized, and pushed through multiple channels. The system automatically associates the most recent defect characteristics, key sensor data trends, and historical failure cases of similar equipment for maintenance reference.

[0087] Figure 4 An architecture diagram of an online detection system for complete switchgear based on multi-sensor fusion is shown.

[0088] The second embodiment of the present invention provides an online detection system for complete sets of switchgear based on multi-sensor fusion. The system includes a multi-source data acquisition module, a data preprocessing and fusion module, a fault diagnosis module, a health assessment and life prediction module, and a visualization and early warning module.

[0089] The multi-source data acquisition module uses infrared thermal imaging sensors, partial discharge sensors, vibration sensors, and current sensors to collect real-time data on the temperature distribution, partial discharge signals, mechanical vibration waveforms, and operating current of the switching equipment.

[0090] The data preprocessing and fusion module preprocesses the multi-source data, and extracts key fault features from the time domain, frequency domain, and time frequency domain based on the preprocessed multi-source data, generating a high-dimensional feature matrix.

[0091] The fault diagnosis module constructs a graph convolutional-long short-term memory hybrid network model as the fault diagnosis model. It uses graph structure to model the topological relationship between multiple sensors, integrates spatiotemporal features to perform multi-dimensional joint representation of the state of switching equipment, and identifies faults in switching equipment.

[0092] The health assessment and life prediction module constructs equipment degradation indicators based on defect characteristics to calculate the equipment health index, combines the Wiener degradation model to characterize the equipment performance decline trend, and makes a probabilistic prediction of the remaining service life.

[0093] The visualization and early warning module visualizes multi-sensor data, health index, and lifespan prediction trends. When a defect is detected or the lifespan is below a threshold, an early warning signal is triggered, and early warning visualization and push notifications are provided.

[0094] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for an online detection method of a complete set of switchgear based on multi-sensor fusion. When the program for the online detection method of a complete set of switchgear based on multi-sensor fusion is executed by a processor, it implements the steps of the online detection method of a complete set of switchgear based on multi-sensor fusion.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0096] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online detection of switchgear based on multi-sensor fusion, characterized in that, Includes the following steps: The temperature distribution, partial discharge signal, mechanical vibration waveform and operating current data of the switchgear are collected in real time by infrared thermal imaging sensor, partial discharge sensor, vibration sensor and current sensor, and the multi-source data is preprocessed. Key fault features are extracted from preprocessed multi-source data to generate a high-dimensional feature matrix. A graph convolution-long short-term memory hybrid network model is constructed as a fault diagnosis model, and the graph structure is used to model the topological relationship between multiple sensors. The high-dimensional feature matrix is ​​imported into the fault diagnosis model to extract spatial correlation features and time-dependent features, and the defect detection results are obtained based on the spatiotemporal features. Based on defect characteristics, an equipment degradation index is constructed to calculate the equipment health index. Combined with the Wiener degradation model, the equipment performance degradation trend is characterized, and the remaining service life is predicted probabilistically. When a defect is detected or the service life is lower than the threshold, an early warning signal is triggered. Key fault features are extracted from the preprocessed multi-source data to generate a high-dimensional feature matrix, including: Multi-dimensional features are extracted from preprocessed normal data and abnormal data to be diagnosed to form an initial feature pool, which includes time-domain features, frequency-domain features, time-frequency joint features and statistical features. The mRMR algorithm is used to calculate the importance of each feature in the initial feature pool. The fault sensitivity factor is set by calculating the KL divergence of the feature in the abnormal data sample to be diagnosed. The improved correlation between the feature and the fault category is defined according to the fault sensitivity factor and mutual information. A candidate feature set is obtained, and the feature with the highest improved correlation is selected as the selected feature. For the candidate feature set, the dynamic redundancy penalty term is adjusted according to the source sensor type of the candidate features, and the improved redundancy of the selected features is defined based on the dynamic redundancy penalty term and mutual information. The feature importance score is calculated based on the maximum relevance and minimum redundancy strategy. Features that meet the preset criteria are selected based on the importance score. An adversarial feature selection mechanism is introduced to train the discriminator and attempt to distinguish the feature distribution of normal data from abnormal data. Through adversarial training, the features that best represent the fault but are not easily disturbed by noise are selected. Output the optimal feature subset, and use an adaptive weighting strategy to assign higher weights to the optimal features corresponding to the abnormal data to be diagnosed, thereby constructing a high-dimensional feature matrix and obtaining the weighted high-dimensional feature matrix.

2. The method for online detection of switchgear assembly based on multi-sensor fusion according to claim 1, characterized in that, Real-time acquisition of temperature distribution, partial discharge signals, mechanical vibration waveforms, and operating current data of switchgear; preprocessing of multi-source data, including: Infrared thermal image data of the switchgear is collected to generate a temperature matrix, the temperature value of each pixel is labeled, the temperature distribution of the switchgear is obtained, and the temperature distribution is synchronized with the collected partial discharge signal, mechanical vibration waveform and operating current data in time. The multi-source signals acquired in real time are cleaned, and data features are extracted from the cleaned multi-source signals. Based on the temperature distribution, hot spot temperature rise and regional temperature gradient are extracted as temperature features. Based on the partial discharge signal, the discharge amplitude, repetition rate and phase distribution are statistically analyzed as partial discharge features. Based on the mechanical vibration waveform, the kurtosis index and energy ratio are calculated as vibration features. Based on the operating current data, the dynamic resistance change rate and odd harmonic growth trend are analyzed as current features. The temperature characteristics, partial discharge characteristics, vibration characteristics, and current characteristics are normalized into a standardized matrix. The neighborhood radius and minimum number of samples for DBSCAN clustering are set. Based on the neighborhood radius and minimum number of samples, clustering is performed to automatically identify core sample clusters, boundary samples, and noise points. Abnormal data is identified and marked in the boundary samples and noise points. It is checked whether the abnormal markings appear synchronously in multiple sensors. Data verification is performed based on the inspection results to generate a labeled dataset containing normal data and abnormal data to be diagnosed.

3. The method for online detection of switchgear assembly based on multi-sensor fusion according to claim 1, characterized in that, A graph convolutional-long short-term memory hybrid network model is constructed as a fault diagnosis model. This model utilizes graph structures to model the topological relationships between multiple sensors, including: Each sensor is treated as a node, and the node connection relationship is defined based on physical correlation to construct a sensor graph. The input features and initial weights of each node are obtained based on the weighted high-dimensional feature matrix, and the adjacency matrix corresponding to the sensor graph is obtained. A graph convolutional-long short-term memory hybrid network model is constructed as a fault diagnosis model. The adjacency matrix is ​​imported into the fault diagnosis model, and graph convolutional layers are used to independently perform graph convolution at each time step to obtain spatial correlation features. The spatial correlation features output by the graph convolutional layer are split according to the node dimension. Each node is independently input into the LSTM network, and the temporal dependency features are output. The final state of each node is obtained based on the spatial correlation features and the temporal dependency features. An attention mechanism is introduced to calculate the importance weight of nodes in combination with the initial weights. The final state of the nodes is then weighted and fused using the importance weights to obtain global features. These global features are then input into a fully connected layer to obtain the probability distribution of fault categories. If the probability of a fault category consistently exceeds the initial threshold but does not reach the fault threshold, it is marked as a latent fault defect.

4. The method for online detection of switchgear assembly based on multi-sensor fusion according to claim 1, characterized in that, Based on defect characteristics, equipment degradation indicators are constructed to calculate the equipment health index, including: The n sensor nodes that contribute the most to the current fault category are identified by the importance weight in the fault diagnosis model and labeled. The high-dimensional features corresponding to the labeled sensor nodes are used as the characterization features of the fault category. The 95th percentile of the data from the first three months of equipment operation is used as a reference value. Each characteristic is dynamically normalized, and a degradation index is constructed by combining the importance weight and the normalized characteristic value. The larger the degradation index, the further the switching equipment deviates from the normal state. The Sigmoid function is used to convert degradation indicators into equipment health indices, and a time decay factor is introduced to correct for lifespan decay, generating dynamically corrected equipment health indices.

5. The method for online detection of switchgear assembly based on multi-sensor fusion according to claim 1, characterized in that, By combining the Wiener degradation model to characterize the device performance degradation trend, probabilistic predictions of remaining useful life are made, including: According to the device health index As a degradation trajectory observation, a Wiener process is defined based on the degradation trajectory observation, denoted as: wherein is a linear drift coefficient, is a diffusion coefficient, is a standard Brownian motion; Maximum likelihood estimation is used to obtain the linear drift coefficient characterizing the average degradation rate and the diffusion coefficient characterizing the intensity of random fluctuations based on historical equipment lifecycle data. Predefined failure threshold The first crossing time is calculated according to the degradation trajectory observation value and the failure threshold, and the probability density function of the remaining service life is obtained , which is expressed as: calculating a probabilistic prediction of remaining useful life at a future time using the probability density function based on the current device health index. calculating a probabilistic prediction of remaining useful life at a future time using the probability density function based on the current device health index.

6. The method for online detection of switchgear equipment based on multi-sensor fusion according to claim 1, characterized in that, Warning signals are triggered when defects are detected or the lifespan falls below a threshold, including: Different levels of early warning are set according to the severity of defects and the urgency of remaining service life. The equipment health index is obtained based on real-time multi-source data, and the probability distribution of remaining service life is updated daily based on the Wiener degradation model. The system acquires the equipment type and operating environment of the complete set of switchgear, dynamically adjusts the adaptive failure threshold based on the equipment type and operating environment according to historical detection data, outputs early warning signals based on the adaptive failure threshold, and performs visualization processing and multi-channel push.

7. A multi-sensor fusion based switchgear online detection system, characterized in that, The system is used to implement the online detection method for complete switchgear based on multi-sensor fusion as described in any one of claims 1-6. The system includes a multi-source data acquisition module, a data preprocessing and fusion module, a fault diagnosis module, a health assessment and life prediction module, and a visualization and early warning module. The multi-source data acquisition module uses infrared thermal imaging sensors, partial discharge sensors, vibration sensors, and current sensors to collect real-time data on the temperature distribution, partial discharge signals, mechanical vibration waveforms, and operating current of the switching equipment. The data preprocessing and fusion module preprocesses the multi-source data, extracts key fault features based on the preprocessed multi-source data, and generates a high-dimensional feature matrix. The fault diagnosis module constructs a graph convolutional-long short-term memory hybrid network model as the fault diagnosis model. It uses graph structure to model the topological relationship between multiple sensors, integrates spatiotemporal features to perform multi-dimensional joint representation of the state of switching equipment, and identifies faults in switching equipment. The health assessment and life prediction module constructs equipment degradation indicators based on defect characteristics to calculate the equipment health index, and combines the Wiener degradation model to characterize the equipment performance degradation trend and make a probabilistic prediction of the remaining service life. The visualization and early warning module visualizes multi-sensor data, health index, and lifespan prediction trends. When a defect is detected or the lifespan is below a threshold, an early warning signal is triggered, and early warning visualization and push notifications are provided.

Images