A switch cabinet fault prediction method based on multi-source data fusion

Abstract

The application belongs to the technical field of fault prediction, and particularly relates to a switch cabinet fault prediction method based on multi-source data fusion, steps of which comprise: collecting multi-source state data in the operation process of a switch cabinet, obtaining standardized data after preprocessing, and performing feature extraction to construct a multi-source feature set; based on a combined weighting method of analytic hierarchy process and entropy weight method, weight coefficients of various features in the multi-source feature set are obtained, and fusion processing is performed to obtain a comprehensive state feature vector; a preset time window length is used to slide and intercept the comprehensive state feature vectors that are continuous in time sequence to construct a time window sequence; the time window sequence is input into a pre-trained fault prediction model to output a comprehensive fault risk value of the switch cabinet at the current time or a future preset time step; based on the comprehensive state feature vector, a device health index is calculated, risk joint determination is performed, and fault early warning information of the switch cabinet is output. The application can realize accurate prediction, conflict review and hierarchical early warning of switch cabinet faults.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction technology, specifically relating to a switchgear fault prediction method based on multi-source data fusion. Background Technology

[0002] Switchgear is a crucial power transmission and distribution equipment in power systems, used for power distribution, control, and protection. Its operational reliability directly affects power supply security and grid stability. Switchgear operates under conditions of high voltage, high current, strong electromagnetic fields, and complex environments for extended periods, making it prone to faults or defects such as contact overheating, insulation dampness, condensation, mechanical jamming, abnormal opening and closing, and fluctuations in electrical parameters.

[0003] In switchgear, vacuum circuit breakers and their key components such as contacts, contact arms, busbars, and insulation parts have a significant impact on overall safety. Abnormal mechanical characteristics of circuit breakers can usually be reflected by stroke-time curves, opening and closing coil current curves, opening and closing times, contact movement speed, and mechanical vibration signals; thermal faults are usually related to load current, ambient temperature, busbar temperature, upper contact arm temperature, hot spot distribution within the cabinet, and contact temperature rise; humidity and condensation risks are closely related to the relative humidity within the cabinet, the distribution of local temperature and humidity fields, and the surface condition of insulation parts.

[0004] Existing switchgear condition monitoring methods mostly analyze a single object or a single data source. For example, they determine the circuit breaker's operating status solely based on mechanical characteristic curves, or detect abnormal temperature rises using only temperature sensors or infrared thermometers. While these methods can reflect the operating status from one perspective, they struggle to simultaneously utilize mechanical, thermal, humidity, and electrical operating information, and cannot fully characterize the coupling relationships between multiple data sources.

[0005] Meanwhile, in practical engineering applications, the internal structure of switchgear is compact and it operates under power, making the contact points generally unsuitable as stable and essential direct measurement points. Although some equipment is equipped with wireless temperature measurement or fiber optic temperature measurement devices, the number of measurement points, installation location, online status, and sampling reliability are greatly affected by the equipment structure and operating environment. Therefore, direct temperature measurement values ​​are more suitable as calibration samples than as necessary basic inputs.

[0006] Furthermore, existing early warning methods often employ single-threshold or simple dual-threshold approaches for judgment. For example, a failure probability exceeding a threshold is considered high-risk, or a health index below a threshold is considered severe. Such methods lack a verification mechanism when the failure prediction model output differs from the health index evaluation results. This can lead to situations where a very high failure probability is directly classified as severe when the health index remains good, or where a significant degradation in the health index is not adequately indicated when the model probability is low, thus affecting the rationality and interpretability of the early warning results.

[0007] Therefore, it is necessary to propose a switchgear fault prediction method that can integrate multi-source heterogeneous data, clearly process some collectable and unmeasurable key parameters, and refine the judgment of the consistency between fault probability and health index, so as to improve the accuracy, stability and engineering feasibility of switchgear fault prediction. Summary of the Invention

[0008] In view of the shortcomings of the prior art, the purpose of this invention is to provide a switchgear fault prediction method based on multi-source data fusion. This method integrates heterogeneous data from multiple sources, including mechanical, thermal, humidity, and electrical data. It solves the problem of unmeasurable key parameters by inverting contact temperature and integrating reliability data. Furthermore, it introduces a consistency verification mechanism between fault probability and health index to achieve accurate prediction, conflict verification, and graded early warning of switchgear faults.

[0009] To achieve the above objectives, this invention provides a switchgear fault prediction method based on multi-source data fusion, comprising the following steps: S1. Collect multi-source status data during the operation of the switchgear, including mechanical status data, measurable thermal status data, humidity status data, and electrical operation data; S2. Perform time synchronization, missing value processing, collection anomaly identification, and normalization on multi-source state data to obtain standardized data; S3. Based on standardized data, feature extraction is performed on mechanical condition data, thermal condition measurable data, humidity condition data and electrical operation data respectively to construct a multi-source feature set; S4. Based on the combined weighting method of the analytic hierarchy process and the entropy weighting method, the weight coefficients of various features in the multi-source feature set are obtained and fused to obtain the comprehensive state feature vector. S5. Slide and truncate the continuous comprehensive state feature vectors in time sequence with a preset time window length to construct a time window sequence; S6. Input the time window sequence into the pre-trained fault prediction model and output the comprehensive fault risk value of the switchgear at the current moment or at a preset future time step. S7. Calculate the equipment health index based on the comprehensive state feature vector, obtain the equipment degradation degree using the equipment health index, and combine the comprehensive fault risk value, equipment degradation degree and the consistency index of the two to make a joint risk judgment and output the fault warning information of the switchgear.

[0010] As a preferred embodiment of the present invention, in S1, the mechanical state data includes the travel-time curve, opening and closing coil current curve, opening and closing time, contact movement speed, and mechanical vibration signal during the circuit breaker opening and closing process; the measurable thermal state data includes ambient temperature, busbar temperature, upper contact arm temperature, temperature distribution of hot spots in the cabinet, and load-related thermal condition data; the humidity state data includes the relative humidity of the air in the cabinet, humidity distribution parameters, and the surface humidity and condensation state of key insulating components collected when measuring points are available; the electrical operation data includes load current, voltage, power, load rate, and operating condition parameters. Key insulation components include vacuum interrupter insulating bushings, post insulators, contact boxes, busbar insulating sheaths, insulating tie rods, wall bushings, and cable terminals.

[0011] As a preferred embodiment of the present invention, the process of obtaining standardized data in S2 is as follows: Time synchronization: Using the high-precision clock of the unified data acquisition controller of the switch cabinet as a reference, data with different acquisition frequencies and different transmission delays are calibrated to the same time axis; Missing value handling: For data gaps, a weighted linear interpolation method based on a sliding time window is used to fill in the gaps, and the interpolation weight is determined by the time distance between the adjacent valid samples before and after the missing point. Anomaly identification: For outliers, the 3σ criterion, the characteristics of abrupt changes in temporal slope, and the consistency of multi-source physical correlation are used to distinguish them: when a single data source has an isolated outlier and does not meet the continuity of adjacent time or the correlation of other data sources, it is marked as an acquisition anomaly and corrected or removed; when multiple data sources change in correlation over time and meet the thermal, electrical, humidity or mechanical constraint relationship, they are retained as samples of potential fault precursors. Normalization: The maximum-minimum normalization method is used to map parameters of different dimensions to the interval [0, 1] to obtain standardized data.

[0012] As a preferred embodiment of the present invention, in S3, for mechanical state data, dynamic time warping and curve morphology feature fusion extraction are adopted. Based on the standard opening and closing stroke-time curve and the standard coil current curve, the dynamic time warping distance between the measured curve and the standard curve is calculated, and the opening and closing time, overtravel, number of bounces, contact movement speed gradient, mechanical vibration peak value and frequency domain energy features are extracted to obtain the mechanical state feature vector. For measurable thermal state data, a joint extraction of contact temperature equivalent inversion and thermodynamic features is adopted. An indirect contact temperature inversion model is constructed based on ambient temperature, load current, upper contact arm temperature, and busbar temperature. The inverted contact temperature value is used as the basic output. When calibration samples with measured contact temperatures are available, the inversion values ​​are fused to obtain the contact equivalent temperature features. At the same time, the hot spot temperature distribution and temperature rise rate in the cabinet are inverted through electro-thermal-fluid coupling simulation and regression model. The temperature gradient, hot spot offset, and heat accumulation features are extracted to obtain the thermal state feature vector. For humidity status data, when measurement points are unavailable, a fusion inversion method combining multiphysics constraints and physical information neural networks (PINN) is used. The coupling laws of electric field, temperature field, flow field, and humidity field serve as physical constraints to invert the surface humidity, local condensation probability, humidity non-uniformity, and critical condensation duration of key insulating components, yielding a humidity status feature vector. When measurement points are available, based on the collected surface humidity and condensation status of key insulating components, as well as the relative humidity and humidity distribution parameters of the air inside the cabinet, the local condensation probability, humidity non-uniformity, and critical condensation duration are obtained. For electrical operation data, operating condition adaptation and time-series fluctuation feature extraction are adopted to extract the effective value of load current, voltage distortion rate, power fluctuation coefficient, load rate time-series features, operating condition adaptation degree and electrical parameter mutation features to obtain electrical operation feature vector; The multi-source feature set is represented as: ; In the formula, Represents the set of multi-source features at sampling time t; , , , These represent the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector at sampling time t, respectively.

[0013] As a preferred embodiment of the present invention, when performing feature extraction on mechanical state data, the dynamic time-normalized distance between the measured curve X and the standard curve Y is: ; In the formula, This represents the optimal time warping path from X to Y; K is the total number of matching points on the optimal warping path, and k is the index of the matching point; This represents the measured data point corresponding to the k-th matching point. Compared with standard data points The Euclidean distance between them; This means taking the minimum value among all possible matching paths; When performing feature extraction on measurable thermal state data, the contact temperature is used as the base value based on the inverted value. for: ; In the formula, This represents the m-th standardized input variable, including ambient temperature, load current, upper contact arm temperature, and busbar temperature, where M is the number of standardized input variables. For the corresponding The regression coefficients, This is the error term; When a contact temperature measuring device is installed in the switchgear, the measured contact temperature is used as a calibration sample. The calibration sample and the inversion value are fused according to the confidence coefficient c to obtain the equivalent temperature characteristics of the contact. ; In the formula, Indicates the contact equivalent temperature; Indicates calibration sample; When the switch cabinet is not equipped with a contact temperature measuring device, or the calibration sample is unreliable, c=0; When extracting features from humidity state data, the loss function L of the physical information neural network is: ; In the formula, This represents the data error term between the model-predicted humidity value and the measured humidity data; Represents the physical constraint residuals of the coupled control equations of the electric field-temperature field-flow field-humidity field; , They are respectively , The weighting coefficients.

[0014] In a preferred embodiment of the present invention, in step S4, the combined weighting specifically involves obtaining the subjective weights of each feature vector based on the analytic hierarchy process (AHP), obtaining the objective weights of each feature vector based on the entropy weighting method, and finally obtaining the comprehensive weights of each feature vector. ; In the formula, This represents the combined weight of the i-th eigenvector. ; This represents the subjective weight of the i-th eigenvector; Represents the objective weight of the i-th eigenvector; Indicates the balance coefficient; The comprehensive state feature vector at sampling time t Represented as: ; In the formula, , , , These represent the combined weights of the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector, respectively.

[0015] In a preferred embodiment of the present invention, in step S5, the comprehensive state feature vector is slidably truncated using a preset time window length L to construct a time window sequence: ; In the formula, This represents a time window sequence at sampling time t; , These represent the combined state feature vectors at sampling times t-L+1 and t-L+2, respectively.

[0016] As a preferred embodiment of the present invention, in S6, the fault prediction model adopts a multi-input single-output time series neural network model based on a gated recurrent GRU unit. During pre-training, a training set is constructed using historical operating data of the switchgear, including historical multi-source feature sequences and corresponding fault labels. The cross-entropy loss function is used for supervised training, and the model parameters are optimized through the backpropagation algorithm. When making online predictions, As input to the model, the GRU unit receives the comprehensive state feature vector at the current time step. Hidden state from the previous moment Update to hidden state The hidden states within the time window are aggregated, and the failure probability at a future preset time step is output through a fully connected layer. ; In the formula, Indicates a future preset time step The probability of failure, when When the value is 0, output the comprehensive fault risk value at the current moment. ; These are the output layer weights; For bias terms; pool indicates pooling operation; This represents the hidden state at sampling time t-L+1; To and The corresponding bias parameters; The output failure probability is used as the comprehensive failure risk value.

[0017] As a preferred embodiment of the present invention, in S7, the joint risk determination specifically refers to, for Calculate the device health index at the current moment. Equipment degradation as well as and Consistency indicators between : ; ; ; In the formula, This represents the standard comprehensive state feature vector of the switchgear under healthy operating conditions; dist represents the Euclidean distance; clip represents the interval truncation function. This represents the sum of all comprehensive state feature vectors in historical operational data. The maximum Euclidean distance between them; Calculate the fusion risk value at the current moment. : ; In the formula, , They represent , Weighting coefficients; Joint determination based on hierarchical thresholds: when and ,or and At that time, a severe warning will be issued; when and ,or When, issue an early warning; when and ,or and When the conflict occurs, output the model conflict review status and the dominant abnormal data source causing the conflict; otherwise, output the normal status. in, , The fault probability threshold and , , The equipment degradation threshold and , , To integrate risk thresholds and , This is the consistency threshold; Similarly, for Calculate the future preset time step The risk value of fusion is determined and a joint judgment is made based on the hierarchical threshold.

[0018] As a preferred embodiment of the present invention, in S7, the fault warning information includes risk level, comprehensive fault risk value, equipment health index, consistency index, dominant abnormal data source, source identifier of contact equivalent temperature, fault type, and maintenance suggestion; the fault type includes abnormal mechanical characteristics of circuit breaker, contact overheating, abnormal temperature rise of hot spots in cabinet, risk of insulation dampness or condensation, and abnormal electrical operation; the maintenance suggestion is generated based on the risk level and dominant abnormal data source and is used to guide the predictive maintenance of switchgear.

[0019] The beneficial effects of this invention are: This invention collaboratively models mechanical, thermal, humidity, and electrical operating conditions, enabling it to reflect the true operating status of the switchgear from multiple perspectives, such as motion characteristics, temperature rise characteristics, insulation moisture risk, and operating condition fluctuations, thus avoiding the problem of insufficient diagnostic information from a single data source.

[0020] This invention clarifies the processing method for contact temperature: the contact temperature is used as the target parameter to be inverted, and the basic value is obtained by inverting measurable parameters such as ambient temperature, load current, upper contact arm temperature, and busbar temperature; when some equipment is equipped with a reliable contact temperature measuring device, the direct temperature measurement value is only used as a calibration sample to participate in the credibility fusion, forming a unified contact equivalent temperature feature, thereby eliminating the technical contradiction between data acquisition and unmeasurable inversion.

[0021] This invention distinguishes between collected anomalies and fault precursors during the preprocessing stage, avoiding the accidental deletion of multi-source related mutations in the actual fault development process and improving the sample effectiveness of the fault prediction model. It adopts a combined weighting method of analytic hierarchy process and entropy weighting method, taking into account both expert experience and objective differences in data, so that the multi-source feature fusion results are more stable and better meet the needs of equipment condition evaluation.

[0022] This invention uses a time window-based GRU time series model to output the failure probability, which can learn the trend of switchgear status evolution over time and realize risk prediction for future preset time steps. In the joint risk judgment, the invention introduces equipment degradation degree and consistency index, which can perform conflict review on the inconsistency between failure probability and health index, and will not directly judge serious warning due to excessively high single probability index, thereby improving the rationality, interpretability and engineering usability of the warning results. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the principle of this invention; Figure 2 This is the verification diagram of contact temperature equivalent inversion and calibration fusion in Embodiment 1 of the present invention. Detailed Implementation

[0024] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, a switchgear fault prediction method based on multi-source data fusion includes the following steps: S1. Collect multi-source status data during the operation of the switchgear, including mechanical status data, measurable thermal status data, humidity status data, and electrical operation data; S2. Perform time synchronization, missing value processing, collection anomaly identification, and normalization on multi-source state data to obtain standardized data; S3. Based on standardized data, feature extraction is performed on mechanical condition data, thermal condition measurable data, humidity condition data and electrical operation data respectively to construct a multi-source feature set; S4. Based on the combined weighting method of the analytic hierarchy process and the entropy weighting method, the weight coefficients of various features in the multi-source feature set are obtained and fused to obtain the comprehensive state feature vector. S5. Slide and truncate the continuous comprehensive state feature vectors in time sequence with a preset time window length to construct a time window sequence; S6. Input the time window sequence into the pre-trained fault prediction model and output the comprehensive fault risk value of the switchgear at the current moment or at a preset future time step. S7. Calculate the equipment health index based on the comprehensive state feature vector, obtain the equipment degradation degree using the equipment health index, and combine the comprehensive fault risk value, equipment degradation degree and the consistency index of the two to make a joint risk judgment and output the fault warning information of the switchgear.

[0025] In S1, mechanical status data includes the travel-time curve, opening and closing coil current curve, opening and closing time, contact movement speed, and mechanical vibration signal during the circuit breaker's opening and closing processes; measurable thermal status data includes ambient temperature, busbar temperature, upper contact arm temperature, temperature distribution of hot spots inside the cabinet, and load-related thermal condition data; humidity status data includes the relative humidity of the air inside the cabinet, humidity distribution parameters, and, when measurement points are available, the surface humidity and condensation status of key insulating components; electrical operation data includes load current, voltage, power, load rate, and operating condition parameters. Key insulation components include vacuum interrupter insulating bushings, post insulators, contact boxes, busbar insulating sheaths, insulating tie rods, wall bushings, and cable terminals.

[0026] Contact temperature is not a mandatory direct acquisition quantity, but is used as a target parameter for inversion. When the switchgear is equipped with a reliable contact temperature measurement device, its output contact temperature value is used as a calibration sample for subsequent reliability fusion. When no measurement point is configured, the measurement point is limited, the measurement point is offline, or the sampling fluctuation exceeds the reliability range, the direct contact temperature value is not used. For the surface humidity and condensation state of critical insulation components where measurement points cannot be arranged, these are used as the target parameter for humidity inversion.

[0027] Load-related thermal condition data supplements the temperature measurement point information in the measurable thermal condition data. It is used to characterize the heat generation and heat dissipation status of equipment under different load conditions. This includes temperature data of measurable points of key conductive components (such as contact arms, busbars, and cable joint temperatures), heat dissipation condition data such as cabinet and ambient temperature and ventilation speed, as well as the cumulative equivalent thermal effect data calculated based on the load. The cumulative equivalent thermal effect calculated based on the load is obtained by converting the load current in the electrical operation data using the thermal effect formula.

[0028] In S2, the process of obtaining standardized data is as follows: Time synchronization: Using the high-precision clock of the unified data acquisition controller of the switch cabinet as a reference, data with different acquisition frequencies and different transmission delays are calibrated to the same time axis; Missing value handling: For data gaps, a weighted linear interpolation method based on a sliding time window is used to fill in the gaps, and the interpolation weight is determined by the time distance between the adjacent valid samples before and after the missing point. Anomaly identification: For outliers, the 3σ criterion, the characteristics of abrupt changes in temporal slope, and the consistency of multi-source physical correlation are used to distinguish them: when a single data source has an isolated outlier and does not meet the continuity of adjacent time or the correlation of other data sources, it is marked as an acquisition anomaly and corrected or removed; when multiple data sources change in correlation over time and meet the thermal, electrical, humidity or mechanical constraint relationship, they are retained as samples of potential fault precursors. For a time series from a single data source, adjacent time points should exhibit a smooth transition and continuity during normal operation. If the difference between the current data point and the data from its immediate neighbors exceeds a preset continuity threshold, or if the standard deviation within the sliding window significantly deviates from the normal fluctuation range, it indicates that the data point has experienced a physically meaningless abrupt change and does not meet the requirement of continuity between adjacent time points.

[0029] Based on historical health data, a benchmark correlation model is established between different data sources, such as the correlation or regression relationship between load current and temperature rise, mechanical stroke and current curve, and ambient humidity and leakage current. When the current data of a certain data source is substituted into the correlation model, if there is a significant deviation from the measured values ​​of other correlated data sources and the deviation exceeds the preset threshold, it indicates that the data point does not meet the normal correlation change with other data sources.

[0030] Thermal, electrical, humidity, or mechanical constraints refer to the inherent constraints that must be satisfied by physical mechanisms between thermal, electrical, humidity, and mechanical data during the operation of the switchgear. Electrical and thermal constraints: The Joule heat of a conductor is determined by the square of the current and the resistance. The temperature rise of the equipment should have a stable linear or exponential relationship with the change of the load current. An abnormal increase in current should be accompanied by a synchronous increase in the temperature of the corresponding part, and vice versa. Humidity and heat constraints: Increased ambient humidity will accelerate condensation and contamination on the surface of insulating components, resulting in a corresponding increase in leakage current and partial discharge. Furthermore, temperature rise changes will affect the conditions for condensation formation. The coupled changes in humidity and temperature rise must conform to the law of thermal and humidity balance. Electromechanical constraints: The stroke and speed curves of the opening and closing mechanical actions should be strictly matched with the electrical signals such as coil current and contact bounce current in terms of timing. Mechanical jamming and bounce will cause current waveform distortion and increased contact resistance. The timing deviation between mechanical characteristics and electrical signals should be controlled within the preset range. Temperature and humidity and insulation constraints: Abnormal temperature and humidity inside the cabinet will affect the thermal aging and moisture absorption process of the insulation components, thereby changing their insulation resistance and dielectric loss characteristics. A stable negative or positive correlation must be maintained between temperature and humidity and insulation electrical parameters.

[0031] Normalization: The maximum-minimum normalization method is used to map parameters of different dimensions to the interval [0, 1] to obtain standardized data.

[0032] In S3, for mechanical state data, dynamic time warping and curve morphology feature fusion extraction are adopted. Based on the standard opening and closing stroke-time curve and the standard coil current curve, the dynamic time warping distance between the measured curve and the standard curve is calculated, and the opening and closing time, overtravel, number of bounces, contact movement speed gradient, mechanical vibration peak value and frequency domain energy features are extracted to obtain the mechanical state feature vector. For measurable thermal state data, a joint extraction of contact temperature equivalent inversion and thermodynamic features is adopted. An indirect contact temperature inversion model is constructed based on ambient temperature, load current, upper contact arm temperature, and busbar temperature. The inverted contact temperature value is used as the basic output. When calibration samples with measured contact temperatures are available, the inversion values ​​are fused to obtain the contact equivalent temperature features. At the same time, the hot spot temperature distribution and temperature rise rate in the cabinet are inverted through electro-thermal-fluid coupling simulation and regression model. The temperature gradient, hot spot offset, and heat accumulation features are extracted to obtain the thermal state feature vector. The indirect temperature inversion model can use linear regression model, support vector regression (SVR) model, or radial basis function (RBF) neural network model based on the principle of thermal balance. It takes measurable parameters such as load current, ambient temperature, and upper contact arm temperature as inputs, and fits the unmeasurable contact temperature and hot spot temperature to achieve rapid and accurate inversion of the internal temperature of the switch cabinet.

[0033] For example, the implementation process of the temperature indirect inversion model based on radial basis function (RBF) neural network is as follows: First, a sample dataset is constructed using measurable parameters such as load current, ambient temperature, and upper contact arm temperature as input features, and contact and hot spot temperatures obtained from finite element simulation or calibration tests as output labels. Then, the center and width of the radial basis function are determined using the K-means clustering algorithm, and a three-layer RBF neural network containing an input layer, a hidden layer (radial basis layer), and an output layer is constructed. Next, the weights from the hidden layer to the output layer are trained using the least squares method to establish a nonlinear mapping relationship between the measurable parameters and the internal hot spot temperature. After the model training is completed, the measurable parameters collected in real time on site are input into the trained RBF neural network, which can quickly output the contact and hot spot temperatures inside the switchgear that cannot be directly measured, realizing online and non-intrusive temperature inversion.

[0034] The electro-thermal-fluid coupled simulation and regression model first performs coupled calculations of electric field, temperature field and fluid field through finite element simulation to obtain the internal temperature distribution and hot spot data of the switch cabinet. Then, it uses regression algorithms such as support vector regression and neural network to fit the mapping relationship between measurable parameters and internal hot spot temperature, so as to quickly and accurately invert the contact and hot spot temperature that cannot be directly measured in the engineering site.

[0035] The finite element simulation process can be based on the actual structure and geometric parameters of the switchgear (such as the KYN series high-voltage switchgear) to build a three-dimensional geometric model, which is then imported into the finite element simulation software to obtain the simulation analysis model. The material properties of each component are set and physical fields (electric field, temperature field, fluid field) are added. The computational domain and boundary conditions of each physical field are set, and after meshing, the simulation solution (transient solution) can be performed to obtain the internal temperature distribution and hotspot data of the switchgear.

[0036] For humidity status data, when measurement points are unavailable, a fusion inversion method combining multi-physics constraints and physical information neural networks is used. The coupling laws of electric field, temperature field, flow field, and humidity field serve as physical constraints to invert the surface humidity, local condensation probability, humidity non-uniformity, and critical condensation duration of key insulating components, yielding a humidity status feature vector. When measurement points are available, based on the collected surface humidity and condensation status of key insulating components, as well as the relative humidity and humidity distribution parameters of the air inside the cabinet, the local condensation probability, humidity non-uniformity, and critical condensation duration are obtained. A multi-physics-constrained humidity inversion method based on PINN can be adopted. The coupled control equations of electric field, temperature field, flow field and humidity field and their boundary conditions are embedded as physical constraints into the neural network training process. Measurable load current, ambient temperature and humidity, and air temperature inside the cabinet are used as inputs, and state variables such as surface humidity of key insulating components and local condensation probability are used as outputs. By minimizing the weighted loss function of the network prediction value, control equation residual and boundary condition residual, the accurate inversion of humidity distribution and condensation state inside the switch cabinet under the condition of no humidity measurement points is achieved, thereby obtaining the humidity state feature vector.

[0037] For electrical operation data, operating condition adaptation and time-series fluctuation feature extraction are adopted to extract the effective value of load current, voltage distortion rate, power fluctuation coefficient, load rate time-series features, operating condition adaptation degree and electrical parameter mutation features to obtain electrical operation feature vector; The multi-source feature set is represented as: ; In the formula, Represents the set of multi-source features at sampling time t; , , , These represent the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector at sampling time t, respectively. When the dimensions of the source features are inconsistent, dimensionality coordination is first achieved through feature mapping or principal component transformation, followed by weighted concatenation or weighted fusion.

[0038] When performing feature extraction on mechanical condition data, the dynamic time-warped distance between the measured curve X and the standard curve Y is: ; In the formula, This represents the optimal time warping path from X to Y; K is the total number of matching points on the optimal warping path, and k is the index of the matching point; This represents the measured data point corresponding to the k-th matching point. Compared with standard data points The Euclidean distance between them; This means taking the minimum value among all possible matching paths; When performing feature extraction on measurable thermal state data, the contact temperature is used as the base value based on the inverted value. for: ; In the formula, This represents the m-th standardized input variable, which includes at least ambient temperature, load current, upper contact arm temperature, and busbar temperature, where M is the number of standardized input variables. For the corresponding The regression coefficients, This is the error term; When a contact temperature measuring device is installed in the switchgear, the measured contact temperature is used as a calibration sample. The calibration sample and the inversion value are fused according to the confidence coefficient c to obtain the equivalent temperature characteristics of the contact. ; In the formula, Indicates the contact equivalent temperature; Indicates calibration sample; c is determined based on the online status of the measuring points, sampling stability, and historical calibration errors. When no contact temperature measuring device is installed in the switch cabinet, or the calibration samples are unreliable, c=0. Unreliable calibration samples specifically refer to contact temperature sample data used for training or calibrating the temperature inversion model. Due to sensor failure, installation position deviation, calibration failure, environmental interference, or abnormal operating conditions, these samples cannot accurately reflect the actual temperature state of the contacts, resulting in a significant deviation between the samples and the true physical state, rendering them unworthy of reference for model training or calibration. c=0 is also set when measuring points are limited or offline.

[0039] To verify the effectiveness of the contact temperature equivalent inversion and credibility fusion mechanism, simulation data was used to generate contact temperature change scenarios. Figure 2 The blue curve represents the contact temperature inversion value obtained through an indirect inversion model based on ambient temperature, load current, upper contact arm temperature, and busbar temperature; the green scatter plot represents the calibration samples collected when simulating the configuration of the contact temperature measuring device, and its distribution simulates the situation where the online status of the measuring points is unstable and the sampling is intermittently missing in actual engineering; the red curve represents the equivalent contact temperature after credibility fusion, and the safe temperature threshold is an optional judgment criterion mentioned in the subsequent early warning judgment.

[0040] The verification results show that during the time period from t=0 to t=60, the calibration samples are sparse and discretely distributed, the confidence coefficient c approaches 0, and the contact equivalent temperature is... With inversion value The near-overlapping values ​​indicate that the system can degenerate into a pure inversion mode when no reliable measurement points are available, ensuring the continuity of temperature monitoring. After t=60, the calibration samples gradually become denser. exist The correction towards the calibration sample demonstrates the role of credibility fusion in improving inversion accuracy. This verification example shows that the contact temperature processing method in this embodiment can adaptively handle various engineering scenarios such as "with / without / intermittent measurement points," eliminating the rigid constraint of direct temperature measurement as a necessary input and improving the feasibility of the solution.

[0041] When extracting features from humidity state data, the loss function L of the physical information neural network is: ; In the formula, This represents the data error term between the model-predicted humidity value and the measured humidity data; Represents the physical constraint residuals of the coupled control equations of the electric field-temperature field-flow field-humidity field; , They are respectively , The weighting coefficients.

[0042] The residuals are obtained by substituting the humidity field distribution and its spatial gradient predicted by the physical information neural network into the coupled control equations of the electric field-temperature field-flow field-humidity field. These residuals include the residuals of the Poisson equation for the electric field, the residuals of the heat conduction equation, the residuals of the fluid continuity equation, the residuals of the humidity diffusion equation, and the residuals of the coupled boundary conditions. Each residual is obtained by automatically differentiating and calculating the spatial / temporal partial derivatives of the neural network output. Their weighted sum constitutes the physical constraint term, which is used to constrain the neural network prediction results to satisfy the physical laws of multi-physics coupling.

[0043] In S4, the combined weighting specifically involves obtaining the subjective weights of each eigenvector based on the analytic hierarchy process (AHP), obtaining the objective weights of each eigenvector based on the entropy weighting method, and finally obtaining the combined weights of all eigenvectors. ; In the formula, This represents the combined weight of the i-th eigenvector. ; This represents the subjective weight of the i-th eigenvector; Represents the objective weight of the i-th eigenvector; Indicates the balance coefficient; The comprehensive state feature vector at sampling time t Represented as: ; In the formula, , , , These represent the combined weights of the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector, respectively.

[0044] In S5, the comprehensive state feature vector is truncated using a preset time window length L to construct a time window sequence: ; In the formula, This represents a time window sequence at sampling time t; , These represent the combined state feature vectors at sampling times t-L+1 and t-L+2, respectively.

[0045] The four types of feature vectors—mechanical, thermal, humidity, and electrical—are all standardized vectors with fixed dimensions. Their dimensions are pre-defined by the feature extraction algorithm of the corresponding data source. The weighted feature vectors are concatenated in a preset order to form a comprehensive state feature vector in the form of a single vector. No additional normalization is required during the splicing process. Dimension alignment is achieved for different switch cabinet models and sensor configurations only through preset dimension filling or truncation processing (such as zero-value filling or feature truncation), ensuring that the vector dimensions of the input GRU model are fixed and uniform.

[0046] In S6, the fault prediction model adopts a multi-input single-output time series neural network model based on gated recurrent GRU units. During pre-training, a training set is constructed using historical operating data of the switchgear, including historical multi-source feature sequences and corresponding fault labels. The cross-entropy loss function is used for supervised training, and the model parameters are optimized through the backpropagation algorithm. When making online predictions, As input to the model, the GRU unit receives the comprehensive state feature vector at the current time step. Hidden state from the previous moment Update to hidden state The hidden states within the time window are aggregated, and the failure probability at a future preset time step is output through a fully connected layer. ; In the formula, Indicates a future preset time step The probability of failure, when When the value is 0, output the comprehensive fault risk value at the current moment. ; These are the output layer weights; For bias terms ( The baseline bias parameter is set to 0 and is jointly optimized with other model parameters during training; pool represents the pooling operation (average pooling). This represents the hidden state at sampling time t-L+1; To and The corresponding bias parameters are used to learn the risk evolution characteristics under different prediction step sizes; Sigmoid is the Sigmoid activation function. The output failure probability is used as the comprehensive failure risk value, and its value range is [0, 1].

[0047] When performing a future preset time step When predicting the probability of failure, a recursive autoregressive approach can be further adopted to achieve multi-step prediction: first, based on the historical true feature sequence at the current moment ( Predict the probability of failure at the next moment. After each prediction step, the predicted probability is converted into the corresponding state feature estimate through inverse mapping, and this estimate replaces the true feature in the input sequence, thus achieving closed-loop recursion until completion. Step prediction; at the same time, to control the accumulation of error during the recursive process, limits are imposed. The time step is set to a small value (e.g., 1-3 time steps), and a prediction confidence assessment mechanism is introduced. When the confidence of prediction results for multiple consecutive steps is lower than the preset threshold, the prediction is automatically terminated and the current result is used as an early warning reference to avoid error accumulation leading to prediction failure.

[0048] In S7, the joint risk assessment specifically refers to... Calculate the device health index at the current moment. Equipment degradation as well as and Consistency indicators between : ; ; ; In the formula, This represents the standard comprehensive state feature vector of the switchgear under healthy operating conditions; dist represents the Euclidean distance; clip represents the interval truncation function. This represents the sum of all comprehensive state feature vectors in historical operational data. The maximum Euclidean distance between them; This is a baseline feature vector formed by standardizing, extracting, and weighting data from four categories—mechanical, thermal, humidity, and electrical—under long-term, stable, and fault-free normal operating conditions of the switchgear. It reflects the normal correlation patterns and distribution characteristics among the various indicators of the switchgear, serving as a benchmark for switchgear health assessment and anomaly identification. To address environmental changes brought about by different seasons, separate settings can be configured for normal operating conditions in different seasons. .

[0049] Calculate the fusion risk value at the current moment. : ; In the formula, , They represent , Weighting coefficients; Joint determination based on hierarchical thresholds: when and ,or and At that time, a severe warning will be issued; when and ,or When, issue an early warning; when and ,or and When the conflict occurs, the output model conflict verification status is displayed, and the dominant abnormal data source causing the conflict (such as mechanical source, heat source, humidity source or electrical source) is also displayed, but it is not directly judged as a serious warning; otherwise, the normal status is displayed. in, , The fault probability threshold and , , The equipment degradation threshold and , , To integrate risk thresholds and , This is the consistency threshold; Similarly, for Calculate the future preset time step The risk value of fusion is determined and a joint judgment is made based on the hierarchical threshold.

[0050] when A significant decrease, but At lower levels, the system can also output model conflict verification status to verify sensors, models, and historical samples. Based on the collected data and inversion results, when the contact equivalent temperature exceeds the safe temperature threshold, the local condensation probability exceeds the insulation safety threshold, the mechanical jamming index exceeds the action reliability threshold, or the electrical protection index exceeds the safety threshold, the system directly outputs a severe warning. The mechanical jamming index is a key feature extracted from the collected data on opening and closing stroke, speed curve, coil current waveform, and action time, including opening and closing time deviation, stroke curve slope anomaly, action speed / acceleration deviation, and coil current peak deviation, reflecting the risk of mechanism jamming, wear, or component failure. The electrical protection index is a safety constraint index obtained from the collected load current, contact resistance, partial discharge, leakage current, and bus voltage data, including circuit resistance exceeding the limit, partial discharge exceeding the limit, leakage current exceeding the limit, and load current / short circuit current exceeding the limit, directly reflecting the electrical insulation and circuit safety status of the equipment.

[0051] When a model conflict is identified as being in a verification state, the primary abnormal data source causing the conflict is first located: the Mahalanobis distance between the feature vectors of each data source and the corresponding health baseline feature vectors (the feature vectors obtained when the switchgear is in a healthy operating state) is calculated, and the data source with the largest distance exceeding a preset threshold is selected as the primary abnormal data source; then, anomaly verification is performed on this data source, including checking the sensor data acquisition status, communication link integrity, data preprocessing correction results, and whether the multiphysics constraint relationships are abnormal, eliminating acquisition-related anomalies such as sensor failure, communication interference, and data correction failure; if the verification confirms that the data is true and valid, the model verification process is triggered, the failure probability or equipment degradation degree corresponding to the indicator is recalculated, and the consistency verification results are updated; if the conflict still exists after verification, a prompt of "abnormal status, on-site inspection recommended" is output to avoid misjudging it as a serious warning.

[0052] The fault probability threshold, equipment degradation threshold, fusion risk threshold, and consistency threshold are obtained through statistical analysis of historical equipment health data and fault sample data. Based on the distribution characteristics of indicators under normal operating conditions and fault conditions, the thresholds are determined using the quantile method or ROC curve method. The fault probability threshold and equipment degradation threshold correspond to the critical values ​​of fault probability and degradation under different risk levels, respectively. The fusion risk threshold corresponds to the graded critical value of the fusion risk indicator. The consistency threshold is the critical judgment value of the consistency indicator of fault probability and degradation. Each threshold can be adaptively adjusted according to on-site operation feedback.

[0053] For example, when maintenance personnel confirm on-site that an alert is a false alarm, the corresponding fault probability, degradation degree, and consistency indicators are added to the false alarm sample library. Periodically (e.g., quarterly), the ROC curve is recalculated based on the updated false alarm sample library and the historical fault sample library. The threshold point that significantly reduces the false alarm rate is selected as a new threshold candidate. If the deviation between the new and old thresholds exceeds a preset proportion (e.g., 10%), a threshold update process is triggered. This update takes effect after manual review, and the threshold change is recorded in the threshold change log. If the deviation does not exceed the preset proportion, the existing threshold remains unchanged to avoid frequent adjustments that could lead to instability in the alert strategy. Similarly, when on-site personnel confirm that an alert is a missed alarm, the corresponding indicators are added to the missed alarm sample library. Based on this library, the threshold is analyzed to determine whether it needs to be lowered to improve sensitivity. Through this closed-loop feedback mechanism, the threshold settings continuously adapt to the actual operating characteristics of the on-site equipment and maintenance experience.

[0054] The fault warning information includes risk level, comprehensive fault risk value, equipment health index, consistency index, dominant abnormal data source, source identifier of contact equivalent temperature, fault type, and maintenance recommendations. Fault types include at least abnormal mechanical characteristics of circuit breakers, contact overheating, abnormal temperature rise of hot spots inside the cabinet, risk of insulation dampness or condensation, and abnormal electrical operation. Maintenance recommendations are generated based on the risk level and dominant abnormal data source and are used to guide predictive maintenance of switchgear.

[0055] The specific method for generating maintenance recommendations is as follows: When the risk level is a severe warning, targeted maintenance recommendations are generated based on the dominant abnormal data sources (such as contact temperature, condensation status, mechanical action, and electrical parameters). For example, if the contact temperature exceeds the standard, it is recommended to shut down the power and inspect the contact status and heat dissipation channels; if the condensation probability exceeds the standard, it is recommended to check the temperature and humidity control device and insulation components inside the cabinet for moisture; if the mechanical jamming occurs, it is recommended to inspect the opening and closing mechanism and perform lubrication and maintenance; if the electrical protection indicators are abnormal, it is recommended to check the circuit insulation and protection device status. When the risk level is a warning, status tracking and periodic inspection recommendations are generated. When the model is in conflict verification status, maintenance recommendations for prioritizing the verification of sensors, communication links, and data acquisition equipment are generated, thereby achieving accurate predictive maintenance guidance based on risk and abnormal sources.

[0056] Example 2: A switchgear fault prediction device based on multi-source data fusion, comprising: One or more processors; Memory, used to store one or more computer programs; When one or more programs are executed by one or more processors, the one or more processors execute the method in Example 1.

[0057] Example 3: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method in Example 1.

[0058] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can make equivalent substitutions or modifications based on the technical solution and concept of the present invention within the scope of the technology disclosed in the present invention, and such modifications should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method for predicting switchgear faults based on multi-source data fusion, characterized in that... Includes the following steps: S1. Collect multi-source status data during the operation of the switchgear, including mechanical status data, measurable thermal status data, humidity status data, and electrical operation data; S2. Perform time synchronization, missing value processing, collection anomaly identification, and normalization on multi-source state data to obtain standardized data; S3. Based on standardized data, feature extraction is performed on mechanical condition data, thermal condition measurable data, humidity condition data and electrical operation data respectively to construct a multi-source feature set; S4. Based on the combined weighting method of the analytic hierarchy process and the entropy weighting method, the weight coefficients of various features in the multi-source feature set are obtained and fused to obtain the comprehensive state feature vector. S5. Slide and truncate the continuous comprehensive state feature vectors in time sequence with a preset time window length to construct a time window sequence; S6. Input the time window sequence into the pre-trained fault prediction model and output the comprehensive fault risk value of the switchgear at the current moment or at a preset future time step. S7. Calculate the equipment health index based on the comprehensive state feature vector, obtain the equipment degradation degree using the equipment health index, and combine the comprehensive fault risk value, equipment degradation degree and the consistency index of the two to make a joint risk judgment and output the fault warning information of the switchgear.

2. The switchgear fault prediction method based on multi-source data fusion according to claim 1, characterized in that, In S1, the mechanical status data includes the travel-time curve, opening and closing coil current curve, opening and closing time, contact movement speed, and mechanical vibration signal during the circuit breaker's opening and closing processes; the measurable thermal status data includes ambient temperature, busbar temperature, upper contact arm temperature, temperature distribution of hot spots in the cabinet, and load-related thermal condition data; the humidity status data includes the relative humidity of the air inside the cabinet, humidity distribution parameters, and, when measurement points are available, the surface humidity and condensation status of key insulating components; the electrical operation data includes load current, voltage, power, load rate, and operating condition parameters. Key insulation components include vacuum interrupter insulating bushings, post insulators, contact boxes, busbar insulating sheaths, insulating tie rods, wall bushings, and cable terminals.

3. The switchgear fault prediction method based on multi-source data fusion according to claim 1, characterized in that, In S2, the process of obtaining standardized data is as follows: Time synchronization: Using the high-precision clock of the unified data acquisition controller of the switch cabinet as a reference, data with different acquisition frequencies and different transmission delays are calibrated to the same time axis; Missing value handling: For data gaps, a weighted linear interpolation method based on a sliding time window is used to fill in the gaps, and the interpolation weight is determined by the time distance between the adjacent valid samples before and after the missing point. Anomaly identification: For outliers, the 3σ criterion, the characteristics of abrupt changes in temporal slope, and the consistency of multi-source physical correlation are used to distinguish them: when a single data source has an isolated outlier and does not meet the continuity of adjacent time or the correlation of other data sources, it is marked as an acquisition anomaly and corrected or removed; when multiple data sources change in correlation over time and meet the thermal, electrical, humidity or mechanical constraint relationship, they are retained as samples of potential fault precursors. Normalization: The maximum-minimum normalization method is used to map parameters of different dimensions to the interval [0, 1] to obtain standardized data.

4. The switchgear fault prediction method based on multi-source data fusion according to claim 2, characterized in that, In S3, for mechanical state data, dynamic time warping and curve morphology feature fusion extraction are adopted. Based on the standard opening and closing stroke-time curve and the standard coil current curve, the dynamic time warping distance between the measured curve and the standard curve is calculated, and the opening and closing time, overtravel, number of bounces, contact movement speed gradient, mechanical vibration peak value and frequency domain energy features are extracted to obtain the mechanical state feature vector. For measurable thermal state data, a combination of contact temperature equivalent inversion and thermodynamic feature extraction is adopted. An indirect contact temperature inversion model is constructed based on ambient temperature, load current, upper contact arm temperature and busbar temperature. The inverted contact temperature value is used as the basic output. When there are calibration samples with measured contact temperature, the inversion value is fused with credibility to obtain the contact equivalent temperature feature. Simultaneously, by using electro-thermal-fluid coupling simulation and regression model to invert the hot spot temperature distribution and temperature rise rate inside the cabinet, the characteristics of temperature gradient, hot spot offset and heat accumulation are extracted to obtain thermal state feature vector; For humidity status data, when measurement points are unavailable, a fusion inversion method combining multi-physics constraints and physical information neural networks is used. The coupling laws of electric field, temperature field, flow field, and humidity field serve as physical constraints to invert the surface humidity, local condensation probability, humidity non-uniformity, and critical condensation duration of key insulating components, yielding a humidity status feature vector. When measurement points are available, based on the collected surface humidity and condensation status of key insulating components, as well as the relative humidity and humidity distribution parameters of the air inside the cabinet, the local condensation probability, humidity non-uniformity, and critical condensation duration are obtained. For electrical operation data, operating condition adaptation and time-series fluctuation feature extraction are adopted to extract the effective value of load current, voltage distortion rate, power fluctuation coefficient, load rate time-series features, operating condition adaptation degree and electrical parameter mutation features to obtain electrical operation feature vector; The multi-source feature set is represented as: ； In the formula, Represents the set of multi-source features at sampling time t; , , , These represent the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector at sampling time t, respectively.

5. The switchgear fault prediction method based on multi-source data fusion according to claim 4, characterized in that, When performing feature extraction on mechanical condition data, the dynamic time-warped distance between the measured curve X and the standard curve Y is: ； In the formula, This represents the optimal time warping path from X to Y; K is the total number of matching points on the optimal warping path, and k is the index of the matching point; This represents the measured data point corresponding to the k-th matching point. Compared with standard data points The Euclidean distance between them; This means taking the minimum value among all possible matching paths; When performing feature extraction on measurable thermal state data, the contact temperature is used as the base value based on the inverted value. for: ； In the formula, This represents the m-th standardized input variable, including ambient temperature, load current, upper contact arm temperature, and busbar temperature, where M is the number of standardized input variables. For the corresponding The regression coefficients, This is the error term; When a contact temperature measuring device is installed in the switchgear, the measured contact temperature is used as a calibration sample. The calibration sample and the inversion value are fused according to the confidence coefficient c to obtain the equivalent temperature characteristics of the contact. ； In the formula, Indicates the contact equivalent temperature; Indicates calibration sample; When the switch cabinet is not equipped with a contact temperature measuring device, or the calibration sample is unreliable, c=0; When extracting features from humidity state data, the loss function L of the physical information neural network is: ； In the formula, This represents the data error term between the model-predicted humidity value and the measured humidity data; Represents the physical constraint residuals of the coupled control equations of the electric field-temperature field-flow field-humidity field; , They are respectively , The weighting coefficients.

6. The switchgear fault prediction method based on multi-source data fusion according to claim 4, characterized in that, In S4, the combined weighting specifically involves obtaining the subjective weights of each eigenvector based on the analytic hierarchy process (AHP), obtaining the objective weights of each eigenvector based on the entropy weighting method, and finally obtaining the comprehensive weight of each eigenvector. ； In the formula, This represents the combined weight of the i-th eigenvector. ; This represents the subjective weight of the i-th eigenvector; Represents the objective weight of the i-th eigenvector; Indicates the balance coefficient; The comprehensive state feature vector at sampling time t Represented as: ； In the formula, , , , These represent the combined weights of the mechanical state feature vector, thermal state feature vector, humidity state feature vector, and electrical operation feature vector, respectively.

7. The switchgear fault prediction method based on multi-source data fusion according to claim 6, characterized in that, In step S5, the comprehensive state feature vector is slidably truncated using a preset time window length L to construct a time window sequence: ； In the formula, This represents a time window sequence at sampling time t; , These represent the comprehensive state feature vectors at sampling times t-L+1 and t-L+2, respectively.

8. The switchgear fault prediction method based on multi-source data fusion according to claim 7, characterized in that, In S6, the fault prediction model adopts a multi-input single-output time series neural network model based on a gated recurrent GRU unit. During pre-training, a training set is constructed using historical operating data of the switchgear, including historical multi-source feature sequences and corresponding fault labels. The cross-entropy loss function is used for supervised training, and the model parameters are optimized through the backpropagation algorithm. When making online predictions, As input to the model, the GRU unit receives the comprehensive state feature vector at the current time step. Hidden state from the previous moment Update to hidden state The hidden states within the time window are aggregated, and the failure probability at a future preset time step is output through a fully connected layer. ； In the formula, Indicates a future preset time step The probability of failure, when When the value is 0, output the comprehensive fault risk value at the current moment. ; These are the output layer weights; For bias terms; pool indicates pooling operation; This represents the hidden state at sampling time t-L+1; To and The corresponding bias parameters; The output failure probability is used as the comprehensive failure risk value.

9. A switchgear fault prediction method based on multi-source data fusion according to claim 8, characterized in that, In S7, the joint risk determination specifically refers to, for Calculate the device health index at the current moment. Equipment degradation as well as and Consistency indicators between : ； ； ； In the formula, This represents the standard comprehensive state feature vector of the switchgear under healthy operating conditions; dist represents the Euclidean distance; clip represents the interval truncation function. This represents the sum of all comprehensive state feature vectors in historical operational data. The maximum Euclidean distance between them; Calculate the fusion risk value at the current moment. : ； In the formula, , They represent , Weighting coefficients; Joint determination based on hierarchical thresholds: when and ,or and At that time, a severe warning will be issued; when and ,or When, issue an early warning; when and ,or and When the conflict occurs, output the model conflict review status and the dominant abnormal data source causing the conflict; otherwise, output the normal status. in, , The fault probability threshold and , , The equipment degradation threshold and , , To integrate risk thresholds and , This is the consistency threshold; Similarly, for Calculate the future preset time step The risk value of fusion is determined and a joint judgment is made based on the hierarchical threshold.

10. A switchgear fault prediction method based on multi-source data fusion according to claim 9, characterized in that, In S7, the fault warning information includes risk level, comprehensive fault risk value, equipment health index, consistency index, dominant abnormal data source, source identifier of contact equivalent temperature, fault type, and maintenance suggestions. Fault types include abnormal mechanical characteristics of circuit breakers, contact overheating, abnormal temperature rise of hot spots in the cabinet, risk of insulation dampness or condensation, and abnormal electrical operation. Maintenance suggestions are generated based on risk level and dominant abnormal data source and are used to guide predictive maintenance of switchgear.

Images