Gas insulated switchgear fault diagnosis method and system based on multi-agent

Abstract

The application discloses a gas insulated switchgear fault diagnosis method and system based on multiple agents, and relates to the technical field of intelligent operation and maintenance of electric power equipment, and comprises the following steps: acquiring signal data of a target device, performing feature extraction on the signal data, and constructing a multi-modal feature matrix; extracting time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establishing a GIS propagation model, and solving the spatial coordinate position of the discharge source through a wave field inversion algorithm; combining the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, inputting the fusion feature vector into a dynamic Bayesian model, and outputting a fault type label and a corresponding confidence; migrating the dynamic Bayesian model based on a transfer learning mechanism, and dynamically updating a classification threshold; and inputting a diagnosis history sequence into a time series prediction model to predict a future operation state; through multi-modal fusion and intelligent reasoning, the application realizes accurate positioning and prediction of GIS faults, and solves the problems of low diagnosis precision and poor adaptability in the prior art.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent operation and maintenance of power equipment, in particular to a gas insulated switchgear fault diagnosis method and system based on multiple agents. BACKGROUND

[0002] As a key equipment in a high-voltage power transmission and transformation system, a gas insulated switchgear (GIS) is widely used in power grid substations and large industrial power systems, and has the advantages of compact structure, reliable operation and small land occupation. However, as the GIS equipment is operated in a high-voltage and high-pressure environment for a long time, the internal insulation medium and conductor components are easily affected by factors such as partial discharge, insulation aging, mechanical looseness and pollution, leading to potential insulation deterioration and discharge fault. Once a fault occurs, it may cause arc discharge or equipment explosion, resulting in serious power grid shutdown and safety accidents. Therefore, it is of great significance to carry out high-precision and real-time GIS fault diagnosis and prediction research to ensure the safe and stable operation of the power system.

[0003] For example, the invention patent with the publication number CN120046085B discloses a kind of mechanical and electrical equipment fault monitoring and early warning method, comprising: collecting strong current data and weak current data of subway mechanical and electrical equipment, respectively pre-processing and time synchronization to strong current data and weak current data, obtain strong current data sequence and weak current data sequence;Strong current data sequence and weak current data sequence are input into fault diagnosis model after attention cross fusion, and the confidence of fault type is obtained;The fault type is input into the pre-constructed fault propagation knowledge graph, and the potential propagation path of the fault is queried;Strong current data sequence, weak current data sequence, fault type and its confidence, fault potential propagation path are input into Bayesian network, and the fault propagation probability is dynamically calculated and early warning level is generated.

[0004] For example, the invention patent with the publication number CN120632643B discloses a kind of high-voltage cable insulation fault intelligent diagnosis method and system based on multi-information fusion, comprising the following steps: S1: obtaining the operation multi-source data of high-voltage cable, and constructing multi-source original data matrix;S2: pre-processing to obtain time-space aligned standardized data matrix;S3: feature extraction is carried out, multi-dimensional feature vector is obtained, and multi-modal deep network is used to learn the internal correlation between features to obtain joint multi-modal features;S4: based on Bayesian inference and Monte Carlo sampling, a fault type identification model is constructed, and a fault type classification result is obtained;S5: combining deep learning and physical model, the interval and positioning information of fault occurrence and fault level are obtained;S6: based on the fault type classification result and the interval and positioning information of fault occurrence, the fault level is obtained, and the diagnosis report is generated for early warning push. Realize high-precision identification, positioning and risk assessment of high-voltage cable insulation fault.

[0005] At least the following technical problems exist in the above disclosed technical solutions:

[0006] In the conventional fault diagnosis technology of gas insulated switchgear, the positioning of diagnosis mostly relies on a single signal source of acoustic wave or electromagnetic wave, and the positioning algorithm is mostly linear estimation, which is difficult to achieve accurate positioning in a complex structure and a multi-path propagation environment, and the conventional GIS fault diagnosis is mostly static classification, lacking the prediction ability of the evolution trend of the equipment state, and being unable to identify potential hidden dangers in advance.

[0007] In view of the above problems, the present application provides a solution. SUMMARY

[0008] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a gas insulated switchgear fault diagnosis method and system based on multiple agents, which realizes accurate positioning and prediction of GIS faults through multi-modal fusion and intelligent reasoning, and solves the problems of low diagnosis accuracy and poor adaptability of the conventional diagnosis.

[0009] To achieve the above object, the present application provides the following technical solutions:

[0010] The gas insulated switchgear fault diagnosis method based on multiple agents comprises: acquiring signal data of a target equipment, performing feature extraction on the signal data, and constructing a multi-modal feature matrix; extracting time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establishing a GIS propagation model, and solving the spatial coordinate position of the discharge source through a wave field inversion algorithm; combining the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, inputting the fusion feature vector into a dynamic Bayesian model, and outputting a fault type label and a corresponding confidence; migrating the dynamic Bayesian model based on a transfer learning mechanism, and dynamically updating a classification threshold; and inputting the fault type and the confidence result output by the migrated and optimized model and corresponding time sequence features into a diagnosis history sequence, and inputting the diagnosis history sequence into a time sequence prediction model to predict a future operation state.

[0011] In a preferred embodiment, the acquisition of the signal data of the target equipment is specifically as follows: according to the electric field distribution simulation result of the gas insulated switchgear, the positions of the sensors are optimized by using a genetic algorithm to form a multi-modal adaptive sensor network, and the signal data of the target equipment is collected; the collected signal data is corrected, the time delay difference between different signals is calculated by using a cross-correlation entropy algorithm, and the time is aligned; the intrinsic mode functions of each signal are decomposed by using an empirical mode decomposition method, and frequency domain matching is realized according to the coherent spectral density to obtain multi-modal signals with time-frequency dual-domain synchronization.

[0012] In a preferred embodiment, the feature extraction of the signal data and the construction of a multimodal feature matrix are specifically as follows: Feature extraction is performed on the multimodal signals, and the features of each type of signal are paired in the time dimension to form a time-step feature set, with feature vectors for each signal at each time step; based on the time-step feature set, a cross-modal graph structure is constructed, where each modal feature vector serves as a node, and weighted edges are constructed between different modalities based on feature similarity and mutual information to form a feature association graph; the feature association graph is input into a graph neural network model, and the weights between nodes are calculated through an attention mechanism, and nonlinear coupling and feature enhancement are performed on the multimodal features to obtain the fused feature matrix. Node feature representation: The fused node features are aggregated in the time dimension, forming a multi-time-step feature sequence through a sliding window, and a three-dimensional feature tensor is constructed in time-step order; the dimensionality of the three-dimensional feature tensor is reduced by the t-SNE algorithm to obtain a low-dimensional dense feature representation; a dynamic feature selection algorithm based on mutual information and the Relief-F evaluation function is used to perform sensitivity analysis and importance ranking on the low-dimensional dense feature representation, and the feature subset with the highest sensitivity to partial discharge is selected; the selected feature subset is organized into a feature vector stream according to the time series, and aggregated into a multimodal feature matrix that can be input into the subsequent diagnostic model through a sliding window.

[0013] In a preferred embodiment, the step of optimizing the sensor placement based on the electric field distribution simulation results of the gas-insulated switchgear to form a multimodal adaptive sensing network is as follows: A three-dimensional electric field simulation model is established using the finite element method or boundary element method based on the geometric structure and conductor layout of the gas-insulated switchgear. A rated voltage boundary condition is applied to the conductors to obtain the electric field intensity distribution inside the equipment. Gradient analysis is performed on the electric field intensity distribution results to calculate the rate of change of electric field intensity at each node in space. Threshold filtering and clustering algorithms are used to extract abrupt electric field changes and concentrated areas, forming a set of candidate sensor placement areas. With the overall coverage and response sensitivity of sensors in high electric field regions as the main optimization objectives, a multi-objective weighting function is constructed. A genetic algorithm is used for site optimization, encoding candidate sensor sites as binary genes to construct an initial population. Crossover and mutation probabilities are set to form an evolutionary search mechanism. In each generation of evolution, the fitness values ​​of each deployment scheme are output based on the sensor coverage algorithm. Selection, crossover, and mutation operations are sequentially performed based on the fitness values ​​to generate new schemes. The algorithm convergence is determined by the change in fitness. If the change in fitness values ​​is less than a set threshold or the maximum number of iterations is reached for several consecutive generations, the algorithm is considered to have converged, and the current deployment result is output.

[0014] In a preferred embodiment, the time delay features of acoustic and electromagnetic signals are extracted from the multi-modal feature matrix, and a GIS propagation model is established, as follows: the time-synchronized and spatially registered multi-modal feature matrix is obtained, and the original waveform data of the acoustic signal and electromagnetic signal corresponding to the time period are selected therefrom as the input for propagation modeling and time delay analysis; the generalized cross-correlation algorithm based on mutual information enhancement is used to calculate the time difference of arrival between acoustic signals, and the signal confidence weight is combined to suppress the influence of abnormal channels, thereby generating an acoustic time delay matrix; phase difference calculation based on joint time-frequency analysis is performed on electromagnetic signals, the envelope phase change trend is extracted through short-time Fourier transform, and the electromagnetic time delay matrix is formed through phase difference curve fitting; the acoustic and electromagnetic time delay matrices are fused based on a confidence adaptive weighting strategy, the modal weight is dynamically adjusted through a confidence weight function, and a multi-modal time delay feature set is generated; a GIS propagation model is established based on the three-dimensional structure model of the gas insulated switchgear and the gas medium distribution data; the measured temperature, pressure and humidity data are input into the GIS propagation model, and the sound speed and dielectric constant are dynamically corrected through the thermodynamic state equation.

[0015] In a preferred embodiment, the spatial coordinate position of the discharge source is solved by a wave field inversion algorithm, as follows: the multi-modal time delay feature set is input into the GIS propagation model, and the theoretical propagation time delay of the discharge source position is output according to the device structure and medium parameters; the difference between the actual observed time delay and the theoretical propagation time delay is used as the core to construct an objective function; the discharge source position is initialized as a multi-point candidate coordinate set inside the device, the objective function is scanned by a global search algorithm to determine a high confidence area; in the high confidence area, a sparse constraint TDOA inversion method is used to obtain the initial position of the discharge source with the minimum sparsity as the target; the initial position is input into the fine positioning stage, a nonlinear least squares inversion algorithm based on full waveform matching is used to compare the consistency of the phase difference and amplitude difference between the simulated waveform and the measured waveform, and the spatial coordinates of the discharge source are locally fine positioned; the residual mean square value of each sensing channel is dynamically output during the inversion iteration process, and the channel weight is updated according to the residual distribution; adaptive residual rejection is performed after each iteration, and the residual matrix is re-output; when the objective function converges and the residual change rate is lower than a threshold value, the inversion solution vector is obtained, which includes the spatial coordinate position of the discharge source and the corresponding energy data.

[0016] In a preferred embodiment, the spatial coordinate position is combined with the multi-modal feature matrix into a complete fusion feature vector, and input into a dynamic Bayesian model to output fault type label and corresponding confidence, as follows: time synchronization and scale normalization are performed on the spatial coordinate position, energy data of the discharge source, and features in the multi-modal feature matrix to form an initial fusion vector; based on the structure and operation mechanism of the gas insulated switchgear, a dynamic Bayesian model is constructed with fault type variable as core node and multi-modal observation features as visible node; the spatial coordinate of the discharge source and the energy data are mapped into spatial prior distribution, the probability correlation between the discharge source position and the fault type is established through kernel function, and the prior distribution is injected into the conditional probability table of the dynamic Bayesian model; in the prior stage of the model, different fault types are set according to historical operation data and expert rule base; the fusion feature vector is input into the dynamic Bayesian model, the likelihood function value of each node is calculated through forward propagation, and backward correction is performed under the constraint of spatial prior to form joint posterior distribution; the joint posterior distribution is marginalized on the spatial coordinate and discharge energy degree to obtain the marginal posterior probability of each fault type; the marginal posterior probability of each fault type is normalized to obtain the posterior probability distribution, and the fault type label is determined through the maximum posterior criterion; the confidence of the diagnosis result is calculated according to the entropy value of the posterior probability distribution.

[0017] In a preferred embodiment, the dynamic Bayesian model is migrated based on the transfer learning mechanism, and the classification threshold is dynamically updated, as follows: a set of pre-trained dynamic Bayesian model parameters on the GIS device is extracted as a source domain knowledge benchmark; multi-modal feature distribution alignment analysis is performed on the target GIS device, and MMD and KL divergence between the source domain and target domain multi-modal feature matrices are output to quantify the distribution difference degree of the two domains in the feature space; according to the distribution difference result, domain adaptation weight is applied to the dynamic Bayesian model parameters through the weighted Bayesian prior mechanism to form an initial prior set of the target device; the initial prior set is input to update the dynamic Bayesian model parameters through the variational Bayesian online learning algorithm, and the labeled samples collected at the initial stage of the target device are used as observation data to perform parameter posterior update and introduce a forgetting factor to dynamically attenuate the weight of historical samples; after updating the dynamic Bayesian model parameters, sparse constraint optimization is performed on the Bayesian network structure; according to the probability distribution output by the Bayesian model optimized by the sparse constraint, a classification threshold optimization objective function is constructed; the classification threshold optimization objective function is iteratively solved under different parameter weights through the NSGA-II algorithm to obtain a plurality of non-dominated threshold solution sets, and the global compromise threshold is selected as the current discriminant standard of the dynamic Bayesian model through the Pareto optimality criterion.

[0018] In a preferred embodiment, the fault type and confidence result output by the migration-optimized model are combined with the corresponding time sequence features to form a diagnostic history sequence, which is input into a time sequence prediction model to predict the future operating state, specifically as follows: the migration-optimized dynamic Bayesian model is continuously run on the target GIS device, and real-time output of the fault type label and the corresponding confidence value at each time step is obtained, and the multi-modal feature matrix at that time is recorded synchronously and integrated by time step to form a diagnostic history sequence; the diagnostic history sequence is input into the time sequence prediction model, the time sequence prediction model learns the time sequence evolution rule of the diagnostic state, and outputs a predicted state vector of future multiple time steps; in the prediction phase, the time sequence prediction model generates a probability distribution of the future state according to the current diagnostic state and the historical window data, and outputs a confidence interval; the prediction result is subjected to abnormal trend detection, and potential abnormal evolution signals are identified according to the confidence interval drift rate, and when the drift rate exceeds a set threshold, the system is marked as a potential fault trend and an early warning is issued.

[0019] The system of the information equipment full life cycle digital management method comprises a data acquisition module, a solving module, a fusion module, an updating module and a prediction module, and there is a connection between the modules; the data acquisition module is used for acquiring signal data of a target device, performing feature extraction on the signal data, and constructing a multi-modal feature matrix; the solving module is used for extracting time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establishing a GIS propagation model, and solving the spatial coordinate position of the discharge source through a wave field inversion algorithm; the fusion module is used for combining the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, and inputting the fusion feature vector into a dynamic Bayesian model to output a fault type label and a corresponding confidence; the updating module is used for migrating the dynamic Bayesian model based on a transfer learning mechanism to dynamically update a classification threshold; and the prediction module is used for combining the fault type and the confidence result output by the migration-optimized model with the corresponding time sequence features to form a diagnostic history sequence, inputting the diagnostic history sequence into a time sequence prediction model, and predicting a future operating state.

[0020] The gas insulated switchgear fault diagnosis method and system based on multiple agents have the following technical effects and advantages:

[0021] 1. The application establishes a multi-modal adaptive sensing network through a multi-agent collaborative perception mechanism, optimizes the sensor distribution position based on the electric field distribution simulation results of the gas insulated switchgear by using a genetic algorithm, so that the sensing nodes realize optimal coverage of the high electric field area in space. This scheme can adaptively adjust the sensor distribution density and distribution form, improve the response sensitivity and spatial resolution to complex internal electric field changes, and improve the representativeness and reliability of signal acquisition from the source.

[0022] 2. The application introduces a GIS propagation model based on time delay characteristics, forms a multi-modal time delay feature set by fusing the time delays of acoustic and electromagnetic signals, and accurately solves the discharge source spatial coordinates by combining the wave field inversion algorithm of full waveform matching. This method not only overcomes the problem of large noise interference in single modal positioning, but also effectively improves the positioning accuracy and stability through sparse constraint inversion and residual adaptive weight adjustment mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of the gas insulated switchgear fault diagnosis method based on multiple agents of the application.

[0024] Figure 2 A system structure diagram of the gas insulated switchgear fault diagnosis method based on multiple agents of the application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0026] Embodiment 1, Figure 1 The gas insulated switchgear fault diagnosis method based on multiple agents of the application is given, which comprises:

[0027] S1, acquiring signal data of a target device and performing feature extraction on the signal data to construct a multi-modal feature matrix;

[0028] In this embodiment, the signal data of the target device is acquired, the signal data is subjected to feature extraction, and a multi-modal feature matrix is constructed, specifically as follows:

[0029] According to the electric field distribution simulation results of the gas insulated switchgear, the positions of the UHF, acoustic and gas sensors are optimized by using a genetic algorithm to form a multi-modal adaptive sensor network, and the signal data of the target device is collected, wherein the signal data comprises acoustic signals, UHF signals, gas signals, temperature and mechanical vibration states;

[0030] The multi-modal adaptive sensor network performs an online self-calibration process before sampling, estimates the noise power spectral density of the UHF channel and corrects the signal-to-noise ratio by using an adaptive Kalman filter, compensates for time drift and dynamically equalizes the gain of the acoustic channel, performs temperature and humidity coupling compensation on the gas array channel, and performs baseline fitting and residual constraint on the temperature and vibration channels;

[0031] The collected signal data is band-pass filtered, denoised and baseline corrected, the time delay difference between different signals is calculated by cross-correlation entropy algorithm and time alignment is performed;

[0032] The intrinsic mode functions of each signal are decomposed by empirical mode decomposition method, and frequency domain matching is realized according to coherent spectral density, so as to obtain multi-modal signals synchronized in time-frequency dual domain;

[0033] The multi-modal signals are extracted, including:

[0034] The transient energy spectral density, phase envelope and pulse peak rate of the UHF signal are extracted by wavelet packet and Hilbert transform;

[0035] The short-time Fourier transform is performed on the acoustic signal, and the mel cepstral coefficient is output;

[0036] The nonlinear concentration features of the gas signal are extracted by sparse autoencoder;

[0037] The multi-scale entropy index is calculated for the temperature and vibration signals;

[0038] The features of each type of signal are correspondingly paired in the time dimension to form a time step feature set, each time step containing feature vectors of electromagnetic, acoustic, gas, temperature and vibration signals;

[0039] According to the time step feature set, a cross-modal graph structure is constructed, in which each modal feature vector is a node, and a weighted edge is constructed between different modalities according to feature similarity and mutual information to form a feature correlation graph;

[0040] The feature correlation graph is input into a graph neural network model, the weights between nodes are calculated through an attention mechanism, and the multi-modal features are nonlinearly coupled and enhanced to obtain the fused node feature representation;

[0041] The fused node features are aggregated in the time dimension, a multi-time step feature sequence is formed through a sliding window method, and a three-dimensional feature tensor is constructed in time steps, with the tensor dimensions including time, modal and feature dimensions;

[0042] The three-dimensional feature tensor is reduced by t-SNE algorithm to obtain a low-dimensional dense feature representation;

[0043] The low-dimensional dense feature representation is analyzed for sensitivity and importance ranking by a dynamic feature selection algorithm based on mutual information and Relief-F evaluation function, and a feature subset with the highest sensitivity to partial discharge is selected;

[0044] The selected feature subset is organized into a feature vector stream in time sequence, and is aggregated by a sliding window into a multi-modal feature matrix that can be input into a subsequent diagnostic model.

[0045] In this embodiment, according to the electric field distribution simulation results of the gas insulated switchgear, the positions of the UHF, acoustic and gas sensors are optimized by using the genetic algorithm to form a multi-modal adaptive sensor network, as follows:

[0046] According to the geometric structure and conductor layout of the gas insulated switchgear, a three-dimensional electric field simulation model is established by using the finite element method or the boundary element method, and the conductors are loaded with rated voltage boundary conditions, and the electric field intensity distribution in the device is solved;

[0047] The gradient analysis is performed on the electric field intensity distribution results, the electric field intensity change rate of each node in the space is calculated, and the electric field mutation area and the concentrated area are extracted by threshold screening and clustering algorithm to form a candidate sensor layout area set;

[0048] For UHF, acoustic and gas sensors, the corresponding perception coverage models are established, wherein the effective monitoring range of the UHF sensor is determined according to the electromagnetic wave propagation attenuation characteristics, the sensitive interval of the acoustic sensor is determined according to the sound pressure attenuation and reflection characteristics, and the detection radius of the gas sensor is determined according to the diffusion model;

[0049] Taking the comprehensive coverage and response sensitivity of the sensors to the high electric field area as the main optimization target, a multi-objective weighted function is constructed;

[0050] The genetic algorithm is used for point layout optimization, the sensor candidate layout points are coded into binary genes, an initial population is constructed, and the crossover probability and mutation probability are set to form an evolutionary search mechanism;

[0051] In each generation evolution, the fitness value of each layout scheme is output based on the sensor coverage algorithm, and the selection, crossover and mutation operations are sequentially executed according to the fitness value to generate new schemes, and the convergence of the algorithm is judged by the fitness change;

[0052] If the fitness value changes less than a set threshold value for a plurality of generations or reaches the maximum number of iterations, it is determined that the algorithm converges, and the current layout result is output to obtain the optimal combination scheme of the types, spatial coordinates and installation directions of the UHF, acoustic and gas sensors, and realize the adaptive optimal coverage of the multi-modal monitoring area.

[0053] S2, the time delay features of acoustic and electromagnetic signals are extracted from the multi-modal feature matrix, a GIS propagation model is established, and the spatial coordinate position of the discharge source is solved by using the wave field inversion algorithm;

[0054] In this embodiment, the time delay features of acoustic and electromagnetic signals are extracted from the multi-modal feature matrix, and a GIS propagation model is established, as follows:

[0055] The time-synchronized and spatially registered multi-modal feature matrix is obtained, from which the original waveform data of the acoustic signal and the electromagnetic signal corresponding to the time period are selected as the input of the propagation modeling and time delay analysis;

[0056] Adaptive sparse reconstruction denoising is performed on the acoustic and electromagnetic signals respectively, the multi-scale wavelet packet denoising is used for the acoustic part, and the compressed sensing reconstruction is used for the electromagnetic part, so as to maintain the high-frequency details and eliminate the background noise interference;

[0057] The generalized cross-correlation algorithm based on mutual information enhancement is used to calculate the time difference of arrival between the acoustic signals, and the signal confidence weight is used to suppress the influence of abnormal channels, thereby generating an acoustic time delay matrix;

[0058] The phase difference calculation based on time-frequency joint analysis is performed on the electromagnetic signal, the envelope phase change trend is extracted through the short-time Fourier transform (STFT), and the high-precision electromagnetic time delay matrix is formed through the phase difference curve fitting;

[0059] The acoustic and electromagnetic time delay matrices are fused based on the confidence adaptive weighting strategy, the modal weight is dynamically adjusted through the confidence weight function, and a unified multi-modal time delay feature set is generated;

[0060] The GIS propagation model is established based on the three-dimensional structure model and the gas medium distribution data of the gas insulated switchgear, the sound field distribution is solved by the finite element method, and the finite difference time domain (FDTD) method is used to simulate the electromagnetic wave propagation path, thereby realizing the coupling modeling of the acoustic and electric multi-physical fields;

[0061] The measured temperature, pressure and humidity data are input into the GIS propagation model, and the sound speed and dielectric constant are dynamically corrected through the thermodynamic state equation.

[0062] In this embodiment, the spatial coordinate position of the discharge source is solved by the wave field inversion algorithm, and the specific process is as follows:

[0063] The multi-modal time delay feature set is input into the GIS propagation model, and the theoretical propagation time delay of the discharge source position is output according to the device structure and medium parameters;

[0064] The difference between the actual observation time delay and the theoretical propagation time delay is used as the core to construct the objective function, wherein the objective function includes a time delay residual term, an energy constraint term and a smoothing regularization term, which are used to constrain the discharge source position and the discharge energy parameter at the same time;

[0065] The discharge source position is initialized as a multi-point candidate coordinate set inside the device, the global search algorithm (such as particle swarm or differential evolution) is used to preliminarily scan the objective function, and the high-confidence area is determined to avoid falling into local optimum;

[0066] In the high-confidence area, the initial position of the discharge source is obtained by the TDOA inversion method with sparse constraints, and the minimization of sparsity is taken as the goal, so as to realize the rapid screening of a large area;

[0067] The initial position is taken as the input to enter the fine positioning stage, and the nonlinear least square inversion algorithm based on full waveform matching is adopted to realize the local fine positioning of the spatial coordinates of the discharge source by comparing the consistency of the phase difference and amplitude difference between the simulated waveform and the measured waveform;

[0068] In the inversion iteration process, the residual mean square value of each sensing channel is dynamically output, and the channel weight is updated according to the residual distribution, so that the weight of the channel with larger residual is automatically reduced in the subsequent iteration, thereby improving the convergence stability;

[0069] After each iteration, adaptive residual rejection is performed to reject abnormal samples deviating from the median by more than a set multiple, and the residual matrix is recalculated to ensure the robustness of the objective function;

[0070] When the objective function converges or the residual change rate is lower than the threshold, the inversion solution vector is obtained, which includes the spatial coordinate position of the discharge source and the corresponding energy data.

[0071] The objective function taking the difference between the actual observation time delay and the theoretical propagation time delay as the core is as follows:

[0072]

[0073] In the formula: is the objective function, indicating the comprehensive inversion error measurement function, which is used to constrain the spatial position and discharge energy of the discharge source, is the spatial coordinate of the discharge source, is the discharge energy, is the mode set, is the mode index, is the number of sensors, is the sensor number index, is the weight of the time delay residual term, indicating the relative weight of the time delay residual in the objective function, which is used to adjust the contribution of each channel to the inversion, is the actual observed signal arrival time, indicating the signal arrival time observed by mode m at sensor r, is the theoretical propagation time delay, is the energy item weight coefficient, which is used to adjust the influence strength of the energy constraint item in the overall objective function, and the value range is usually , is the weight of the energy constraint item, indicates the observation amplitude of sensor r to mode m signal, is the residual loss function, is a transmission gain function, representing signal attenuation and spatial propagation characteristics from discharge source to sensor r, is a regularization term weight coefficient, used to control the constraint strength of the smoothing regularization term, to prevent overfitting and numerical instability, is a smoothing regularization term, representing prior constraints of discharge source location and energy or spatial smoothing constraints, is the propagation speed of mode m.

[0074] S3, combine the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, and input it into the dynamic Bayesian model to output the fault type label and the corresponding confidence;

[0075] In this embodiment, the spatial coordinate position and the multi-modal feature matrix are combined into a complete fusion feature vector, and input into the dynamic Bayesian model to output the fault type label and the corresponding confidence, as follows:

[0076] The spatial coordinate position, energy data of the discharge source and the features in the multi-modal feature matrix are time-synchronized and scaled to form an initial fusion vector containing spatial, energy and multi-source signal information;

[0077] Based on the structure and operation mechanism of gas insulated switchgear, a dynamic Bayesian model is constructed with fault type variable as core node and multi-modal observation features as visible node;

[0078] The discharge source spatial coordinate and energy data are mapped to spatial prior distribution, the probability association between discharge source location and fault type is established through kernel function, and the prior distribution is injected into the conditional probability table of the dynamic Bayesian model;

[0079] In the prior stage of the model, different fault types (such as partial discharge, poor grounding, insulation breakdown, metal particle ionization, etc.) are set according to historical operation data and expert rule base;

[0080] The fusion feature vector is input into the dynamic Bayesian model, the likelihood function value of each node is calculated through forward propagation, and backward correction is performed under the constraint of spatial prior to form joint posterior distribution;

[0081] The joint posterior distribution is marginalized on the spatial coordinate and discharge energy scale to obtain the marginal posterior probability of each fault type;

[0082] The marginal posterior probability of each fault type is normalized to obtain the posterior probability distribution, and the fault type label is determined through the maximum a posteriori criterion (MAP);

[0083] The confidence of the diagnosis result is calculated based on the entropy value of the posterior probability distribution, and a binary output result containing “fault type-confidence” is generated.

[0084] The joint posterior distribution is marginalized on the spatial coordinates and discharge energy, as follows:

[0085]

[0086] wherein: is the marginal posterior probability, is a discrete variable of the i-th fault type, represents the observed fusion feature vector, including UHF features, acoustic features, gas components, temperature / vibration features, is the spatial coordinate of the discharge source, is the discharge energy, represents the observation and the discharge source is located at , with energy , the conditional probability of the occurrence of the i-th fault.

[0087] S4, migrating the dynamic Bayesian model based on a transfer learning mechanism to dynamically update the classification threshold;

[0088] In this embodiment, the dynamic Bayesian model is migrated based on a transfer learning mechanism to dynamically update the classification threshold, as follows:

[0089] Extract the pre-trained dynamic Bayesian model parameter set on the GIS (Gas Insulated Switchgear), which includes the structural topology, conditional probability distribution, prior weight, and fault type threshold mapping function, as the source domain knowledge benchmark;

[0090] Perform multi-modal feature distribution alignment analysis on the target GIS device, output the MMD (Maximum Mean Discrepancy) and KL divergence between the source domain and target domain multi-modal feature matrices, to quantify the distribution difference degree of the two domains in the feature space, and provide a basis for transfer strategy selection;

[0091] According to the distribution difference result, apply domain adaptation weight to the dynamic Bayesian model parameters through a weighted Bayesian prior mechanism to form the initial prior set of the target device, wherein a lower transfer weight is given to high offset feature dimensions, and a high weight is maintained for stable feature dimensions, thereby balancing between knowledge transfer and domain adaptation;

[0092] Take the initial prior set as input, update the dynamic Bayesian model parameters through a variational Bayesian online learning algorithm, take the labeled samples collected by the target device initially as observation data, perform parameter posterior update, and introduce a forgetting factor to dynamically attenuate the weight of historical samples;

[0093] After the dynamic Bayesian model parameter update, sparse constraint optimization is performed on the Bayesian network structure, the network connection is dynamically adjusted according to the mutual information gain between feature nodes, redundant dependency relationships are deleted, key causal links are retained, and the model structure stability and interpretability are improved.

[0094] According to the Bayesian model output probability distribution of the sparse constraint optimization, a classification threshold optimization objective function is constructed.

[0095] The NSGA-II algorithm is used to iteratively solve the classification threshold optimization objective function under different parameter weights, and a plurality of sets of non-dominated threshold solutions are obtained. The global compromise threshold is selected by the Pareto optimality criterion as the current discriminant standard of the dynamic Bayesian model.

[0096] During the operation of the device, the model performance change rate and the confidence entropy are monitored in real time. When the performance decreases or the output entropy increases to a set threshold, the migration update module is automatically triggered to perform local parameter fine tuning to offset the influence of domain drift

[0097] During the migration update process, the drift factor between the posterior distributions of consecutive time periods is calculated. If the drift exceeds a threshold, low-rank matrix constraint is used to adjust the high-sensitive parameters, so as to keep the main structure of the model unchanged and realize rapid self-adaptation.

[0098] The updated model re-outputs the posterior distribution for the latest feature input, and calculates the trend variance in different time windows. The confidence regression prediction threshold drift direction is adjusted in advance to adjust the classification threshold of the next stage.

[0099] S5, the fault type and confidence result output by the migration optimized model and the corresponding time sequence features are combined to form a diagnostic history sequence, which is input into a time sequence prediction model to predict the future running state.

[0100] In this embodiment, the fault type and confidence result output by the migration optimized model and the corresponding time sequence features are combined to form a diagnostic history sequence, which is input into a time sequence prediction model to predict the future running state, as follows:

[0101] The migration optimized dynamic Bayesian model is continuously run on the target GIS device, and real-time output of each time step of the fault type label and the corresponding confidence value is obtained. The multi-modal feature matrix at that time is recorded synchronously, and is integrated according to the time step to form a diagnostic history sequence.

[0102] The diagnostic history sequence is input into the time sequence prediction model, the time sequence prediction model learns the time sequence evolution rule of the diagnostic state, and outputs a prediction state vector of a plurality of future time steps.

[0103] In the prediction stage, the time series prediction model generates a probability distribution of the future state according to the diagnostic state of the current moment and the historical window data, and outputs a confidence interval at the same time;

[0104] An abnormal trend detection is performed on the prediction result, potential abnormal evolution signals are identified according to a confidence interval drift rate, and when the drift rate exceeds a set threshold, the system is marked as a potential fault trend and an early warning is issued.

[0105] Embodiment 2, Figure 2 The system of the gas insulated switchgear fault diagnosis method based on the multi-agent is given, which comprises a data acquisition module, a solving module, a fusion module, an updating module and a prediction module, and there is a connection between the modules;

[0106] The data acquisition module is used for acquiring signal data of a target device, performing feature extraction on the signal data, and constructing a multi-modal feature matrix.

[0107] The solving module is used for extracting time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establishing a GIS propagation model, and solving the spatial coordinate position of the discharge source through a wave field inversion algorithm.

[0108] The fusion module is used for combining the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, inputting the fusion feature vector into a dynamic Bayesian model, and outputting a fault type label and a corresponding confidence.

[0109] The updating module is used for migrating the dynamic Bayesian model based on a transfer learning mechanism, and dynamically updating a classification threshold.

[0110] The prediction module is used for inputting a fault type and a confidence result output by the model after migration and optimization, and corresponding time series features into a diagnostic history sequence, inputting the diagnostic history sequence into a time series prediction model, and predicting a future operation state.

[0111] The above formulas are all dimensionless numerical calculations, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0112] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part.

[0113] Those skilled in the art can understand that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0114] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0115] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0116] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for diagnosing a fault of a gas-insulated switchgear based on multi-agent, characterized in that, The method comprises the following steps: acquiring signal data of a target device and performing feature extraction on the signal data to construct a multi-modal feature matrix; extracting time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establishing a GIS propagation model, and solving the spatial coordinate position of the discharge source through a wave field inversion algorithm; combining the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, inputting the fusion feature vector into a dynamic Bayesian model, and outputting a fault type label and a corresponding confidence level; transferring the dynamic Bayesian model based on a transfer learning mechanism to dynamically update the classification threshold; inputting the fault type and confidence level output by the transferred and optimized model and the corresponding time sequence features into a diagnostic history sequence, and inputting the diagnostic history sequence into a time sequence prediction model to predict a future operation state; The signal data of the target device is acquired in the following way: According to the electric field distribution simulation results of the gas insulated switchgear, the genetic algorithm is used to optimize the sensor layout position to form a multi-modal adaptive sensor network, and the signal data of the target device is collected; The collected signal data is corrected, the time delay difference between different signals is calculated by the cross-correlation entropy algorithm, and the time is aligned; The intrinsic mode function of each signal is decomposed by the empirical mode decomposition method, and the frequency domain matching is realized according to the coherent spectral density to obtain the multi-modal signal with time-frequency dual-domain synchronization.

2. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 1, characterized in that, The signal data is feature extracted and a multi-modal feature matrix is constructed in the following way: The multi-modal signal is feature extracted, and the features of each type of signal are correspondingly paired in the time dimension to form a time step feature set, and the feature vector of each signal at each time step; According to the time step feature set, a cross-modal graph structure is constructed, wherein each modal feature vector is taken as a node, and a weighted edge is constructed between different modal feature vectors according to the feature similarity and mutual information to form a feature correlation graph; The feature correlation graph is input into a graph neural network model, the weight between nodes is calculated through an attention mechanism, and the multi-modal features are nonlinearly coupled and enhanced to obtain the fused node feature representation; The fused node features are aggregated in the time dimension, a multi-time step feature sequence is formed through a sliding window method, and a three-dimensional feature tensor is constructed in the time step sequence; The three-dimensional feature tensor is reduced in dimension by the t-SNE algorithm to obtain a low-dimensional dense feature representation; The low-dimensional dense feature representation is analyzed for sensitivity and importance by a dynamic feature selection algorithm based on mutual information and a Relief-F evaluation function, and a feature subset with the highest sensitivity to partial discharge is screened out; The screened feature subset is organized into a feature vector stream in the time sequence, and is aggregated into a multi-modal feature matrix that can be input into a subsequent diagnostic model through a sliding window.

3. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 2, characterized in that, According to the electric field distribution simulation results of the gas insulated switchgear, the genetic algorithm is used to optimize the sensor layout position to form a multi-modal adaptive sensor network, and the signal data of the target device is collected; According to the geometric structure and conductor layout of the gas insulated switchgear, a three-dimensional electric field simulation model is established by using the finite element method or the boundary element method, and the conductor is loaded with a rated voltage boundary condition to solve the internal electric field intensity distribution of the device; Gradient analysis is performed on the electric field intensity distribution results to calculate the electric field intensity variation rate of each node in the space, and the electric field mutation area and concentration area are extracted through threshold screening and clustering algorithm to form a candidate sensor layout area set; Taking the comprehensive coverage degree and response sensitivity of the sensor to the high electric field area as the main optimization target, a multi-objective weighted function is constructed; The genetic algorithm is used for point layout optimization, the sensor candidate layout points are coded into binary genes, the initial population is constructed, and the crossover probability and mutation probability are set to form an evolutionary search mechanism; In each generation evolution, the fitness value of each layout scheme is output based on the sensor coverage algorithm, and the selection, crossover and mutation operations are sequentially executed according to the fitness value to generate a new scheme, and the convergence of the algorithm is judged through the change of the fitness value; If the fitness value changes less than a set threshold for several generations or reaches the maximum iteration number, the algorithm is determined to be converged, and the current layout result is output.

4. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 3, characterized by, The time delay features of acoustic and electromagnetic signals are extracted from the multi-modal feature matrix, and the GIS propagation model is established, which is as follows: The time-synchronized and spatially registered multi-modal feature matrix is obtained, and the original waveform data of the corresponding time period of the acoustic signal and the electromagnetic signal are selected as the input of the propagation modeling and time delay analysis; The generalized cross-correlation algorithm based on mutual information enhancement is used to calculate the time difference of arrival of acoustic signals, and the signal confidence weight is combined to suppress the influence of abnormal channels, and an acoustic time delay matrix is generated; The phase difference calculation based on time-frequency joint analysis is performed on the electromagnetic signal, the envelope phase change trend is extracted through short-time Fourier transform, and the electromagnetic time delay matrix is formed through phase difference curve fitting; The acoustic and electromagnetic time delay matrices are fused based on the confidence adaptive weighting strategy, the modal weight is dynamically adjusted through the confidence weight function, and a multi-modal time delay feature set is generated; The GIS propagation model is established based on the three-dimensional structure model and the gas medium distribution data of the gas insulated switchgear; The measured temperature, pressure and humidity data are input into the GIS propagation model, and the sound speed and dielectric constant are dynamically corrected through the thermodynamic state equation.

5. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 4, characterized in that, The spatial coordinate position of the discharge source is solved by the wave field inversion algorithm, which is as follows: The multi-modal time delay feature set is input into the GIS propagation model, and the theoretical propagation time delay of the discharge source position is output according to the device structure and medium parameters; The difference between the actual observation time delay and the theoretical propagation time delay is used as the core to construct the objective function; The discharge source position is initialized as a multi-point candidate coordinate set inside the device, and the global search algorithm is used to scan the objective function to determine the high confidence area; In the high confidence area, the initial position of the discharge source is obtained by the sparse constraint TDOA inversion method with the minimum sparsity as the target; The initial position is input into the fine positioning stage, and the nonlinear least squares inversion algorithm based on full waveform matching is used to compare the consistency of the phase difference and amplitude difference between the simulated waveform and the measured waveform to perform local fine positioning on the spatial coordinates of the discharge source; The residual mean square value of each sensing channel is dynamically output during the inversion iteration process, and the channel weight is updated according to the residual distribution; Self-adaptive residual rejection is performed after each iteration, and the residual matrix is output again; When the objective function converges and the residual rate of change is lower than a threshold value, an inversion solution vector is obtained, the inversion solution vector including spatial coordinate positions of the discharge source and corresponding energy data.

6. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 5, characterized in that, The spatial coordinate positions and the multi-modal feature matrix are combined into a complete fusion feature vector, and the complete fusion feature vector is input into a dynamic Bayesian model to output a fault type label and a corresponding confidence. The spatial coordinate positions of the discharge source, the energy data, and the features in the multi-modal feature matrix are time-synchronized and scale-normalized to form an initial fusion vector. Based on the structure and operation mechanism of the gas insulated switchgear, a dynamic Bayesian model is constructed with a fault type variable as a core node and multi-modal observation features as visible nodes. The spatial coordinate positions of the discharge source and the energy data are mapped into a spatial prior distribution, a probability correlation between the discharge source positions and the fault types is established through a kernel function, and the prior distribution is injected into a conditional probability table of the dynamic Bayesian model. In the prior stage of the model, different fault types are set according to historical operation data and an expert rule base. The fusion feature vector is input into the dynamic Bayesian model, a likelihood function value of each node is calculated through forward propagation, and a joint posterior distribution is formed through backward correction under the constraint of the spatial prior. The joint posterior distribution is marginalized on the spatial coordinate and the discharge energy degree to obtain an edge posterior probability of each fault type. The edge posterior probability of each fault type is normalized to obtain a posterior probability distribution, and a fault type label is determined through a maximum posterior criterion. The confidence of the diagnosis result is calculated according to the entropy value of the posterior probability distribution.

7. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 6, characterized in that, The dynamic Bayesian model is migrated based on a transfer learning mechanism, and a classification threshold value is dynamically updated as follows. A set of pre-trained dynamic Bayesian model parameters on the body insulation switchgear is extracted as a source domain knowledge benchmark. A multi-modal feature distribution alignment analysis is performed on the target GIS device, and MMD and KL divergence between the source domain and the target domain multi-modal feature matrices are output to quantify the distribution difference degree of the two domains in the feature space. According to the distribution difference result, a domain adaptation weight is applied to the dynamic Bayesian model parameters through a weighted Bayesian prior mechanism to form an initial prior set of the target device. The initial prior set is input, the dynamic Bayesian model parameters are updated through a variational Bayesian online learning algorithm, the labeled samples collected at the initial stage of the target device are used as observation data, the parameter posterior update is performed, and a forgetting factor is introduced to dynamically attenuate the weight of the historical samples. After the dynamic Bayesian model parameters are updated, a sparse constraint optimization is performed on the Bayesian network structure. According to the probability distribution output by the Bayesian model optimized by the sparse constraint, a classification threshold value optimization objective function is constructed. The classification threshold value optimization objective function is iteratively solved under different parameter weights through the NSGA-II algorithm to obtain a plurality of non-dominated threshold solution sets, and a global compromise threshold value is selected as the current discriminant standard of the dynamic Bayesian model through the Pareto optimality criterion.

8. The multi-agent based gas insulated switchgear fault diagnosis method according to claim 7, characterized in that, The fault type and the confidence result output by the migrated and optimized model are combined with the corresponding time sequence features to form a diagnosis history sequence, which is input into a time sequence prediction model to predict future operation states. The dynamic Bayesian model after migration optimization is continuously run on the target GIS device, real-time outputs the fault type label and the corresponding confidence value of each time step, synchronously records the multi-modal feature matrix at that time, and integrates according to the time step to form a diagnosis history sequence; The diagnosis history sequence is input into the time series prediction model, the time series prediction model learns the time sequence evolution rule of the diagnosis state, and outputs a prediction state vector of future multiple time steps; In the prediction stage, the time series prediction model generates a probability distribution of the future state according to the current diagnosis state and the historical window data, and outputs a confidence interval; Anomaly trend detection is performed on the prediction result, potential abnormal evolution signals are identified according to the drift rate of the confidence interval, and when the drift rate exceeds a set threshold, the system is marked as a potential fault trend and an early warning is issued.

9. A system using the multi-agent based gas insulated switchgear fault diagnostic method according to any one of claims 1 to 8, characterized in that, It includes a data acquisition module, a solving module, a fusion module, an updating module and a prediction module, and there is a connection between the modules; The data acquisition module is used to acquire signal data of the target device, and to extract features from the signal data to construct a multi-modal feature matrix; The solving module is used to extract the time delay features of acoustic and electromagnetic signals from the multi-modal feature matrix, establish a GIS propagation model, and solve the spatial coordinate position of the discharge source by wave field inversion algorithm; The fusion module is used to combine the spatial coordinate position and the multi-modal feature matrix into a complete fusion feature vector, and input it into the dynamic Bayesian model to output the fault type label and the corresponding confidence; The updating module is used to migrate the dynamic Bayesian model based on the transfer learning mechanism, and dynamically update the classification threshold; The prediction module is used to input the fault type and confidence result output by the migrated and optimized model and the corresponding time series features into the diagnosis history sequence, and input it into the time series prediction model to predict the future running state.

Images