Intelligent fault diagnosis method and system for electrical equipment

Abstract

The application relates to the technical field of electrical equipment fault diagnosis, in particular to an intelligent fault diagnosis method and system for electrical equipment, which comprises the following steps: unifiedly fusing equipment state information, electrical distance weighted connection and phase dynamic coupling relationship by constructing a multi-dimensional tensor model, and extracting an abnormal propagation mode by using high-order singular value decomposition; designing a space-time-frequency coupling interference stripping mechanism, combining structure-guided disturbance deconstruction, multi-scale dictionary learning and sparse low-rank decomposition to accurately separate propagating and non-propagating interference; reconstructing a fault trajectory based on a generative adversarial mechanism, coupling a graph structure dynamic encoder, a topological consistency discriminator and a time-controllable generator, and restoring a real propagation path; and finally integrating tensor semantic compression, a three-view graph neural network and fault label back-projection interpretation through a multi-source semantic fusion mechanism. The application realizes cross-time-space and cross-structure fault diagnosis and tracing, and improves accuracy and interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment fault diagnosis, and particularly relates to an intelligent fault diagnosis method and system for electrical equipment. BACKGROUND

[0002] The research background of the intelligent fault diagnosis technology for electrical equipment focuses on the objective development demand of the continuous expansion of the scale of the modern power system and the significant improvement of the coupling complexity of the equipment; with the increasing complexity of the power grid topology structure and the multi-dimension diversification of the equipment state monitoring, the traditional diagnosis method based on a single index threshold or local state analysis faces challenges in the ability to depict the fault propagation across time and space.

[0003] A power distribution network optical fiber composite fault intelligent positioning and diagnosis method based on hybrid sensing is disclosed in Chinese Invention Patent No. CN119881542B, which comprises acquiring and preprocessing optical fiber and electrical sensing data, extracting time-frequency domain features through a multi-channel convolutional neural network, calculating dynamic weight coefficients using an attention mechanism for feature fusion, constructing a dynamic fault propagation graph, combining graph convolution and causal convolution for time-space feature extraction, and realizing fault positioning and type identification.

[0004] The strong electrical distance correlation, dynamic phase coupling and nonlinear interference superposition existing between electrical equipment require the diagnosis model to have efficient fusion capability for multi-dimensional heterogeneous data and precise modeling capability for topology propagation rules; at the same time, the demand for explainability of fault tracing and active operation and maintenance decision support in industrial sites further promotes the evolution of the diagnosis technology to an intelligent collaborative paradigm that integrates physical mechanisms and data-driven; under this background, it is urgent to build a multi-dimensional analysis framework that can uniformly represent equipment state evolution, topology connection weight and phase dynamic response, develop interference stripping and trajectory reconstruction technology to cope with data missing and noise interference scenarios, and realize high-precision and strong-explainability output of the diagnosis results through cross-domain semantic fusion mechanism, so as to form an intelligent diagnosis system supporting the safe operation of the power grid. SUMMARY

[0005] The present application aims to solve the problems in the background art and proposes an intelligent fault diagnosis method and system for electrical equipment.

[0006] The technical solution of the present application is an intelligent fault diagnosis method for electrical equipment, comprising the following specific implementation steps:

[0007] S1, collecting electrical equipment operation data through a sensor network, constructing a normalized state vector, constructing a topology weight matrix based on electrical distance and physical connection, defining a phase coupling matrix using a phase difference cosine response, fusing the normalized state vector, the topology weight matrix and the phase coupling matrix, constructing a three-dimensional dynamic tensor for topology evolution modeling, and extracting a dominant abnormal propagation pattern through high-order singular value decomposition;

[0008] S2, extract the node disturbance intensity and change rate sequence from the three-dimensional dynamic tensor, obtain the base function response and residual through wavelet residual multi-scale decomposition, construct the disturbance response similarity matrix using topology guidance, identify the propagation interference, generate the interference frequency band set, and strip the system disturbance through low-rank sparse decomposition of the residual matrix, and finally generate the purified topology evolution tensor;

[0009] S3, based on the topology evolution tensor, construct a sliding window topology perception adjacency matrix to generate structure embedding, use the generator to combine device static characteristics and time offset to recover missing trajectories, ensure trajectory structure consistency through discriminator adversarial training, fuse candidate paths based on similarity and propagation force score, and optimize candidate trajectory reconstruction output;

[0010] S4, compress the device dynamic trajectory into a semantic vector by weighting the response field intensity, construct a three-view adjacency matrix by fusing structure topology, semantic similarity and attribute homogeneity, generate node embedding representation through graph attention network, map fault probability and locate gradient sensitive dimension to trace back to the source.

[0011] Preferably, the construction process of the three-dimensional dynamic tensor is as follows:

[0012] The sensor network is laid out to collect the running data of any node of the electrical equipment at a certain time, including: voltage amplitude, current amplitude, phase angle, frequency and temperature, and a normalized state vector is constructed;

[0013] Based on the physical connection mark and the electrical distance, the topological weight matrix between devices is defined by a natural exponential decay function, which quantifies the connection density and influence strength;

[0014] For the phase coupling characteristics between electrical equipment, a phase response matrix is defined, whose elements are driven by the phase difference between devices, and the cosine function is used to nonlinearly compress the phase difference to calculate the phase coupling response value;

[0015] Fusion of normalized state vector, topological weight matrix and phase coupling matrix, definition of three-dimensional tensor :

[0016] ;

[0017] Among them, represents the three-dimensional dynamic state tensor in N nodes and T time steps; represents the comprehensive state influence strength of node i on node j at time t k ; represents the transposition operation on the state vector ; represents the electrical connection relationship strength weight between devices i and j; represents the electrical connection relationship strength weight between devices i and j;k , the phase coupling response value between device i and device j.

[0018] Preferably, the dominant anomalous propagation mode extraction process is: ;

[0019] wherein, denotes the singular value of the rth principal component; denotes the feature vector of the rth mode in the device dimension; denotes the feature vector of the rth mode in the connection relationship dimension; denotes the feature vector of the rth mode in the time dimension; R denotes the number of retained principal propagation modes; denotes the outer product operation of the tensor.

[0020] Preferably, the propagating interference identification process is:

[0021] The disturbance comprehensive influence quantity of each node is extracted from the three-dimensional dynamic tensor, a first-order difference disturbance change rate is introduced, the tensor is projected to the node-time plane to obtain a disturbance sequence, and the overall disturbance intensity and the perturbation intensity are quantified: ; ;

[0022] wherein, denotes the overall disturbance intensity perceived by node i at time t k ; denotes the disturbance change rate of node i at the adjacent time;

[0023] The disturbance sequence is decomposed into a multi-scale disturbance subspace and a non-explained residual, and a sparse representation of the node disturbance is constructed: ;

[0024] wherein, M denotes the total number of scale levels of decomposition; denotes the mth disturbance basis function; denotes the response coefficient of node i to the mth layer scale disturbance basis function; denotes the disturbance residual;

[0025] Based on the node disturbance basis function component and the state of the adjacent node, similarity propagation matching is performed, topological information is introduced, and a response similarity matrix is constructed: ;

[0026] If is greater than a set threshold value, the frequency band is marked as a propagating interference, and the interference frequency band set is recorded.

[0027] wherein, denotes the disturbance basis function component of device i at scale m. represents the perturbation basis function component of device j at scale m; represents the topological guidance similarity of the perturbation morphology of device i and device j at scale m; represents the perturbation shape similarity measure function; represents the electrical topological connection weight of device node i and node j.

[0028] Preferably, the generation process of the purification topological evolution tensor is as follows:

[0029] Collect node perturbation residuals Combine the two-dimensional matrix, constrain the nuclear norm and L1 norm by low-rank sparse joint decomposition, ensure smoothness by time difference gradient, strip systematic interference to capture shared trends, and retain sparse local mutation anomalies:

[0030] wherein, represents the sparse term adjustment coefficient; represents the trend smoothing adjustment coefficient; represents the nuclear norm, i.e., the sum of singular values of the matrix; represents the L1 norm; represents the time difference gradient of L; L represents the low-rank term of the residual matrix; S represents the sparse term of the residual matrix; represents the perturbation residual matrix, each element , i.e., the residual value;

[0031] Strip the identified interference components from the original evolution tensor to construct the purified tensor:

[0032]

[0033] wherein, represents the purified device topological evolution tensor; represents the low-rank background disturbance term.

[0034] Preferably, the optimization process of the candidate trajectory reconstruction output is as follows:

[0035] Construct a sliding window topologically aware adjacency matrix to quantify the inter-node propagation probability, based on the time window evolution topology, and combine the node feature input graph attention network to generate structure embedding, capturing local structure semantics and propagation influence;

[0036] Based on the extracted structure embedding representation, construct a generator network that integrates time logic, input historical structure embedding vectors, device static feature vectors, and time offset variables, output device state reconstruction vectors, and model time evolution trends and topological semantics through graph neural networks and gated units to restore the physical logic consistent evolution trajectory of the missing segment;

[0037] ​​The constructed topology consistency discriminator evaluates the structural consistency of the generated trajectory and the real propagation path through an adversarial training mechanism, inputs the reconstructed state and the adjacency matrix, and outputs the authenticity probability, so as to optimize the generator to output a physically logical trajectory;

[0038] Based on the structural similarity and the propagation flow intensity, the confidence score of each candidate trajectory is calculated, and the optimal state estimation, i.e. the reconstruction output of the candidate trajectory, is output through weighted fusion.

[0039] Preferably, the adversarial loss function of the adversarial training mechanism is:

[0040] ;

[0041] Among them, represents the adversarial loss function; represents the real observed device state vector; represents the pseudo sample device state vector generated by the generator; and represents the expected value calculation; A represents the topology adjacency matrix at the current time point.

[0042] Preferably, the three-view adjacency matrix construction process is:

[0043] The semantic compression vector of each device node i is defined as: ;

[0044] The weight is defined as: ;

[0045] Among them, represents the semantic compression vector of device i; represents the semantic contribution weight of device i at time step t; represents the reconstructed state vector of device i at time t; represents the L2 norm of the vector; represents the full-cycle average state of device i; T represents the set time step number;

[0046] The three-view adjacency matrix is constructed:

[0047] The structure diagram G S : multi-level electrical topology evolution structure, adjacency matrix A S ;

[0048] The semantic graph G E : calculate the cosine similarity of the semantic vectors of any two devices:

[0049] ;

[0050] The attribute graph G A: meta-attribute matching constructs binary adjacency: ;

[0051] constructing fused adjacency matrix: ;

[0052] using graph attention network for node embedding:

[0053] ;

[0054] wherein, represents the similarity of node i and j in semantic space; represents the transposition operation on the semantic compression vector of node i; represents the attribute graph adjacency matrix; I represents the indicator function, which is 1 when the condition is met, otherwise 0; type(i) represents the type identification of node i; represents the adjacency matrix of the multi-view fusion graph; 、 and represent the weight factor; A S represents the topological structure graph adjacency matrix; A E represents the semantic similarity graph, i.e. the similarity of device evolution trend; A A represents the device attribute graph; represents the fused semantic representation of node i after graph attention network; represents the k-th layer graph attention network.

[0055] Preferably, the positioning process of locating the gradient sensitive dimension back to the source is:

[0056] define the final classification head as a linear projector: ;

[0057] set the threshold threshold Th, if , trigger the diagnostic label;

[0058] wherein, represents the fault probability vector of node i, with C classes; represents the weight matrix of the classification head, which maps the fused features to C-dimensional output; represents the classification head bias term;

[0059] Using reverse activation weight analysis to locate the dimension with the most diagnostic weight in the fused semantic vector, that is: calculate the gradient of the classification output to each dimension of the input semantic vector: ;

[0060] Accordingly: find the top k dimensions with the highest activation and associate their upstream adjacent nodes in the graph structure to assist in judging the potential fault source;

[0061] wherein, representing the fault class c against the semantic vector F i sensitivity of each dimension.

[0062] The technical scheme of the present application: an intelligent fault diagnosis system of an electrical equipment, which is used to execute the above-mentioned intelligent fault diagnosis method of an electrical equipment, comprising:

[0063] A multi-level topology evolution modeling module acquires operation data of the electrical equipment based on the laid sensor network, and constructs a multi-level electrical topology evolution tensor based on the electrical equipment operation data and the system topology structure, to accurately depict the device state change and the abnormal propagation path in the topology;

[0064] A heterogeneous interference deconstruction and multi-scale disturbance stripping module strips non-fault-related interference from the original state tensor, and retains key disturbance features that truly reflect fault evolution;

[0065] A structure-guided fault evolution trajectory reconstruction module is used to reconstruct a complete, coherent and structurally consistent fault evolution trajectory on the state data after disturbance stripping;

[0066] A multi-graph semantic fusion and fault output module performs deep semantic fusion analysis on the reconstructed state trajectory, and completes the final fault type determination and reverse interpretation.

[0067] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0068] The present application designs an intelligent fault diagnosis method and system of an electrical equipment, which unifies the electrical equipment state information, the topology weight and the phase dynamic coupling relationship through multi-dimensional tensor modeling, significantly improves the abnormal propagation depiction ability of a complex coupling scene, accurately separates the propagating and non-propagating interference sources based on the interference stripping technology of the space-time-frequency coupling mechanism combined with the topology-guided similarity analysis, greatly suppresses the misjudgment caused by non-fault disturbance, realizes the physical logic consistent reconstruction of the fault trajectory by fusing the graph structure dynamic encoder, the topology consistency discriminator and the time controllable generator, effectively solves the diagnosis blind area in the data missing or fuzzy scene, and synchronously improves the accuracy and explainability in the cross-time-space cross-structure diagnosis through the multi-source semantic fusion mechanism integrating the tensor compression, the three-view graph neural network and the gradient back projection explanation, to provide the intelligent decision support for the electrical equipment fault tracing and active operation and maintenance, which takes into account the theoretical rigor and engineering practicability. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 A method flowchart of an intelligent fault diagnosis method of an electrical equipment proposed by the present application is provided.

[0070] Figure 2A system architecture diagram of an intelligent fault diagnosis system of an electrical equipment is proposed for the present application. DETAILED DESCRIPTION

[0071] Embodiment one, as shown in the present application, an intelligent fault diagnosis method of an electrical equipment, comprising the following specific implementation steps: Figure 1

[0072] S1, construct a multi-dimensional tensor model, unify the multi-dimensional state information, the electrical distance weighted connection and the phase dynamic coupling into the space-time evolution tensor, reveal the abnormal propagation through tensor decomposition, and dynamically and comprehensively reflect the complex coupling relationship between the electrical equipment and its time sequence change, realize the efficient capture of abnormal fault information and the description of the propagation path, and the specific implementation process is as follows:

[0073] S11, lay out the sensor network, for electrical equipment node i, collect its running state index at time t k , construct state vector :

[0074] ;

[0075] Among them, represents the multi-dimensional normalized running state vector of the electrical equipment node i at time t k ; represents the voltage amplitude of device i at time t k , reflecting the stability and integrity of power supply; represents the current amplitude of device i at time t k , which is an important index to measure the device load and power consumption behavior; represents the phase angle of device i at time t k , which is used to analyze the phase difference between voltage and current; represents the frequency of device i at time t k , which reflects the stability of power grid frequency; represents the temperature of device i at time t k , which is an important external state signal of device running health; represents the transpose operation;

[0076] S12, based on the physical connection and electrical distance of the device, define the topological relationship weight matrix between the devices , wherein each element: ;

[0077] Among them, represents the electrical connection relationship strength weight between device i and j, which is used to express the connection density and influence degree of each device in the topological structure; represents the electrical distance between device i and j;​ denotes the distance decay control parameter, controlling the exponential decay degree of the middle distance term; denotes the physical connection flag quantity, binary (0 or 1), 1 if there is a physical connection between device i and device j, otherwise 0; denotes the natural exponential function;

[0078] S13, for the phase coupling characteristics between electrical devices, define the phase response matrix , the elements of which are driven by the phase difference between devices: ;

[0079] wherein, denotes the phase coupling response value between device i and device j at time t k , reflecting the phase synchronization relationship between the two devices, using the cosine function to nonlinearly compress the phase difference, the value closer to 1 indicates the more synchronized phase; denotes the phase angle difference between device i and j at time t k ; N denotes the total number of electrical device nodes, i.e. the number of devices monitored, modeled and diagnosed in the entire electrical network;

[0080] S14, fuse the above normalized state vector, topology weight matrix and phase coupling matrix to define a three-dimensional tensor:

[0081] ;

[0082] Specifically, the tensor calculation process is: for each time point t k , traverse all device pairs (i, j), calculate the product of the normalized state vector , , , form the tensor element , update in the continuous time sliding window, realize the topology dynamic evolution modeling;

[0083] wherein, denotes the three-dimensional dynamic state tensor in N nodes and T time steps, used for high-dimensional modeling of the multi-element state evolution process and abnormal propagation rule in the electrical system; denotes the comprehensive state influence intensity of node i on node j at time t k ; denotes the transposition operation on the state vector ;

[0084] S15, through the time sequence change of the tensor , use the tensor decomposition method (this embodiment uses high-order singular value decomposition HOSVD) to extract the dominant abnormal propagation mode:

[0085] ;

[0086] where, denotes the singular value of the rth principal component, measuring the weight or significance of the propagation pattern in the whole tensor, the larger the value, the more important the pattern is; denotes the eigenvector of the rth pattern in the device dimension (node dimension), representing which devices are most affected by the pattern; denotes the eigenvector of the rth pattern in the connection relationship dimension, reflecting which connection relationships the propagation is mainly concentrated on; denotes the eigenvector of the rth pattern in the time dimension, depicting the evolution trajectory of the abnormal propagation pattern in time; R denotes the number of retained principal propagation patterns, usually set according to the cumulative contribution rate of singular values (95% energy coverage is adopted in this embodiment), determining how many high-order propagation patterns are retained for anomaly extraction; denotes the outer product operation of the tensor, used to combine the vectors in each dimension to form a tensor approximation.

[0087] S2, design an interference modeling and stripping method based on a space-time-frequency coupling mechanism, through structure-guided disturbance decomposition + multi-scale disturbance dictionary learning + sparse low-rank anomaly recognition, the specific implementation process is as follows:

[0088] S21, from the constructed three-dimensional dynamic state tensor , extract the comprehensive influence of the disturbance of each node (i.e. the topology propagation disturbance aggregation value): ;

[0089] Then, in order to further depict the volatility of the disturbance, the disturbance change rate (i.e. the first-order difference) is introduced:

[0090] ;

[0091] Accordingly: project the tensor from three dimensions to the node-time plane to obtain a disturbance sequence with physical meaning;

[0092] where, denotes the overall disturbance intensity perceived by node i at time t k ; denotes the disturbance change rate (perturbation intensity) of node i at the adjacent time;

[0093] S22, input the disturbance sequence to the data-driven wavelet residual decomposer, based on the data-driven disturbance dictionary learning mechanism, decompose the disturbance sequence into a multi-scale disturbance subspace and a non-explanatory residual, and construct a sparse representation of the node disturbance:

[0094] ;

[0095] where M represents the total number of scale levels of wavelet / perturbation decomposition; represents the mth perturbation basis function, representing the typical perturbation pattern at this scale, which is adaptively learned from historical samples (such as transformer switching signals, lightning harmonic waveforms, etc.) through data-driven or sparse learning; represents the response coefficient of node i to the mth layer scale perturbation basis function, reflecting the significant degree or active degree of the perturbation pattern at the current node; represents the perturbation residual, i.e. the “unexplained component” that does not belong to the existing perturbation dictionary;

[0096] S23, the perturbation characteristics of node i are matched with the state of its adjacent nodes for similarity propagation, structural topology information is introduced to enhance the accuracy of interference recognition, and a perturbation response similarity matrix is constructed:

[0097] ;

[0098] wherein, represents the perturbation basis function component of device i at scale m; represents the perturbation basis function component of device j at scale m; represents the topological guidance similarity of the perturbation pattern of device i and device j at scale m; represents the perturbation shape similarity measure function, and the cosine similarity is adopted in the embodiment; represents the electrical topology connection weight of device node i and node j;

[0099] If multiple perturbation components exist among a group of strongly connected devices (i.e.: greater than the set threshold value) at the same time, the frequency band is marked as propagated interference, and recorded into the interference frequency band set ;

[0100] S24, the perturbation residual term of each node i is combined into a two-dimensional matrix ;

[0101] Then, low-rank sparse joint decomposition is performed: ;

[0102] Accordingly, the estimated low-rank term L is separated from the perturbation response of the original tensor, and a more true non-propagated mutation anomaly is retained;

[0103] wherein, represents the sparse term adjustment coefficient; represents the trend smoothing adjustment coefficient; denotes the nuclear norm, i.e., the sum of singular values of a matrix, to constrain the low-rank property of L; denotes the L1 norm, i.e., the sum of absolute values of all elements, to constrain the sparsity of S; denotes the time-difference gradient of L, i.e., the trend of L along the time axis, to constrain its stationarity (i.e., should not have dramatic jumps); L denotes the low-rank term of the residual matrix, capturing the shared disturbance trend existing in all nodes, such as systematic factors like thermal drift, long-period noise, device aging trend, etc.; S denotes the sparse term of the residual matrix, i.e., sparse local jump abnormal signals like short-time arc, contact fault, electromagnetic interference, etc.; denotes the disturbance residual matrix, with rows representing nodes and columns representing time; each element denotes the residual value;

[0104] S25, finally integrates the above analysis results, and removes the identified interference components from the original evolution tensor to construct a purified tensor: ;

[0105] wherein, denotes the purified device topology evolution tensor; denotes the low-rank background disturbance term, i.e., the systematic long-term trend or periodic disturbance component of device i at time t k .

[0106] S3, based on the purified topology evolution tensor , a new trajectory reconstruction method is constructed, which integrates structure prior, time dependence and generative adversarial mechanism, couples the graph structure guided dynamic encoder, structure consistency discriminator and time controllable trajectory generator into one, accurately reconstructs the missing or fuzzy historical state segment of the electrical equipment, and restores the real fault propagation trajectory and path evolution sequence, and the specific implementation process is as follows:

[0107] S31, by constructing dynamic structure embedding, the evolution relationship of the actual propagation path between devices over time is extracted, thereby providing accurate topological guidance for subsequent trajectory generation, specifically:

[0108] a sliding window topology-aware adjacency matrix is constructed for each time t k :

[0109] ;

[0110] wherein, denotes the propagation probability of node i to node j in the topology adjacency matrix at time point k, i.e., the possible influence intensity of node i to j at that time; denotes the sliding time window width, i.e., the time span considered forward from the current time, used to construct the time sequence-aware topology evolution;

[0111] Input the graph neural network module (graph attention network is adopted in this embodiment) with the adjacency matrix and node feature at each moment to generate a structure embedding:

[0112] ;

[0113] wherein, represents a state feature matrix of each device at moment k, N is the total number of devices, and d is the state dimension; represents a structure embedding representation extracted under the graph attention network (GAT), reflecting the local structure semantics and context propagation influence of each node at the current moment;

[0114] S32, based on the extracted structure embedding representation, a generator network based on fusion time logic is used to restore the evolution trajectory on the missing or abnormal segment so as to make it comply with the physical evolution sequence and propagation direction, and a generator function is defined The output of the generator function is: ;

[0115] wherein, represents a state reconstruction vector of device i output by the trajectory generator at time t; represents a structure embedding vector of device i at time t-1, reflecting the topological propagation state semantics before the current moment of the device; represents a static structure feature vector of device i, including device type (transformer, switch, etc.), importance rating, historical fault frequency, etc., for personalized trajectory generation; represents a time offset of the current reconstruction time point t relative to the starting point t s of the fault segment, for simulating the time evolution trend;

[0116] It should be noted that the generator is a core module in the structure-guided fault evolution trajectory reconstruction method, which aims to dynamically restore the state evolution trajectory of the electrical device in the missing or abnormal segment by using the historical structure embedding information, device static features and time offset variable. The input of the generator includes the structure embedding vector of the previous moment, the device static attribute vector and the time offset of the current moment and the starting point of the trajectory missing By fusing the structure semantics extracted by the graph attention network (GAT) and the modeling of the time evolution trend by the GRU type gating unit, the generator can output continuous, structure-consistent and time-evolution-logic state estimation values;

[0117] S33, to ensure that the generated trajectory is consistent with the real propagation path in structure, a topological consistency discriminator The core goal is to improve the structural credibility of the trajectory in the generation-discrimination adversarial mechanism, specifically:

[0118] The discrimination probability is defined as:

[0119] The output value represents whether the reconstructed state conforms to the propagation logic under the current topology structure (if close to 1, it means that the reconstructed state is consistent with the real topology propagation trend, conforming to the fault evolution logic; if close to 0, it means that the state may deviate from the network propagation path, which is a false reconstruction);

[0120] The training objective is to minimize the following adversarial loss:

[0121]

[0122] where, represents the credibility score output by the discriminator at time t (between 0-1), and represents the reconstructed state whether it conforms to the topology propagation rule; represents the adversarial loss function, which is used to optimize the training process of the generator and the discriminator, making the generator generate more realistic and credible results. This loss function consists of two parts: the first part is to let the discriminator maximize the "true" output (close to 1) for real samples, and the second part is to let the discriminator minimize the acceptance of generated samples (close to 0); represents the real observed device state vector, which is used as a positive sample in the training of the discriminator; represents the pseudo-sample device state vector generated by the generator, which is used as a negative sample in the training of the discriminator; and represent the expected value calculation, i.e., solving the average loss of the sample; A represents the topology adjacency matrix at the current time point, which is used to constrain whether the generated state is consistent with the physical connection between devices, as a structural prior injection;

[0123] It should be noted that the topology consistency discriminator is a discrimination mechanism for evaluating the consistency between the fault evolution trajectory and the actual propagation structure of the electrical network. Its core role is to ensure that the state sequence constructed by the trajectory generator is not only numerically reasonable, but also structurally consistent with the physical topology and propagation logic. The discriminator takes the generated state vector at each time point and the corresponding topology adjacency matrix as input, and outputs a probability value representing the authenticity of the trajectory. Through adversarial training, the generated trajectory is forced to maintain continuity while strictly following the network propagation rules. Specifically, the discriminator uses a structure-aware neural network to encode the input state with topology features, extracts the topology profile of fault evolution by combining the historical evolution semantics of the propagation path, and effectively identifies trajectories that do not conform to physical logic such as structural drift and fake paths; ​​

[0124] S34, introduce a path confidence fusion mechanism, based on structural similarity and propagation force score to optimize multiple candidate trajectories, specifically:

[0125] Define a consistency score function: ;

[0126] The final output fusion result: ;

[0127] Where, represents the reconstruction output of the rth path of the candidate trajectory; represents the consistency confidence score of the state estimation of device i at time t generated by the rth candidate trajectory; represents the topological adjacency probability vector of device i at time t, that is, its out-edge propagation probability distribution in the graph structure; represents the propagation flow intensity score of device i at time t, which measures the structural activity or signal flux as a propagation source / intermediary.

[0128] S4, on the basis of fault trajectory reconstruction under the guidance of structure, facing the final diagnosis output target, a multi-source semantic fusion and fault mode output method is constructed, which fuses tensor semantic compression, three-view graph neural network fusion and fault label back-projection interpretation mechanism, completes the information integration and reasoning across space-time, indicators and structures in electrical systems, so that the diagnosis result is not only accurate, but also has strong explainability and generalization. The specific implementation process is:

[0129] S41, a tensor semantic compression mechanism based on response field intensity weighting is constructed, which compresses the dynamic trajectory of multiple time steps and multiple devices into a joint semantic vector in the response domain, extracts the semantic response main line of the device at the key moment, specifically:

[0130] Define the semantic compression vector of each device node i as: ;

[0131] The weight is defined as: ;

[0132] Where, represents the semantic compression vector of device i, which represents the comprehensive expression of the important state evolution of the device in the entire time period T, and is the input feature of the subsequent diagnosis model; represents the semantic contribution weight of time step t to device i, reflecting the deviation degree of the state at that moment from its mean value; represents the reconstruction state vector of device i at time t, which comes from the fault trajectory reconstruction output in step S3; represents the L2 norm of the vector, that is, the Euclidean distance, which characterizes the difference between the state at a certain time point and the mean value; T represents the set time step number; and

[0133] S42, a multi-view graph fusion mechanism is constructed to map the structure topology, attribute homogeneity and state semantic similarity to a unified graph representation, and learn a final fusion representation based on a graph attention mechanism, specifically:

[0134] A three-view adjacency matrix is constructed:

[0135] Structure graph G S : multi-level electrical topology evolution structure from step S1, adjacency matrix A S ;

[0136] Semantic graph G E : calculate the cosine similarity of the semantic vectors of any two devices:

[0137] ;

[0138] Attribute graph G A : device category (such as transformer, circuit breaker), installation environment (indoor / outdoor) and other meta-attribute matching binary adjacency: ;

[0139] Construct a fusion adjacency matrix: ;

[0140] Use a graph attention network for node embedding:

[0141] ;

[0142] wherein, represents the similarity of nodes i and j in the semantic space, and the larger the value, the more similar the state evolution; represents the transpose operation on the semantic compression vector of node i; represents that when the attributes of devices i and j are the same (such as type, installation environment, etc.), it is 1, otherwise it is 0, representing structural similarity; I represents an indicator function, which is 1 when the condition is met, otherwise it is 0; type(i) represents the type identification of node i (for example, circuit breaker, bus, transformer, etc.); represents the adjacency matrix of the multi-view fusion graph; , and represent weight factors; A S represents the topology structure graph adjacency matrix; A E represents the semantic similarity graph, i.e. the similarity of device evolution trend; A A represents the device attribute graph; F i represents a fused semantic representation of node i after the graph attention network, used for final fault identification; G i represents a k-th layer graph attention network used for learning the weighted information propagation between nodes and neighbors;

[0143] S43, after obtaining the fused semantic representation F i , it is mapped to a specific fault label output (multi-label, multi-level), and a reverse interpretation mechanism is attached to enhance the understandability and traceability of the diagnosis result, specifically:

[0144] (1) define the final classification head as a linear projector: ;

[0145] Set the threshold threshold Th, if , trigger the diagnosis label;

[0146] Wherein, F i represents the fault probability vector of node i, C classes in total, and each dimension represents the probability of occurrence of the class; W represents the weight matrix of the classification head, which maps the fused features to C-dimensional output; b represents the bias term of the classification head;

[0147] (2) adopt reverse activation weight analysis to locate the dimension with the highest diagnostic weight in the fused semantic vector, that is, calculate the gradient of each dimension in the input semantic vector to the classification output: ;

[0148] Accordingly: find the top k dimensions with the highest activation and associate them with the upstream adjacent nodes in the graph structure to assist in judging the potential fault source;

[0149] Wherein, F c represents the semantic vector F i of fault class c.

[0150] Embodiment two, as Figure 2 shown, the intelligent fault diagnosis system of the electrical equipment proposed by the application is used to execute the intelligent fault diagnosis method of the electrical equipment proposed in embodiment one, which comprises: a multi-level topology evolution modeling module, a heterogeneous interference decomposition and multi-scale disturbance stripping module, a structure-guided fault evolution trajectory reconstruction module, and a multi-graph semantic fusion and fault output module.

[0151] The multi-level topology evolution modeling module collects the operation data of the electrical equipment based on the laid sensor network, and constructs a multi-level electrical topology evolution tensor based on the electrical equipment operation data and the system topology structure, to accurately depict the device state change and abnormal propagation path in the topology;

[0152] The heterogenous interference deconstruction and multi-scale disturbance stripping module strips non-fault-related interference (including but not limited to dispatch fluctuation, operation disturbance, temperature drift) from the original state tensor, and retains key disturbance characteristics that truly reflect fault evolution;

[0153] The structure-guided fault evolution trajectory reconstruction module is used to reconstruct a complete, coherent and structure-consistent fault evolution trajectory on the state data after disturbance stripping;

[0154] The multi-graph semantic fusion and fault output module performs deep semantic fusion analysis on the reconstructed state trajectory, and completes final fault type determination and reverse interpretation.

[0155] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A method of intelligent fault diagnosis of an electrical device, characterized by, The specific implementation steps include the following: S1. Collect electrical equipment operation data through sensor network, construct normalized state vector, construct topology weight matrix based on electrical distance and physical connection, define phase coupling matrix using phase difference cosine response, fuse normalized state vector, topology weight matrix and phase coupling matrix, construct three-dimensional dynamic tensor for topology evolution modeling, and extract dominant anomaly propagation mode through high-order singular value decomposition. The construction process of the three-dimensional dynamic tensor is as follows: Deploy a sensor network to collect operating data of any node of electrical equipment at a certain moment, including: voltage amplitude, current amplitude, phase angle, frequency and temperature, and construct a normalized state vector; Based on physical connection markers and electrical distances, a topology weight matrix between devices is defined using a natural exponential decay function to quantify connection density and impact intensity. To address the phase coupling characteristics between electrical equipment, a phase response matrix is ​​defined, whose elements are driven by the phase difference between the equipment. The phase coupling response value is calculated by nonlinearly compressing the phase difference using a cosine function. Fusing the normalized state vector, the topological weight matrix, and the phase coupling matrix, a three-dimensional tensor is defined : ； wherein, represents a three-dimensional dynamic state tensor in N nodes, T time steps; represents the state of node i at time t k , the comprehensive state influence strength of node i on node j; represents the transpose operation on the state vector ; represents the electrical connection relationship strength weight between device i and j; represents the phase coupling response value between device i and device j at time t k ; S2. Extract the node perturbation intensity and rate of change sequence from the three-dimensional dynamic tensor, obtain the basis function response and residual through wavelet residual multi-scale decomposition, construct the perturbation response similarity matrix using topology guidance, identify propagating interference, generate interference frequency band set, remove system perturbation by low-rank sparse decomposition of residual matrix, and finally generate purified topology evolution tensor. S3. Based on the topological evolution tensor, a sliding window topologically aware adjacency matrix is ​​constructed to generate structural embeddings. The generator is combined with the device's static features and time offset to recover the missing trajectory. The discriminator is used for adversarial training to ensure the consistency of the trajectory structure. Candidate paths are fused based on similarity and propagation power scores to optimize the candidate trajectory reconstruction output. The optimization process for the candidate trajectory reconstruction output is as follows: A sliding window topology-aware adjacency matrix is ​​constructed to quantify the propagation probability between nodes. The topology is evolved based on a time window, and a structural embedding is generated by combining node feature input graph attention network to capture local structural semantics and propagation effects. A generator network integrating temporal logic is constructed based on the extracted structural embedding representation. The inputs are historical structural embedding vectors, device static feature vectors, and time offset variables. The output is device state reconstruction vector. The temporal evolution trend and topological semantics are modeled collaboratively by graph neural networks and gating units to restore the evolution trajectory that is consistent with the physical logic of the missing segment. A topology consistency discriminator is constructed to evaluate the structural consistency between the generated trajectory and the real propagation path through an adversarial training mechanism. The input is the reconstructed state and the adjacency matrix, and the output is the probability of authenticity. The generator is optimized to output a trajectory with reasonable physical logic. The confidence scores of each candidate trajectory are calculated based on structural similarity and propagation flow intensity, and the optimal state estimate is output by weighted fusion, which is the reconstruction output of the candidate trajectory. S4. The dynamic trajectory of the device is compressed into a semantic vector by weighting the response field intensity. The three-view adjacency matrix is ​​constructed by integrating structural topology, semantic similarity and attribute homogeneity. The node embedding representation is generated by graph attention network, which maps the failure probability and locates the gradient sensitive dimension to trace back to the source.

2. The intelligent fault diagnosis method for electrical equipment according to claim 1, characterized in that, The dominant abnormal propagation pattern extraction process is: ; in, Let represent the singular value of the r-th principal component; This represents the feature vector of the r-th pattern in the device dimension; This represents the feature vector of the r-th pattern in the connection dimension; Represents the feature vector of the r-th mode in the time dimension; R represents the number of main propagation modes retained; This represents the outer product operation of tensors.

3. The intelligent fault diagnosis method for electrical equipment according to claim 2, characterized in that, The process of identifying propagating interference is as follows: The combined perturbation influence of each node is extracted from the three-dimensional dynamic tensor. A first-order differential perturbation rate is introduced, and the tensor is projected onto the node-time plane to obtain the perturbation sequence, quantifying the overall perturbation intensity and the micro-perturbation intensity. ； ； in, This indicates that node i at time t k The perceived overall disturbance intensity; This represents the rate of change of the disturbance at node i in adjacent time intervals; perturbation sequence The perturbation is decomposed into a multi-scale perturbation subspace and non-interpretive residuals, and a sparse representation of the node perturbation is constructed: ; Where M represents the total number of scale levels in the decomposition; This represents the m-th perturbation basis function; This represents the response coefficient of node i to the basis function of the m-th scale perturbation; Indicates the disturbance residual; Similarity propagation matching is performed based on the perturbation basis function components of nodes and the states of neighboring nodes, and topological information is introduced to construct a response similarity matrix: ； like If the value exceeds the set threshold, the frequency band is marked. For propagating interference, the interference frequency bands are recorded. ; in, Represents the perturbation basis function components of device i at scale m; This represents the perturbation basis function components of device j at scale m; This represents the topological guided similarity between device i and device j in terms of perturbation morphology at scale m; This represents a perturbation shape similarity measurement function; This represents the electrical topology connection weight between device node i and node j.

4. The intelligent fault diagnosis method for electrical equipment according to claim 3, characterized in that, The generation process of the purification topology evolution tensor is as follows: Collect node perturbation residuals By combining two-dimensional matrices, constraining the nuclear norm and L1 norm through low-rank sparse joint decomposition, ensuring stationarity through temporal difference gradients, stripping away systematic disturbances to capture shared trends, and preserving sparse local mutation anomalies: ； in, This represents the adjustment coefficient for the sparse term; Indicates the trend smoothing adjustment coefficient; The nuclear norm is the sum of the singular values ​​of a matrix. Represents the L1 norm; Let L represent the time difference gradient; L represents the low-rank term of the residual matrix; S represents the sparse term of the residual matrix. This represents the perturbation residual matrix, where each element... That is, the residual value; The identified interfering components are removed from the original evolutionary tensor to construct a purified tensor: ； in, This represents the topology evolution tensor of the purified equipment. This represents the low-rank background disturbance term.

5. The intelligent fault diagnosis method for electrical equipment according to claim 4, characterized in that, The adversarial loss function for the adversarial training mechanism is: ； in, Represents the adversarial loss function; Represents the device state vector of actual observation; This represents the pseudo-sample device state vector generated by the generator; and represents the expected value calculation; A represents the topological adjacency matrix at the current time point.

6. The intelligent fault diagnosis method for electrical equipment according to claim 5, characterized in that, The process of constructing the adjacency matrix of the three views is as follows: Define the semantic compression vector for each device node i as follows: ; Weight Defined as: ; in, Represents the semantic compression vector of device i; This represents the semantic contribution weight of time step t to device i; This represents the reconstructed state vector of device i at time t; The L2 norm of a vector; This represents the average state of device i over the entire cycle; T represents the set number of time steps. Construct the adjacency matrix of the three views: Structural graph G S : Multistage electrical topology evolution structure, adjacency matrix A S ; Semantic graph G E : Compute cosine similarity of semantic vectors of any two devices: ； Attribute graph G A Meta-attribute matching constructs binary adjacency: ; Constructing a fused adjacency matrix: ; Node embedding using graph attention networks: ； in, This represents the similarity between nodes i and j in the semantic space; This indicates that the semantic compression vector of node i is transposed. Represents the adjacency matrix of the attribute graph; I represents the indicator function, which is 1 when the condition is met and 0 otherwise; type(i) represents the type identifier of node i; The adjacency matrix represents the multi-view fusion graph; , and Indicates the weighting factor; A S Represents the adjacency matrix of the topological graph; A E This represents a semantic similarity graph, i.e., the similarity of device evolution trends; A A Represents the device attribute diagram; This represents the fused semantic representation of node i after it passes through the graph attention network. This represents the attention network of the k-th layer graph.

7. The intelligent fault diagnosis method for electrical equipment according to claim 6, characterized in that, The process of locating the source by tracing back to the gradient-sensitive dimension is as follows: Define the final classification head as a linear projector: ; Set a threshold Th, if If so, the diagnostic label will be triggered; in, Let i represent the failure probability vector of node i, which has C classes. The weight matrix representing the classification head maps the fused features to a C-dimensional output; Indicates the category head bias item; Inverse activation weight analysis is used to locate the dimension with the most diagnostic weight in the fused semantic vector, that is, to calculate the gradient of the classification output with respect to each dimension of the input semantic vector: ; Therefore: find the top k dimensions with the highest activation and associate them with their upstream adjacent nodes in the graph structure to help identify potential sources of failure; in, The fault category c represents the semantic vector F. i Sensitivity of each dimension.

8. An intelligent fault diagnosis system for electrical equipment, used to execute the intelligent fault diagnosis method for electrical equipment according to any one of claims 1 to 7, characterized in that, include: The multi-level topology evolution modeling module collects operating data of electrical equipment based on the deployed sensor network, and constructs a multi-level electrical topology evolution tensor based on the operating data of electrical equipment and the system topology, in order to accurately characterize the changes in equipment state and the propagation path of anomalies within the topology; The heterogeneous interference deconstruction and multi-scale perturbation stripping module strips non-fault-related interference from the original state tensor, while retaining key perturbation features that truly reflect the fault evolution. The structure-guided fault evolution trajectory reconstruction module is used to reconstruct a complete, coherent and structurally consistent fault evolution trajectory on the state data after the disturbance has been removed. The multi-graph semantic fusion and fault output module performs deep semantic fusion analysis on the reconstructed state trajectory and completes the final fault type determination and reverse interpretation.

Images