Power equipment anomaly detection and early warning system and method

Abstract

The invention relates to the technical field of power equipment early warning, in particular to a power equipment anomaly detection early warning system and method. The method specifically comprises the following steps: acquiring multi-source sensing data of target power equipment; performing cross-modal collaborative denoising based on semantic graph modeling and multi-scale residual feedback; fusing the space-time similarity and abnormal diffusion characteristics of the equipment, constructing a weighted relation graph, carrying out abnormal sensitive clustering by using a graph neural network, and identifying a potential abnormal group; abstracting historical anomalies into fault gene vectors, matching an evolution path, and predicting a new anomaly propagation trend; an expert knowledge graph and a graph neural network are introduced, double-domain decisions are dynamically fused through an objection-driven game, and the early warning credibility is improved; grading early warning is triggered according to propagation influence evaluation, abnormal features and evolution paths are written back to a knowledge base in a structured mode, collaborative optimization of the model and a rule base is driven, and a detection-early warning-optimization self-closed-loop system is formed. According to the method, the problems of dynamic relevance modeling, new anomaly prediction and decision credibility are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment early warning, and particularly relates to a power equipment anomaly detection and early warning system and method. Background Art

[0002] As a key link to ensure the safe and stable operation of the power grid, power equipment anomaly detection has long faced challenges such as difficulties in dynamically correlating multi-source data under complex working conditions, insufficient early prediction ability for new anomalies, and limited decision-making credibility; traditional methods mostly rely on single-sensor data or static clustering analysis, and it is difficult to effectively capture the anomaly conduction characteristics in the collaborative operation of equipment groups, and it is prone to false positives and false negatives under the interference of low-quality data; with the intelligent upgrade of the power system, equipment status monitoring data presents characteristics such as multi-modal, high-dimensional, and strong time-series correlation. Existing methods based on rule engines or isolated models have significant limitations in aspects such as anomaly propagation path reasoning, cross-device anomaly resonance analysis, and dynamic knowledge evolution, and it is difficult to meet the requirements of real-time early warning and closed-loop optimization.

[0003] A Chinese invention patent with the publication number CN118898342B discloses a power equipment safety early warning method and system based on multi-modal data, including: constructing a spatial reference distribution coordinate with an environmental attribute deviation vector, an audio attribute deviation vector, and an operating attribute deviation vector, performing case retrieval in a power equipment safety management case library, constructing a recursive network tree for node weight distribution, and obtaining a node weight distribution result; performing inverse growth aggregation on the recursive network tree according to the node weight distribution result to obtain a power equipment anomaly index, and executing power equipment early warning when it is greater than or equal to the power equipment anomaly index threshold; solving the technical problem in the prior art that due to the lack of a comprehensive analysis scheme for multi-modal data, it is impossible to effectively utilize data of multiple attributes for comprehensive judgment, and thus it is difficult to achieve accurate early warning of power equipment.

[0004] In addition, the split application of expert experience and data-driven models leads to the lack of interpretability in the anomaly discrimination process, especially in new anomaly or low-confidence scenarios, it is difficult to form a credible decision; aiming at the above problems, it is urgent to construct an intelligent detection system that integrates multi-source perception, dynamic correlation modeling, and knowledge self-evolution to achieve precise control of the entire life cycle of power equipment anomalies. Summary of the Invention

[0005] The object of the present invention is to propose a power equipment anomaly detection and early warning system and method for the problems existing in the background art.

[0006] The technical solution of the present invention: A power equipment anomaly detection and early warning method includes the following specific implementation steps: S1. Collect the operating status, environmental parameters, and historical maintenance record data of the target power equipment; S2. A three-domain fusion denoising framework for dynamic adjustment of noise suppression weights in different data modalities through data source correlation dynamic measurement, temporal semantic stability discrimination, and graph structure context redundancy suppression, for cross-modal collaborative denoising, introducing a structure-aware multi-scale consistency function to perform feedback correction on the denoising results; S3. Extract device features through multi-scale semantic embedding, construct an elastic association graph by combining semantic similarity, abnormal resonance degree, and redundancy suppression, use a graph neural network to fuse abnormal perturbation signals for feature reconstruction, and achieve dynamic device grouping through multi-objective optimization clustering and edge pruning optimization; S4. Construct a fault gene map through historical fault data, calculate edge weights by fusing frequency and delay factors, activate potential paths in combination with the current abnormal state, estimate the future propagation probability of non-abnormal devices, and generate a final abnormal propagation prediction map; S5. Encode node features and propagation risks through constructing a structured knowledge graph and a graph attention network, design a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model predictions, and fuse logical predicate discrimination and classification results to output a comprehensive risk level; S6. Generate multi-dimensional risk scores through propagation influence evaluation and map them to warning levels, parse abnormal features and propagation paths to generate knowledge entries after triggering a warning, dynamically update the rule base based on credibility evaluation, and design an adjustment mechanism driven by rule usage frequency and prediction consistency to optimize model parameters, forming a dual-domain self-evolving closed loop.

[0007] Preferably, the denoising process of cross-modal collaborative denoising is as follows: S21. Use an attention-aware pooling method to extract the semantic vectors of each perception source at each moment, and calculate the semantic similarity between every two perception sources i and j at the current moment t: ; ; where represents the data sequence of the i-th perception source within the time window [t - k, t]; represents the semantic feature representation of the data sampled by the perception source i at the moment t; represents the semantic feature vector of the i-th perception source at the moment t; represents the attention-aware pooling operation; k represents the length of the historical window; represents the semantic similarity weight between the i-th and j-th perception sources at the moment t; here represents the transpose operation; represents the Euclidean norm; represents the indicator function; S22. Calculate the mean volatility within the sliding window for each perception dimension and set a dynamic enhancement coefficient: ; ; Among them, represents the short-term fluctuation intensity of the current perception dimension; K represents the size of the sliding window; represents the j-th perception index value of the i-th perception source at time t; represents the dynamic enhancement coefficient; is used to regulate the fluctuation compression rate; is used to prevent excessive suppression; represents the natural exponential function; S23. Generate denoised data through a denoising function, construct a semantic residual loss function based on the semantic neighborhood weighted reference value, combine the graph structure edge weight and the attention pooling mapping function, and balance the residual and semantic consistency losses to optimize the denoising process; S24. Extract multi-layer features of the original and denoised data through a multi-scale aggregation function, define the structural consistency loss and integrate it into the comprehensive optimization objective, and use the regulation factor to balance the multi-scale structure preservation intensity to suppress local overfitting; S25. Output the denoised data sequence of any power device i.

[0008] Preferably, the semantic residual loss function is: ; ; ; Among them, represents the semantic neighborhood reference value of the i-th perception source; represents the set of neighbor nodes connected to node i in the perception graph; represents the edge weight in the graph structure, that is, the semantic similarity weight between the i-th and j-th perception sources at time t; represents the denoised data, the denoised value of the j-th dimension of each perception source; represents a hyperparameter; represents the mapping function for extracting semantic embeddings, that is, the attention perception pooling operation ; is the denoising function; represents the denoised data; D represents a three-dimensional tensor used to uniformly express multi-dimensional index data collected from multiple perception nodes over a period of time; S represents the total number of perception sources, ; F represents the number of indicators collected by any perception source, ; T represents the total time series, .

[0009] Preferably, the grouping process of dynamic device grouping is as follows: S31. For the denoised data sequence of each device i Perform multi-scale semantic feature extraction e i : ; Among them, represents the denoised data sequence of device i, with a time window length of k and F perceptual features at each moment; represents the downsampled sequence; and represent the gated recurrent unit; represents the convolutional-attention module; represents the concatenation operation; e i represents the multi-scale semantic embedding vector of device i; S32. Combine multi-scale semantic similarity and abnormal resonance degree to construct an elastic weighted multi-scale similarity graph between devices, construct the edge weight adjacency matrix A = { }, and output the elastic correlation graph between devices ; ; ; ; ; Among them, represents the edge weight value between nodes i and j; represents the Sigmoid function; , , represent adjustable weights; represents the multi-scale semantic embedding vector of device i e i and the multi-scale semantic embedding vector of device j e j The semantic similarity, where T represents the vector transpose operation; represents the abnormal resonance degree; t represents the current moment; represents the time variable in the sliding time window; and respectively represent whether device i is abnormal at time τ and whether device j is abnormal at time τ; represents the indicator function; represents the redundancy; and represent the denoised perceptual data sequence; represents the cosine similarity; represents the logical AND operator; S33. Use the multi-scale semantic embedding vector as the initial feature, introduce a perturbation signal based on the abnormal score, propagate and fuse the adjacent node information through the weighted graph convolution, and output the final node feature after multiple layers of iteration; S34. Initialize the clustering framework through K-Means and construct an anomaly-aware clustering objective function that combines structural consistency and anomaly aggregation. Combine dynamic pruning of redundant edge weights and iteratively update the embedding and graph structure until convergence and optimization.

[0010] Preferably, the output process of the final node features is as follows: A1. Input the initial semantic features. The input feature of each device node i is the embedding representation obtained in the previous step : ; Among them, e i represents the multi-scale semantic embedding vector of device i; A2. For each node, according to its anomaly score score i ∈[0,1], introduce a perturbation signal: ; ; Among them, represents the denoising reconstruction error of the i-th device; represents the mean of the errors of all devices; represents the smoothing factor; represents the Gaussian noise perturbation; A3. Based on the constructed weighted adjacency matrix A, perform graph convolution propagation with edge weights: , after performing L layers, output the final node features: ; Among them, represents the feature vector of node j at the l-th layer; represents the feature vector of node i at the l+1-th layer; represents the value of the i-th row and j-th column in the weighted adjacency matrix, including self-loops, that is , I is the identity matrix; represents the non-linear activation function; represents the weight matrix at the l-th layer; represents the set of adjacent nodes of node i; represents the set of adjacent nodes including self-loops; z i represents the anomaly-sensitive embedding representation of the final device node i; represents the feature vector of node i at the L-th layer.

[0011] Preferably, the construction process of the anomaly-aware clustering objective function is as follows: B1. Based on the final embedding representation obtained after graph convolution propagation of the devices , initially use the K-Means method to divide the devices into M initial clusters: {C 1 , C 2, …, C m , …, C M}, each cluster center is represented as: ; Among them, represents the m-th clustering cluster; represents the center vector of the m-th clustering cluster; B2. Construct an anomaly-aware clustering objective function: ; Among them, represents the indicator function. If i and j belong to the same cluster C m , then take 1, otherwise take 0; represents the anomaly score of device i; represents the variance of the anomaly scores of all devices in the cluster where device i is located; , respectively represent the weights of the structural consistency loss term and the anomaly aggregation degree loss term; L 3 represents the objective function.

[0012] Preferably, the generation process of the final anomaly propagation prediction graph is as follows: S41. Establish a graph-based memory unit based on historical fault evolution behaviors, that is: extract fault sequences from historical fault events, including the sequence and time interval information of anomalies occurring in each device. All historical sequences will be abstracted into path information with time annotations, and then construct a fault gene graph G fault ; Determine the edge weights of this graph, that is, the weights: ; Among them, represents the propagation edge weight from node i to j; represents the number of historical fault events propagated from device i to device j; represents the total number of fault events for which device i has been the propagation starting point; represents the average propagation time delay from i to j; represents the time delay attenuation factor; S42. Project the currently detected abnormal device state onto the fault gene graph, and dynamically activate potential propagation paths by constructing a path activation mechanism jointly driven by anomaly intensity - structural intensity - propagation similarity, so as to realize the reconstruction and extension of the fault chain; Matching score function: ; Among them, P represents a candidate propagation path; represents the elastic dependence strength from i to j in the current anomaly graph; represents the anomaly state difference, that is, the difference degree between the anomaly values of devices i and j; S43. Estimate the fault propagation probability of other devices without anomalies based on the activated set of propagation paths, and construct an evolutionary graph structure G for the future predict : The propagation probability is estimated as: ; Set the risk threshold θ p , and construct the predictive propagation path: ; Generate the final abnormal propagation prediction graph G predict ; Among them, represents the predicted probability that the current device j without anomalies will have anomalies in the future; represents the set of devices that have been detected with anomalies currently; represents the time interval from the occurrence of the anomaly to the current moment for device i; represents the decay time factor; represents the set of future propagation edges obtained by speculation.

[0013] Preferably, the output process of the comprehensive risk level is as follows: S51. Based on the expert experience database in the power industry and the accident diagnosis standard document, abstract the abnormal mode characteristics and logical rule chains of the devices, and construct a structured knowledge graph K exp , and each expert rule is represented as a triple: ; Among them, r k represents the k-th expert experience rule; represents the abnormal feature associated with the k-th rule; represents the duration of this abnormal feature; represents the risk level indicated by this rule; Formalize the rule into a logical predicate function: ; Construct a discriminant vector driven by multiple rules: ; Among them, represents whether the k-th expert rule is satisfied after being applied to the input x; represents the duration of the abnormal behavior; represents a 0 / 1 vector composed of the discrimination results of all expert rules on the current device state, that is, the matching situation of the expert rule set for this data; S52. Construct a lightweight graph neural network structure, and jointly encode the node features, graph structure and propagation risk in the abnormal propagation prediction graph G predict into a low-dimensional semantic embedding h i , and perform risk classification based on the discriminant model : ; ; Among them, X represents the original input features of the node; represents the graph attention network; represents the embedding representation of the i-th device node; represents the downstream classification model with parameters θ, and outputs the abnormal risk level of the node; represents the abnormal classification result of the model for the i-th device, that is, the risk level predicted by the model; S53. Construct an objection-driven collaborative game model to perform interpretive correction on the conflict between expert rule discrimination and model inference results, and introduce a trust-weighted collaborative factor for unified decision-making discrimination: Collaborative discrimination score: ; ; Among them, represents the sigmoid function to normalize the expert score; represents the dynamic weight factor; represents the difference degree between the model and the expert output; represents the hyperparameter to control the collaborative sensitivity; represents the final comprehensive risk score of the i-th device; represents the expert rule weight vector; represents the expert rule matching result vector of the i-th device; represents the fusion discrimination score vector of the i-th device; S54. Output the final decision level : ; Among them, represents the comprehensive fusion score for determining that the device node i belongs to the c-th risk level.

[0014] Preferably, the comprehensive optimization objective is: C1. Use a multi-scale aggregation function to extract multi-layer features: ; ; C2. Define the structural consistency loss: ; C3. Define the comprehensive optimization objective: ; Among them, and respectively represent the original and denoised data structure features extracted at the l-th layer; represents the multi-scale structure feature extractor; L represents the number of multi-scale feature layers; represents the multi-scale structure preservation loss; represents the overall loss; represents the multi-scale consistency regulation factor, i.e., the structure preservation importance coefficient; represents the denoised data; D represents a three-dimensional tensor, which is used to uniformly express the multi-dimensional index data collected from multiple sensing nodes over a period of time.

[0015] The technical solution of the present invention: A power equipment abnormal detection and early warning system, which is used to execute the above-mentioned power equipment abnormal detection and early warning method, includes: A multi-source data acquisition module, which is used to acquire multi-source sensing data of the target power equipment; An edge computing and preprocessing module, which is used to perform initial data screening and denoising at the on-site terminal; An abnormal elastic clustering construction module, which is used to dynamically construct a weighted device relationship graph and perform abnormal-driven clustering analysis based on the spatio-temporal similarity between device operating states and the diffusion characteristics of abnormal signals, and identify potential device abnormal diffusion groups; A fault gene map evolution reasoning module, which is used to abstract the device historical abnormal data into fault gene vectors, construct a map and match the similar evolution paths corresponding to the current potential abnormality based on the graph attention mechanism, so as to realize intelligent analogy and trend prediction of new abnormalities; An expert rule and model collaborative discrimination module, which is used to encode node features and spread risks by constructing a structured knowledge graph and a graph attention network, design a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model predictions, and fuse logical predicate discrimination and classification results to output a comprehensive risk level; A multi-level early warning and knowledge writing-back module, which is used to output multi-level early warning signals according to the fusion judgment result, and write the features, evolution trajectories and response strategies in this round of detection into the system knowledge base in a structured manner to support subsequent model iteration and self-optimization, and realize the rule writing-back mechanism of the abnormal detection and early warning process.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: The present invention designs a power equipment abnormal detection and early warning system and method, which has the capabilities of dynamic association modeling, new abnormal prediction and self-evolution, can greatly reduce the false alarm and missed alarm rates, and improve the real-time performance and reliability of power equipment operation and maintenance. Specifically: (1)An adaptive denoising method for power multi-source perception data through multi-source perception data of target power equipment, based on semantic graph modeling and multi-scale residual feedback, a three-domain fusion denoising framework based on dynamic measurement of perception source correlation + discrimination of temporal semantic stability + suppression of graph structure context redundancy, using a semantic consistency residual regularization mechanism to dynamically adjust the noise suppression weights in different data modalities, realizing cross-modal collaborative denoising, and effectively improving data quality and processing timeliness; (2)Dynamically generate a weighted relationship graph based on the spatio-temporal similarity and abnormal diffusion characteristics of equipment, and combine graph neural networks to achieve anomaly-sensitive clustering and accurately identify potential abnormal groups; (3)Abstract historical anomalies into gene vectors and match the evolution paths to enhance the analog prediction ability for new anomalies; (4)Fuse dual-domain decisions through an objection-driven game mechanism to significantly improve the credibility and interpretability of anomaly discrimination; (5)Trigger hierarchical responses based on the evaluation of propagation influence, and structurally write back the anomaly features and evolution paths to the knowledge base to drive the collaborative optimization of the model and the rule base, forming a self-closed loop iterative system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a system architecture diagram of a power equipment anomaly detection and early warning system proposed by the present invention; Figure 2 It is a method flow diagram of a power equipment anomaly detection and early warning method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Example 1, as Figure 1 shown, a power equipment anomaly detection and early warning system proposed by the present invention includes: a multi-source data acquisition module, an edge computing and preprocessing module, an anomaly elastic clustering construction module, a fault gene map evolution reasoning module, an expert rule and model collaborative discrimination module, and a multi-level early warning and knowledge write-back module.

[0019] The multi-source data acquisition module collects multi-source perception data of the target power equipment, including but not limited to temperature, voltage, current, partial discharge, vibration, and noise; The edge computing and preprocessing module realizes data pre-screening and denoising at the on-site terminal; The anomaly elastic clustering construction module dynamically constructs a weighted equipment relationship graph based on the spatio-temporal similarity between equipment operating states and the diffusion characteristics of abnormal signals, and performs anomaly-driven clustering analysis to identify potential equipment abnormal diffusion groups; The fault gene map evolution reasoning module abstracts the equipment historical anomaly data into "fault gene vectors", constructs a map and matches the similar evolution paths corresponding to the current potential anomalies based on the graph attention mechanism to realize intelligent analogy and trend prediction of new anomalies; The expert rule and model collaborative discrimination module constructs a structured knowledge graph and a graph attention network to encode node features and propagate risks, designs a dual-domain collaborative game mechanism to dynamically adjust the weight differences between expert rules and model predictions, and fuses logical predicate discrimination and classification results to output the comprehensive risk level. The multi-level early warning and knowledge writing-back module outputs multi-level early warning signals according to the fusion judgment results, and writes the features, evolution trajectories, and response strategies in this round of detection into the system knowledge base in a structured manner to support subsequent model iteration and self-optimization, realizing the rule writing-back mechanism in the abnormal detection and early warning process.

[0020] Embodiment 2, as Figure 2 shown, a power equipment abnormal detection and early warning method proposed by the present invention is applied to a power equipment abnormal detection and early warning system proposed in Embodiment 1, and its specific implementation steps are as follows: S1. The multi-source data acquisition module acquires multi-source perception data of the target power equipment, including but not limited to operating status, environmental parameters, and historical maintenance record data, and transmits the multi-source perception data to the edge computing and preprocessing module.

[0021] S2. The edge computing and preprocessing module constructs an adaptive denoising method for power multi-source perception data based on semantic graph modeling and multi-scale residual feedback, a three-domain fusion denoising framework based on dynamic measurement of perception source correlation + temporal semantic stability discrimination + graph structure context redundancy suppression, uses a semantic consistency residual regularization mechanism to dynamically adjust the noise suppression weights in different data modalities to achieve cross-modal collaborative denoising, and introduces a "structure-aware multi-scale consistency function" to perform feedback correction on the denoising results, enabling the denoising process to have a self-supervised feedback mechanism. Its specific implementation process is as follows: S21. Acquire the multi-source perception data of the target power equipment, and define the data from different perception sources as: , and uniformly model it as a three-dimensional tensor: ; Among them, represents the jth perception index value (including but not limited to voltage, current, temperature, and humidity) of the ith sensor (i.e., perception source) at time t; S represents the total number of sensors, ; F represents the number of indicators collected by any sensor, ; T represents the total time series, ; represents the set of real numbers; D represents the three-dimensional tensor data structure composed of multi-source sensors: perception source × time × index dimension, with the dimensions unified; Exemplarily, assume that a power station site deploys 20 sensors (S = 20), each sampling 5 metrics (F = 5, including but not limited to voltage, current, power factor, frequency, temperature and humidity), the data collection period is recorded once per hour, and continuous collection is carried out for 48 hours (T = 48), then: , accordingly: construct a 3D data tensor that encompasses the complete perception data of the entire site over two days; S22. Construct a semantic association graph between perception sources, and dynamically evaluate the redundancy degree and cooperation degree between each perception source, specifically: S2201. Use the attention-aware pooling method to extract the semantic vector of each perception source at each moment: ; Among them, represents the data sequence of the i-th sensor within the time window [t - k, t]; represents the semantic feature representation (vector) of the data sampled by sensor i at time t; represents the semantic feature vector of the i-th sensor at time t; represents the attention-aware pooling operation; k represents the length of the historical window; S2202. Calculate the semantic similarity of every two perception sources i, j at the current moment t: ; Among them, represents the semantic similarity weight of the i-th and j-th sensors at time t; represents the transpose operation; represents the Euclidean norm, that is, the vector norm; represents the indicator function, if i = j, it is 1, otherwise it is 0, which is used to exclude the influence of self-similarity; Accordingly: abstract a high-order semantic representation based on the unified tensor D, and construct a dynamic semantic graph, which directly affects the subsequent "redundancy suppression mechanism" and "reference data estimation"; S23. Construct a mutation noise enhancement recognition mechanism to perform time series fluctuation detection and anomaly enhancement, and detect and enhance the influence weight of mutant noise in the time series trajectory, specifically: S2301. Calculate the sliding stability, and calculate the mean volatility within the sliding window for each perception dimension: ; Among them, represents the short-term fluctuation intensity of the current perception dimension; K represents the size of the sliding window; represents the j-th perception index value of the i-th perception source at time t; S2302 Set the dynamic enhancement coefficient : ; Among them, represents the dynamic enhancement coefficient, which is used to adjust the retention degree of this value; is used to regulate the fluctuation compression rate, ; is used to prevent excessive suppression, ; represents the natural exponential function; S24. To suppress "semantic shift type noise" and at the same time prevent the denoising process from damaging the high-dimensional features of the real signal, semantic residual regularization denoising is performed. Specifically: Define the denoising function , and generate denoised data ; Construct the semantic neighborhood reference data : ; Define the loss function with semantic residuals: ; Among them, represents the semantic neighborhood reference value of the i-th sensor; represents the set of neighbor nodes connected to i in the perception graph; represents the edge weight in the graph structure, that is, the semantic similarity weight between the i-th and j-th sensors at time t; represents the data after denoising, the denoised value of the j-th dimension of each sensor; represents the hyperparameter, which is used to balance the residual loss and the semantic consistency loss; represents the mapping function for extracting semantic embeddings, that is, the attention-aware pooling operation ; S25. Structural multi-scale feedback self-supervised adjustment, that is, to further enhance the overall structural consistency and avoid overfitting of the network to local interference. Specifically: Use the multi-scale aggregation function (graph convolution / convolution + pooling) to extract multi-layer features: ; ; Define the structural consistency loss: ; Comprehensive optimization objective: ; Among them, and respectively represent the original and denoised data structure features extracted from the l-th layer; represents the multi-scale structure feature extractor; L represents the number of multi-scale feature layers; represents the multi-scale structure preservation loss; represents the overall loss; Represents a multi-scale consistency regulator, i.e., the structural preservation importance coefficient; S26. Output the denoised data sequence of any power device i , and transmit the denoised data sequence of any power device i to the abnormal elastic clustering construction module.

[0022] S3. The abnormal elastic clustering construction module constructs a power device elastic association graph construction and adaptive clustering method based on multi-scale semantic reconstruction and abnormal driving mechanism. After completing the denoising of multi-source data, the goal is to depict the implicit structural relationship between devices in functional collaboration and abnormal conduction, and perform dynamic clustering accordingly. The specific implementation process is as follows: S31. Map the multi-dimensional perception state of each power device within the time window into a semantic representation vector, including multi-scale behavior semantics, that is: perform multi-scale semantic feature extraction e on the denoised data sequence of each device i i : ; where k represents F perception features at each moment, that is, the sequence length of the denoised data sequence of each device i ; F represents the number of perception features at each moment; represents the downsampled sequence, used to capture long-period behavior patterns; and represent the Gated Recurrent Unit, respectively modeling fine-grained and long-term dependence features; represents the convolutional-attention module, extracting local key time series change patterns; represents the concatenation operation; e i represents the multi-scale semantic embedding vector of device i; S32. Combine multi-scale semantic similarity and abnormal resonance degree to construct a multi-scale similarity graph between devices with elastic weighting, dynamically fuse functional association and abnormal interaction, and have a "redundancy suppression mechanism": ; ; ; ; where, represents the edge weight value between nodes i and j, that is, the elastic structure strength; represents the Sigmoid function; , , represent adjustable weights, used to balance the influence of similarity, resonance degree and redundancy; represents the multi-scale semantic embedding vector of device i e iThe semantic similarity (cosine similarity) with the multi-scale semantic embedding vector e of device j, where T here represents the vector transpose operation; j denotes the abnormal resonance degree, that is, the proportion of simultaneous abnormalities of two devices, and statistically counts the frequency of co-abnormalities in the past k time instants; k represents the time window length; t represents the current time instant; denotes the time variable in the sliding time window, which is used for summation iteration; and and respectively denote whether device i is abnormal at time τ and whether device j is abnormal at time τ; denotes the indicator function, and returns 1 when the condition holds; denotes the redundancy, the correlation of device sensed values; and denote the denoised sensed data sequence; denotes the cosine similarity; denotes the logical AND operator; Accordingly: construct the edge-weighted adjacency matrix A = { }, and output the elastic association graph between devices ; where V represents the set of device nodes, that is, the set of all monitored power device nodes; denotes the set of elastic edges, that is, the dynamic association relationship existing between power devices. In this embodiment, it is constructed by measuring the multi-scale time series embedding similarity (cosine similarity) and abnormal resonance factor between devices; S33. Abnormality-driven graph convolutional feature reconstruction, that is, on the constructed device elastic association graph utilize the graph neural network (GNN) to propagate structural information and fuse abnormal scoring signals to achieve semantic reconstruction and abnormal difference amplification of devices in the embedding space. Specifically: S3301. Initial semantic feature input: The input feature of each device node i is the embedding representation obtained in the previous step: ; where d represents the feature dimension; e i denotes the multi-scale semantic embedding vector of device i; S3302. Introduce the abnormal perturbation mechanism: For each node, according to its abnormal score score i ∈[0,1], introduce the perturbation signal: ; ; where, denotes the denoising reconstruction error of the i-th device; denotes the mean of the errors of all devices; denotes the smoothing factor, which controls the amplitude of non - linear enhancement; denotes the Gaussian noise perturbation, which simulates the structural sensitivity; S3303. Based on the constructed weighted adjacency matrix A, perform graph convolution propagation with edge weights: , after performing L layers, output the final node features: ; Among them, denotes the eigenvector of node j at the l - th layer; denotes the eigenvector of node i at the (l + 1)-th layer; denotes the value of the i - th row and j - th column in the weighted adjacency matrix, including self - loops (i.e., ), and I is the identity matrix; denotes the non - linear activation function; denotes the weight matrix at the l - th layer; denotes the set of adjacent nodes of node i; denotes the set of adjacent nodes including self - loops; z i denotes the abnormally sensitive embedding representation of the final device node i, with dimension d; denotes the eigenvector of node i at the L - th layer; S34. Form a structure - aware device function cluster in the device embedding space. The clustering result can actively respond to the abnormal conductivity characteristics, perform adaptive pruning optimization on the elastic graph structure between devices, and improve the structural sparsity and expression ability. Specifically: S3401. Initialize the structure - consistency clustering framework: After graph convolution propagation, the device obtains the final embedding representation: , initially use the K - Means method to divide the devices into M initial clusters: {C 1 , C 2 , …, C m , …, C M}, and the center of each cluster is represented as: ; Among them, denotes the m - th clustering cluster, which contains a group of device nodes with similar structures / abnormal characteristics; denotes the center vector of the m - th clustering cluster; S3402. Define the multi - objective optimization function and construct the abnormal - perception clustering objective function: ; Among them, denotes the indicator function. If i and j belong to the same cluster C m , then take 1, otherwise take 0; denotes the abnormal score of device i; denotes the variance of the abnormal scores of all devices in the cluster where device i is located; 、 respectively represent the weights of the structural consistency loss term and the abnormal aggregation degree loss term; L 3 represents the objective function; It should be noted that the objective function introduces a structural consistency mechanism (consistent edge connectivity) and abnormal aggregation degree (consistent abnormal trend), and actively tends to classify devices with "functionally similar + similar abnormal patterns" into one category, realizing the semantic collaborative fusion clustering of structural anomalies, and having the capabilities of anomaly-driven and structure interpretable: ; S3403. To prevent the redundant graph structure from misleading clustering, define the edge pruning constraint: ; ; Among them, represents the edge weight pruning threshold; represents the average value of all edge weights in the current graph; represents the standard deviation of all edge weights in the graph; represents the edge weight threshold adjustment factor; S3404. In each round of iteration, execute in sequence: update the embedding z based on the current graph structure i ; perform structure-aware anomaly clustering based on the embedding; correct the edge weights of the graph structure based on the clustering results; if the objective function converges or reaches the maximum number of rounds, then terminate.

[0023] S4. The fault gene map evolution reasoning module takes the elastic association graph between devices and the output results of anomaly-driven clustering as inputs, and proposes an evolution path reasoning method based on the fault gene map to mine the device-level fault behavior sequence template from historical behavior patterns, and combines the current abnormal state to realize the prediction and deduction of future potential propagation paths, and assist the dispatching control system to identify risk chains in advance. The specific implementation process is as follows: S41. Establish a graph-based memory unit based on historical fault evolution behaviors, that is: extract fault sequences from a large number of historical fault events, including the sequence and time interval information of each device's anomaly, and all historical sequences will be abstracted into path information with time annotations, and then construct the fault gene map G fault ; The edge weights (propagation weights) of this graph include both frequency factors and time delay factors, ensuring that the combined influence of historical occurrence probability and propagation efficiency is reflected: ; Among them, represents the propagation edge weight from node i to j, comprehensively reflecting the frequency and efficiency of propagation from device i to device j in history; represents the number of historical fault events propagated from device i to device j; Denote the total number of fault events where device i was the starting point of propagation; Denote the average propagation time delay from i to j; Denote the time delay attenuation factor (constant), which controls the attenuation degree of propagation delay on the edge weight. It is set through experiments and is empirically set to 10 minutes in this embodiment; S42. Project the currently detected abnormal device status onto the fault gene map, and dynamically activate potential propagation paths by constructing a path activation mechanism jointly driven by abnormal intensity - structural intensity - propagation similarity, so as to realize the reconstruction and extension of the fault chain; Matching score function: ; Among them, P represents a candidate propagation path; Denote the elastic dependence intensity from i to j in the current abnormal graph, which comes from the elastic association graph and is obtained by extracting the collaborative behavior pattern; Denote the abnormal state difference: the difference degree between the abnormal values of devices i and j; S43. Based on the set of the most likely propagation paths activated in the previous step, further estimate the fault propagation probability of other devices that have not shown abnormalities, and construct an evolutionary graph structure G for the future predict : The propagation probability is estimated as: ; Set the risk threshold θ p , and construct a predictive propagation path: ; Generate the final abnormal propagation prediction graph G predict ; Among them, Denote the predicted probability that the currently non - abnormal device j will become abnormal in the future; Denote the set of devices that have currently been detected as abnormal; Denote the time interval from the occurrence of the abnormality to the current moment for device i; Denote the decay time factor, which controls the propagation influence intensity of the most recent abnormality and prevents misleading caused by stale abnormalities; Denote the set of inferred future propagation edges.

[0024] S5. The expert rule and model collaborative discrimination module fuses the expert knowledge rule domain and the model algorithm inference domain, constructs a "human - machine knowledge" dual - channel and multi - perspective joint discrimination mechanism, and significantly improves the interpretability, accuracy, and controllability of discrimination through a dynamic game and correction mechanism of "structural constraint + behavior insight". The specific implementation process is as follows: S51. Based on the expert experience database of the power industry and the accident diagnosis standard document, abstract the abnormal mode characteristics and logical rule chains of devices, and construct a structured knowledge graph K exp, as an explicit constraint mechanism in the discrimination process, each expert rule is represented as a triple: ; where r k represents the k-th expert experience rule; represents the abnormal features associated with the k-th rule (including but not limited to high temperature, current fluctuation, frequency deviation); represents the duration of this symptom; represents the risk level indicated by this rule; Formalize the rule into a logical predicate function: ; Construct a discrimination vector driven by multiple rules: ; where represents whether the k-th expert rule is satisfied after being applied to the input x (1 means satisfied, 0 means not satisfied); represents the duration of abnormal behavior; represents a 0 / 1 vector composed of the discrimination results of all expert rules on the current device state, that is, the matching situation of the expert rule set to this data; S52. Construct a lightweight graph neural network structure (GAT, Graph Attention Networks, graph attention network), and jointly encode the node features, graph structure, and propagation risk in the abnormal propagation prediction graph G predict into a low-dimensional semantic embedding h i , and perform risk classification based on the discrimination model : ; ; where X represents the original input features of the node (including but not limited to abnormal indicators, propagation probability, fault chain depth); represents the graph attention network, which performs structure-sensitive semantic embedding learning on the node; represents the embedding representation of the i-th device node; represents the downstream classification model (MLP), with parameters θ, and outputs the abnormal risk level of this node; represents the abnormal classification result of the model for the i-th device, the risk level predicted by the model; S53. Dual-domain collaborative game discrimination mechanism: Design a dissent-driven collaborative game model to perform interpretive correction on the conflict between the expert rule discrimination and the model inference results, and introduce a trust-weighted collaborative factor for unified decision-making discrimination: Collaborative discrimination score: ; ; where Denotes the sigmoid function normalized expert score; Denotes the dynamic weight factor, determined by the degree of abnormality; Denotes the difference degree between the model and the expert output, measuring the divergence; Denotes the hyperparameter for controlling the collaborative sensitivity; Denotes the final comprehensive risk score of the i-th device; Denotes the expert rule weight vector, i.e., the importance of each rule; Denotes the expert rule matching result vector of the i-th device; Denotes the fusion discrimination score vector of the i-th device; Output the final decision level : ; Among them, Denotes the comprehensive fusion score that device node i is judged to be the c-th risk level (e.g., normal, warning, severe), which is jointly calculated by the expert rule domain and the model reasoning domain through a game weighting mechanism, reflecting the confidence level or score of device i belonging to the c-th category in the collaborative discrimination between the expert rule judgment result and the model prediction result.

[0025] S6. The multi-level warning and knowledge feedback module performs hierarchical response processing of abnormal results and a knowledge layer adaptive evolution feedback mechanism, triggers precise multi-level warning measures according to the abnormal level, and performs structured knowledge extraction and credible rule automatic generation based on the evolution trajectory and symptom combination, forming a continuous linkage and dynamic update closed-loop between the expert rule base and the intelligent model. The specific implementation process is as follows: S61. Based on the discrimination result, combined with the fault gene map, evaluate the propagation influence of each abnormal node to form a multi-dimensional risk scoring index L i , and map it to a multi-level warning level accordingly: ; Among them, L i Denotes the multi-dimensional propagation influence score of node i; d i Denotes the propagation depth of node i in the fault evolution path diagram; p i Denotes the participation frequency of the node in multiple propagation paths; w 1 , w 2 , w 3 Denotes the risk score weighting coefficient; Warning level mapping: ; Among them, Alert i Denotes the warning level; , , Denotes the multi-level warning trigger threshold; S62. After triggering an early warning, structurally analyze the abnormal features, triggering paths, and upstream and downstream associated devices of the early warning nodes, and automatically generate standardized knowledge entries: ; Among them, represents newly generated rule entries; Symptoms i represents a combination of abnormal symptoms (multi-dimensional features); T i represents the abnormal duration; CauseChain i represents the causal path chain that triggers this abnormality, which is inferred from the evolutionary path of the fault gene map and describes the whole process of this abnormality spreading from a certain initial node (such as a component of a device) in the system to the current node (i.e., the location where the fault occurs); Based on a comprehensive evaluation of symptom frequency, path consistency, and historical hit performance, introduce a rule credibility index , and adopt the following mechanism to determine whether to write back to the knowledge base: ; Among them, represents the credibility score of the newly generated rule; represents the expert rule base before the t-th update; represents the updated rule base; S63. To prevent the rule base from expanding and redundant accumulation, design a dynamic adjustment mechanism for rule credibility based on usage frequency and prediction consistency, and use this to guide the collaborative optimization of model parameters: If a newly added rule is frequently hit in the long term and is consistent with the model prediction, then increase its credibility weight w r ; If the rule performance is inconsistent or there are frequent misjudgments, then automatically trigger a weight reduction process; Accordingly: Under this mechanism, a dual-domain self-evolution closed-loop structure of model-assisted rule generation → rule-driven model constraint optimization → abnormal re-discrimination feedback is formed.

[0026] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to this. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the purpose of the present invention.

Claims

1. A method for detecting and warning abnormality of electric power equipment, characterized in that: The specific implementation steps include the following: S1. Collect the operating status, environmental parameters, and historical maintenance record data of the target power equipment; S2. Through the three-domain fusion denoising framework of dynamic measurement of data source relevance, judgment of temporal semantic stability and suppression of graph structure context redundancy, the noise suppression weights under different data modalities are dynamically adjusted to perform cross-modal collaborative denoising, and a structure-aware multi-scale consistency function is introduced to feedback and correct the denoising results; S3. Extract device features through multi-scale semantic embedding, build elastic association graph by combining semantic similarity, abnormal resonance and redundancy suppression, use graph neural network to fuse abnormal disturbance signals for feature reconstruction, and realize dynamic device grouping through multi-objective optimization clustering and edge pruning optimization; S4. Construct a fault gene map through historical fault data, integrate frequency and delay factors to calculate edge weights, activate potential paths based on the current abnormal state, estimate the future propagation probability of non-abnormal devices, and generate the final abnormal propagation prediction map; S5. By constructing a structured knowledge graph and graph attention network to encode node features and propagation risks, a dual-domain collaborative game mechanism is designed to dynamically adjust the weight difference between expert rules and model predictions, and the logical predicate discrimination and classification results are integrated to output a comprehensive risk level; S6. Generate multi-dimensional risk scores and map warning levels through communication influence assessment. After the warning is triggered, analyze the abnormal characteristics and propagation paths to generate knowledge items. Dynamically update the rule base based on credibility assessment. Design an adjustment mechanism driven by rule usage frequency and prediction consistency to optimize model parameters and form a dual-domain self-evolution closed loop.

2. A method for detecting and warning abnormality of electric power equipment according to claim 1, characterized in that: The denoising process of cross-modal collaborative denoising is as follows: S21. Use the attention-aware pooling method to extract the semantic vector of each perception source at each moment, and calculate the semantic similarity of each two perception sources i and j at the current moment t: ； ； in, represents the data sequence of the i-th sensor source in the time window [tk, t]; Represents the semantic feature representation of the data sampled by the perception source i at time t; represents the semantic feature vector of the i-th perception source at time t; represents the attention-aware pooling operation; k represents the length of the history window; represents the semantic similarity weight between the i-th and j-th perception sources at time t; Represents a transpose operation; represents the Euclidean norm; represents the indicator function; S22. Calculate the mean volatility within the sliding window for each perception dimension and set the dynamic enhancement coefficient: ； ； in, Indicates the short-term fluctuation intensity of the current perception dimension; K indicates the sliding window size; represents the jth perception index value of the i-th perception source at time t; represents the dynamic enhancement coefficient; Used to regulate the fluctuating compression ratio; Used to prevent over-compression; represents the natural exponential function; S23, generate denoised data through denoising function, construct semantic residual loss function based on semantic neighborhood weighted reference value, combine graph structure edge weight and attention pooling mapping function, balance residual and semantic consistency loss to optimize denoising process; S24, extract multi-layer features of original and denoised data through multi-scale aggregation functions, define structural consistency loss and integrate it into the comprehensive optimization goal, use regulatory factors to balance the multi-scale structure preservation strength and suppress local overfitting; S25. Output the denoised data sequence of any power device i.

3. A method for detecting and warning abnormality of electric power equipment according to claim 2, characterized in that: The semantic residual loss function is: ； ； ； in, represents the semantic neighborhood reference value of the i-th perception source; represents the set of neighbor nodes connected to node i in the perception graph; represents the edge weight in the graph structure, that is, the semantic similarity weight between the i-th and j-th perception sources at time t; Represents the denoised data, the denoised value of the jth dimension of each perception source; represents a hyperparameter; Represents the mapping function for extracting semantic embedding, i.e., the attention-aware pooling operation ; is the denoising function; represents denoised data; D represents a three-dimensional tensor, which is used to uniformly express the multi-dimensional indicator data collected from multiple sensing nodes over a period of time; S represents the total number of sensing sources, ; F represents the number of indicators collected by any perception source, ; T represents the total time series, .

4. The method for detecting and warning abnormality of electric power equipment according to claim 1, characterized in that: The grouping process of dynamic device grouping is as follows: S31, denoising data sequence for each device i Perform multi-scale semantic feature extraction i : ; in, represents the denoised data sequence of device i, with a time window length of k and F perceptual features at each moment; represents the downsampled sequence; and represents a gated recurrent unit; represents the convolution-attention module; Indicates the splicing operation; e i Represents the multi-scale semantic embedding vector of device i; S32. Combining multi-scale semantic similarity and abnormal resonance, construct an elastically weighted multi-scale similarity graph between devices, and construct an edge weight adjacency matrix A={ }, output the elastic association diagram between devices ; ； ； ； ； in, represents the edge weight between nodes i and j; Represents the Sigmoid function; , , Indicates adjustable weight; Represents the multi-scale semantic embedding vector e of device i i and device j multi-scale semantic embedding vector e j The semantic similarity of , where T represents the vector transposition operation; represents the abnormal resonance degree; t represents the current time; represents the time variable in the sliding time window; and They respectively indicate whether device i is abnormal at time τ and whether device j is abnormal at time τ; represents the indicator function; Indicates redundancy; and represents the denoised sensory data sequence; represents cosine similarity; Represents the logical AND operator; S33, taking the multi-scale semantic embedding vector as the initial feature, introducing the perturbation signal based on the anomaly score, fusing the adjacent node information through weighted graph convolution propagation, and outputting the final node feature after multiple layers of iteration; S34. Initialize the clustering framework through K-Means and construct an anomaly-aware clustering objective function that integrates structural consistency and anomaly aggregation. Combined with dynamic pruning of redundant edge weight constraints, iteratively update the embedding and graph structure until convergence optimization.

5. A method for detecting and warning abnormality of electric power equipment according to claim 4, characterized in that: The output process of the final node feature is: A1. Initial semantic feature input. The input feature of each device node i is the embedded representation obtained in the previous step. : ; Among them, e i Represents the multi-scale semantic embedding vector of device i; A2. For each node, score its abnormality i ∈[0,1], introduce a perturbation signal: ； ； in, represents the denoising reconstruction error of the i-th device; represents the mean of all device errors; represents the smoothing factor; represents Gaussian noise disturbance; A3. Based on the constructed weighted adjacency matrix A, perform edge-weighted graph convolution propagation: , after executing L layers, the final node features are output: ; in, represents the feature vector of node j at layer l; Represents the feature vector of node i at the l+1th layer; Represents the value of the i-th row and j-th column in the weighted adjacency matrix, including self-loops, that is, , I is the identity matrix; represents a nonlinear activation function; represents the weight matrix of the lth layer; Represents the set of adjacent nodes of node i; represents the set of adjacent nodes that contain self-loops; z i represents the abnormality-sensitive embedding representation of the final device node i; Represents the feature vector of node i at the Lth layer.

6. A method for detecting and warning abnormality of electric power equipment according to claim 5, characterized in that: The process of constructing the anomaly-aware clustering objective function is as follows: B1. The final embedding representation is obtained after the device-based graph convolution propagation , the K-Means method is used to initially divide the devices into M initial clusters: {C1, C2,…, C m ,…,C M }, each cluster center is represented as: ; in, represents the mth cluster; Represents the center vector of the mth cluster; B2. Construct anomaly-aware clustering objective function: ； in, Represents the indicator function, if i and j belong to the same cluster C m , then take 1, otherwise take 0; represents the anomaly score of device i; represents the variance of abnormal scores of all devices in the cluster where device i is located; , They represent the weight of the structural consistency loss term and the weight of the abnormal aggregation loss term respectively; L3 represents the objective function.

7. The method for detecting and warning abnormality of electric power equipment according to claim 1, characterized in that: The final anomaly propagation prediction graph is generated as follows: S41. Establish a graph memory unit based on the historical fault evolution behavior, that is, extract the fault sequence from the historical fault events, including the order and time interval information of the abnormalities of each device. All historical sequences will be abstracted into path information with time annotations, and then construct the fault gene map G fault ; Determine the edge weight of the graph: ; in, represents the propagation edge weight from node i to j; represents the number of historical failure events propagated from device i to device j; represents the total number of fault events in which device i has been the starting point of propagation; represents the average propagation time delay from i to j; represents the delay attenuation factor; S42, projecting the currently detected abnormal equipment status to the fault gene map, dynamically activating the potential propagation path by constructing a path activation mechanism jointly driven by abnormal strength, structural strength and propagation similarity, thereby realizing the reconstruction and extension of the fault chain; Matching score function: ; Among them, P represents a candidate propagation path; Indicates the elastic dependence strength from i to j in the current anomaly graph; It represents the abnormal state difference, that is, the difference between the abnormal values ​​of equipment i and j; S43. Based on the set of activated propagation paths, estimate the fault propagation probability of other devices without abnormalities and construct a future-oriented evolutionary graph structure G predict : The transmission probability is estimated as: ; Set the risk threshold θ p , construct a predictive propagation path: ; Generate the final anomaly propagation prediction graph G predict ; in, It represents the predicted probability that the currently normal device j will have an abnormality in the future; Indicates the device collection that has currently detected an anomaly; Indicates the time interval from the occurrence of the abnormality of device i to the current moment; represents the decay time factor; Represents the set of inferred future propagation edges.

8. The method for detecting and warning abnormality of electric power equipment according to claim 7, characterized in that: The output process of the comprehensive risk level is: S51. Based on the power industry expert experience database and accident diagnosis standard documents, the abnormal mode characteristics and logical rule chains of the equipment are abstracted to construct a structured knowledge graph K exp , each expert rule is represented as a triple: ; Among them, r k represents the kth expert experience rule; represents the abnormal feature associated with the kth rule; Indicates the duration of the abnormal feature; Indicates the risk level indicated by the rule; Formalize the rules as logical predicate functions: ; Construct a multi-rule driven discriminant vector: ; in, Indicates whether the logic of the input x is satisfied after applying the kth expert rule; Indicates the duration of abnormal behavior; Represents a 0 / 1 vector consisting of the judgment results of all expert rules on the current device state, that is, the matching status of the expert rule set to the data; S52. Construct a lightweight graph neural network structure to predict the anomaly propagation graph G predict The node features, graph structure and propagation risk in the graph are jointly encoded into a low-dimensional semantic embedding h i , and based on the discriminant model To classify the risks: ； ； Among them, X represents the original input feature of the node; Representation graph attention network; represents the embedded representation of the i-th device node; represents the downstream classification model, with parameter θ, and outputs the abnormal risk level of the node; represents the abnormal classification result of the model for the i-th device, that is, the risk level predicted by the model; S53. Construct a collaborative game model based on objection-driven, interpret the conflict between expert rule judgment and model reasoning results, and introduce trust-weighted collaborative factors for unified decision judgment: Co-discrimination score: ； ； in, Represents the sigmoid function normalized expert score; represents the dynamic weight factor; Indicates the difference between the model and the expert output; represents the hyperparameter controlling the co-sensitivity; represents the final comprehensive risk score of the i-th device; represents the expert rule weight vector; represents the expert rule matching result vector of the i-th device; represents the fusion discriminant score vector of the i-th device; S54, output final decision level : ; in, Indicates the comprehensive fusion score of device node i being judged as the cth risk level.

9. The method for detecting and warning abnormality of electric power equipment according to claim 2, characterized in that: The comprehensive optimization goal is: C1. Use multi-scale aggregation functions Extracting multi-layer features: ； ； C2. Define structural consistency loss: ; C3. Define comprehensive optimization goals: ; in, and Respectively represent the structural features of the original and denoised data extracted from the lth layer; represents a multi-scale structural feature extractor; L represents the number of multi-scale feature layers; represents the multi-scale structure preserving loss; Indicates overall loss; represents the multi-scale consistency regulation factor; represents denoised data; D represents a three-dimensional tensor, which is used to uniformly express the multi-dimensional indicator data collected from multiple sensing nodes within a period of time.

10. An electric power equipment abnormality detection and early warning system, which is used to execute an electric power equipment abnormality detection and early warning method according to any one of claims 1 to 9, characterized in that: include: A multi-source data acquisition module is used to collect multi-source perception data of target power equipment; Edge computing and preprocessing modules are used to implement data screening and denoising at the on-site terminal; Abnormal elastic clustering construction module, which is used to dynamically build a weighted device relationship graph and perform abnormality-driven clustering analysis based on the spatiotemporal similarity between device operating states and the diffusion characteristics of abnormal signals, and identify potential device abnormality diffusion groups; The fault gene graph evolution reasoning module is used to abstract the historical abnormal data of the equipment into fault gene vectors, build a graph, and match similar evolution paths corresponding to the current potential abnormalities based on the graph attention mechanism, so as to realize intelligent analogy and trend prediction of new abnormalities. The expert rule and model collaborative discrimination module is used to encode node features and propagation risks by building a structured knowledge graph and graph attention network, design a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model predictions, and integrate logical predicate discrimination and classification results to output a comprehensive risk level; The multi-level warning and knowledge write-back module is used to output multi-level warning signals based on the fusion judgment results, and write the characteristics, evolution trajectory and response strategy in this round of detection into the system knowledge base in a structured manner, support subsequent model iteration and self-optimization, and realize the rule write-back mechanism of the anomaly detection and warning process.

Images