A power equipment anomaly detection and early warning system and method

Abstract

The application relates to the technical field of power equipment early warning, in particular to a power equipment anomaly detection and early warning system and method; specifically, multi-source sensing data of a target power equipment is acquired; cross-modal collaborative denoising is carried out based on semantic graph modeling and multi-scale residual feedback; a weighted relation graph is constructed by fusing equipment space-time similarity and abnormal diffusion characteristics, and graph neural network is used for abnormal sensitive clustering to identify a potential abnormal group; historical anomalies are abstracted into fault gene vectors, an evolution path is matched, and a new abnormal propagation trend is predicted; expert knowledge graph and graph neural network are introduced, two-domain decisions are dynamically fused through objection-driven game, and early warning credibility is improved; a hierarchical early warning is triggered according to propagation influence, abnormal characteristics and evolution paths are written back to a knowledge base in a structured manner, a model and a rule base are cooperatively optimized, and a detection-early warning-optimization self-closed loop system is formed. The application solves the problems of dynamic correlation modeling, new anomaly prediction and decision credibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment early warning, in particular to a power equipment anomaly detection and early warning system and method. BACKGROUND

[0002] As a key link to ensure the safe and stable operation of the power grid, power equipment anomaly detection has long been faced with challenges such as difficulty in modeling dynamic correlation of multi-source data under complex working conditions, insufficient early prediction ability of new anomalies, and limited decision-making credibility. Traditional methods mostly rely on single sensor data or static clustering analysis, which is difficult to effectively capture the abnormal conduction characteristics in the coordinated operation of equipment groups, and is prone to misjudgment and missed judgment under low-quality data interference. With the upgrading of the intelligentization of the power system, the equipment state monitoring data presents characteristics such as multi-modal, high dimension, and strong time sequence correlation. The existing methods based on rule engine or isolated model have significant limitations in abnormal propagation path reasoning, cross-device abnormal resonance analysis, and dynamic knowledge evolution, and are difficult to meet the needs of real-time early warning and closed-loop optimization.

[0003] The Chinese invention patent with publication number CN118898342B discloses a power equipment safety early warning method and system based on multi-modal data, which includes an environmental attribute deviation vector, an audio attribute deviation vector, and a running attribute deviation vector, constructs a spatial reference distribution coordinate, performs case retrieval in a power equipment safety management case library, constructs a recursive network tree for node weight distribution, and obtains a node weight distribution result. According to the node weight distribution result, the recursive network tree is inversely grown and aggregated to obtain a power equipment anomaly index. If the index is greater than or equal to a power equipment anomaly index threshold, power equipment early warning is executed. This solves the technical problem in the prior art that due to the lack of a comprehensive analysis scheme for multi-modal data, it is difficult to effectively utilize multi-attribute data for comprehensive judgment, thereby making it difficult to achieve precise early warning of power equipment.

[0004] In addition, the fragmented application of expert experience and data-driven models results in a lack of explainability in the anomaly discrimination process, especially in new abnormal or low confidence scenarios, making it difficult to form a credible decision. To address the above problems, it is necessary to build an intelligent detection system that integrates multi-source perception, dynamic correlation modeling, and knowledge self-evolution to achieve precise life cycle management and control of power equipment anomalies. SUMMARY

[0005] The present application aims to solve the problems in the background art by providing a power equipment anomaly detection and early warning system and method.

[0006] The technical solution of the present application is a power equipment anomaly detection and early warning method, which includes the following specific implementation steps:

[0007] S1, collect the running state, environmental parameters, and historical maintenance record data of the target power equipment;

[0008] S2, through the three-domain fusion denoising framework of dynamic data source correlation measurement, time sequence semantic stability discrimination and graph structure context redundancy suppression, dynamically adjusting the noise suppression weight under different data modalities, cross-modal collaborative denoising, introducing a structure-aware multi-scale consistency function to feedback and correct the denoising result;

[0009] S3, through multi-scale semantic embedding to extract device features, combined with semantic similarity, abnormal resonance degree and redundancy suppression to construct an elastic correlation graph, using graph neural network to fuse abnormal disturbance signals for feature reconstruction, and through multi-objective optimization clustering and edge pruning optimization to realize dynamic device grouping;

[0010] S4, through the construction of historical fault data to build a fault gene map, fuse frequency and time delay factors to calculate edge weight, combine the current abnormal state to activate the potential path, estimate the future propagation probability of non-abnormal devices, and generate the final abnormal propagation prediction graph;

[0011] S5, through the construction of structured knowledge graph and graph attention network to encode node features and propagate risks, design a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model prediction, fuse logical predicate discrimination and classification results to output comprehensive risk level;

[0012] S6, through the generation of multi-dimensional risk score and the mapping of warning level by propagation influence evaluation, analyze abnormal features and propagation path to generate knowledge items after triggering the warning, dynamically update the rule library based on credibility evaluation, 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.

[0013] Preferably, the denoising process of cross-modal collaborative denoising is as follows:

[0014] S21, the semantic vector of each perception source at each time is extracted by using attention perception pooling method, and the semantic similarity of each two perception sources i, j at current time t is calculated:

[0015] ;

[0016] ;

[0017] Wherein, represents the data sequence of the i-th perception source in the time window [t-k, t]; represents the semantic feature representation of the sampled data of perception source i at time t; represents the semantic feature vector of the i-th perception source at time t; represents the attention perception pooling operation; k represents the length of the historical window; represents the semantic similarity weight of the i-th and j-th perception sources at time t; here Indicates the transpose operation; Represents the Euclidean norm; Indicates an indicator function;

[0018] S22. Calculate the mean volatility within the sliding window for each perception dimension and set the dynamic enhancement coefficient:

[0019] ;

[0020] ;

[0021] in, This represents the short-term fluctuation intensity of the current perception dimension; K represents the sliding window size. This represents the j-th sensing index value of the i-th sensing source at time t; Indicates the dynamic enhancement coefficient; Used to regulate volatility compression rate; Used to prevent excessive suppression; Represents the natural exponential function;

[0022] S23. Generate denoised data through a denoising function, construct a semantic residual loss function based on semantic neighborhood weighted reference values, and combine graph structure edge weights and attention pooling mapping function to balance residual and semantic consistency loss to optimize the denoising process.

[0023] S24. Extract multi-layer features from the original and denoised data using a multi-scale aggregation function, define structural consistency loss and integrate it into the comprehensive optimization objective, and use a regulatory factor to balance the multi-scale structure to maintain strength and suppress local overfitting.

[0024] S25. Output the denoised data sequence of any power device i.

[0025] Preferably, the semantic residual loss function is:

[0026] ;

[0027] ; ;

[0028] in, This represents the semantic neighborhood reference value of the i-th perceptual source; Let i represent the set of neighboring nodes connected to node i in the perception graph; The edge weights in the graph structure represent the semantic similarity weights of the i-th and j-th perceptual sources at time t. This represents the denoised data, specifically the denoised value of the j-th dimension for each sensing source; Indicates hyperparameters; a mapping function representing the extracted semantic embedding, i.e., an attention-aware pooling operation ; is a denoising function; denotes the denoised data; D denotes a three-dimensional tensor, which is used to uniformly express multi-dimensional index data collected by multiple perception nodes over a period of time; S denotes the total number of perception sources, ; F denotes the number of indexes collected by any perception source, ; T denotes the total time series, .

[0029] Preferably, the grouping process of the dynamic device grouping is as follows:

[0030] S31, for each device i, the denoised data sequence Multi-scale semantic feature extraction e i : ;

[0031] wherein, denotes the denoised data sequence of device i, the time window length is k, and there are F perception features at each time; denotes the down-sampled sequence; and denotes a gated recurrent unit; denotes a convolution-attention module; denotes a concatenation operation; e i denotes the multi-scale semantic embedding vector of device i;

[0032] S32, combining multi-scale semantic similarity and abnormal resonance degree, constructing an elastic weighted multi-scale similarity graph between devices, constructing an edge weight adjacency matrix A={ }, outputting an elastic correlation graph between devices ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] wherein, denotes the edge weight value between nodes i and j; denotes a Sigmoid function; , , denotes an adjustable weight; denotes the multi-scale semantic embedding vector e i of device i and the multi-scale semantic embedding vector ej semantic similarity of the i-th device, where T represents the vector transpose operation; represents the abnormal resonance degree; t represents the current time; 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 represents the denoised perception data sequence; represents the cosine similarity; represents the logical AND operator;

[0038] S33, taking the multi-scale semantic embedding vector as the initial feature, introducing a perturbation signal based on the abnormal score, propagating and fusing the adjacent node information through weighted graph convolution, and outputting the final node feature through multi-layer iteration;

[0039] S34, initializing the clustering framework through K-Means and constructing an abnormal perception clustering objective function that fuses structure consistency and abnormal aggregation degree, combining dynamic pruning and redundant edge weight constraints, and iteratively updating the embedding and graph structure until convergence optimization.

[0040] Preferably, the output process of the final node feature is:

[0041] A1, initial semantic feature input, the input feature of each device node i is the embedding representation obtained in the previous step : ;

[0042] wherein e i represents the multi-scale semantic embedding vector of device i;

[0043] A2, for each node, according to its abnormal score score i ∈[0,1], a perturbation signal is introduced:

[0044] ;

[0045] ;

[0046] wherein, represents the denoising reconstruction error of the i-th device; represents the mean of all device errors; represents the smoothing factor; represents the Gaussian noise disturbance;

[0047] A3, based on the constructed weighted adjacency matrix A, perform edge-weighted graph convolution propagation: , execute L layer, output the final node feature: ;

[0048] where, represents the feature vector of node j at the lth 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 the self-loop, i.e. , I is the unit 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 containing the self-loop; z i represents the anomaly-sensitive embedding representation of the final device node i; represents the feature vector of node i at the Lth layer.

[0049] Preferably, the anomaly-aware clustering objective function construction process is:

[0050] B1, based on the final embedding representation of the device after graph convolution propagation , the K-Means method is used to divide the devices into M initial clusters: {C1, C2,…, C m ,…,C M}, and each cluster center is represented as: ;

[0051] where, represents the mth clustering cluster; represents the center vector of the mth clustering cluster;

[0052] B2, construct an anomaly-aware clustering objective function:

[0053] ;

[0054] where, represents an indicator function, which takes 1 if i and j belong to the same cluster C m , otherwise 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 weight of the structure consistency loss term and the weight of the anomaly aggregation loss term; L3 represents the objective function.

[0055] Preferably, the generation process of the final anomaly propagation prediction graph is as follows:

[0056] S41, based on the historical failure evolution behavior to establish a graph memory unit, that is, to extract the failure sequence from the historical failure events, including the order of abnormality of each device and the time interval information, all historical sequences will be abstracted as path information with time annotation, and then the failure gene graph G is constructed fault ;

[0057] Determine the edge weight of the graph, that is, the weight: ;

[0058] Wherein, represents the propagation edge weight of node i to j; represents the number of historical failure events from device i to device j; represents the total number of failure events in which device i has served as the starting point of propagation; represents the average propagation time delay from i to j; represents the time delay attenuation factor;

[0059] S42, project the current detected abnormal device state to the failure gene graph, and dynamically activate the potential propagation path by constructing an abnormal intensity-structure intensity-propagation similarity joint driven path activation mechanism, so as to realize the reconstruction and extension of the failure chain;

[0060] Matching score function: ;

[0061] Wherein, P represents a candidate propagation path; represents the elastic dependence intensity of i to j in the current abnormal graph; represents the abnormal state difference, that is, the difference between the abnormal values of devices i and j;

[0062] S43, based on the activated propagation path set, estimate the failure propagation probability of other abnormal devices, and construct an evolution graph structure G predict :

[0063] The propagation probability is estimated as: ;

[0064] Set the risk threshold θ p , and construct a predictive propagation path: ;

[0065] Generate the final abnormal propagation prediction graph G predict :

[0066] Wherein, represents the prediction probability of the current non-abnormal device j in the future; represents the set of devices that have been detected to be abnormal; represents the time interval from the occurrence of abnormality of device i to the current time. denotes the decay time factor; denotes the predicted future propagation edge set.

[0067] Preferably, the output process of the comprehensive risk level is:

[0068] S51, based on the power industry expert experience library and the accident diagnosis standard document, abstract the device abnormal mode characteristics and the logical rule chain, and construct into a structured knowledge graph K exp Each expert rule is represented as a triple: ;

[0069] Wherein, r k denotes the kth expert experience rule; denotes the abnormal feature associated with the kth rule; denotes the duration of the abnormal feature; denotes the risk level indicated by the rule;

[0070] Formalize the rule into a logical predicate function: ;

[0071] Construct a multi-rule driven discriminant vector: ;

[0072] Wherein, denotes whether the logic of the kth expert rule is satisfied after applying the input x; denotes the duration of abnormal behavior; denotes a 0 / 1 vector composed of the discriminant results of the current device state by all expert rules, i.e. the matching condition of the expert rule set for the data;

[0073] S52, construct a light graph neural network structure, and jointly encode the node features, graph structure and propagation risk in the abnormal propagation prediction graph G predict into low-dimensional semantic embedding h i , and perform risk classification based on the discriminant model :

[0074] ;

[0075] ;

[0076] Wherein, X denotes the original input feature of the node; denotes the graph attention network; denotes the embedding representation of the ith device node; denotes the downstream classification model, the parameter is θ, and the output is the abnormal risk level of the node; denotes the abnormal classification result of the model for the ith device, i.e. the risk level predicted by the model;

[0077] S53, construct a collaborative game model based on the objection, explain the conflict between the expert rule and the model inference result, and introduce a trust weighted collaborative factor for unified decision making:

[0078] Collaborative decision score:

[0079] ;

[0080] ;

[0081] wherein, represents the sigmoid function normalization expert score; represents the dynamic weight factor; represents the difference between the model and the expert output; represents the super parameter for controlling the sensitivity of collaboration; represents the final comprehensive risk score of the ith device; represents the expert rule weight vector; represents the expert rule matching result vector of the ith device; represents the fusion decision score vector of the ith device;

[0082] S54, output the final decision level : ;

[0083] wherein, represents the comprehensive fusion score of the device node i being judged as the cth risk level.

[0084] Preferably, the comprehensive optimization target is:

[0085] C1, use a multi-scale aggregation function extract multi-layer features:

[0086] ; ;

[0087] C2, define the structure consistency loss: ;

[0088] C3, define the comprehensive optimization target: ;

[0089] wherein, and respectively represent the original and denoised data structure features extracted at the lth layer; represents a 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 multiscale consistency regulation factor, i.e. the structure maintaining importance coefficient; represents the denoised data; D represents a three-dimensional tensor, i.e. a unified expression of multi-dimensional index data collected by multiple perception nodes in a period of time.

[0090] The technical scheme of the present application: a power equipment anomaly detection and early warning system for executing the above-mentioned power equipment anomaly detection and early warning method, comprising:

[0091] A multi-source data acquisition module is configured to acquire multi-source perception data of the target power equipment.

[0092] An edge computing and preprocessing module is configured to realize data preliminary screening and denoising at the field terminal.

[0093] An anomaly elasticity clustering construction module is configured to dynamically construct a weighted device relationship graph and perform anomaly-driven clustering analysis based on the spatiotemporal similarity between device operating states and the diffusion characteristics of abnormal signals, and identify potential device abnormal diffusion groups.

[0094] A fault gene graph evolution reasoning module is configured to abstract device historical abnormal data into a fault gene vector, construct a graph, and match a similar evolution path corresponding to the current potential anomaly based on a graph attention mechanism, so as to realize intelligent analogy and trend prediction of new anomalies.

[0095] An expert rule and model collaborative discrimination module is configured to encode node features and propagate 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.

[0096] A multi-level early warning and knowledge backwriting module is configured to output multi-level early warning signals according to the fusion judgment result, and write the features, evolution trajectory and response strategy in the current detection into the system knowledge base in a structured manner to support subsequent model iteration and self-optimization, and realize a rule backwriting mechanism for the anomaly detection and early warning process.

[0097] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0098] The present application designs a power equipment anomaly detection and early warning system and method, which has dynamic correlation modeling, new anomaly prediction and self-evolution capabilities, can greatly reduce the false alarm and missed alarm rates, and improves the real-time performance and reliability of power equipment operation and maintenance.

[0099] (1) Through the multi-source sensing data of the target power equipment, and the adaptive denoising method of power multi-source sensing data based on semantic graph modeling and multi-scale residual feedback, the three-domain fusion denoising framework based on the dynamic measurement of sensing source correlation + temporal semantic stability discrimination + graph structure context redundancy suppression is used to dynamically adjust the noise suppression weight under different data modes by utilizing the semantic consistency residual regularization mechanism, thereby realizing cross-modal collaborative denoising and effectively improving data quality and processing timeliness.

[0100] (2) Based on the spatiotemporal similarity of equipment and the characteristics of abnormal diffusion, a weighted relationship graph is dynamically generated, and combined with a graph neural network to realize abnormal sensitive clustering and accurately identify potential abnormal groups;

[0101] (3) Abstract historical anomalies into gene vectors and match evolutionary paths to enhance the ability to predict new anomalies by analogy;

[0102] (4) By integrating dual-domain decision-making through objection-driven game mechanism, the credibility and explanatory power of anomaly detection are significantly improved;

[0103] (5) Based on the assessment of the impact of the dissemination, a graded response is triggered, and the abnormal characteristics and evolution path are written back to the knowledge base in a structured manner to drive the collaborative optimization of the model and the rule base, forming a self-closed loop iterative system. Attached Figure Description

[0104] Figure 1 This is a system architecture diagram of a power equipment anomaly detection and early warning system proposed in this invention;

[0105] Figure 2 This is a flowchart of a method for detecting and warning of abnormalities in power equipment proposed in this invention. Detailed Implementation

[0106] Example 1, as Figure 1 As shown, the present invention proposes a power equipment anomaly detection and early warning system, which 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 writing module.

[0107] The multi-source data acquisition module collects multi-source sensing data from the target power equipment, including but not limited to temperature, voltage, current, partial discharge, vibration, and noise.

[0108] The edge computing and preprocessing module performs initial data screening and noise reduction at the field terminal.

[0109] The abnormal elastic clustering construction module dynamically constructs a weighted device relationship graph and performs anomaly-driven clustering analysis based on the spatiotemporal similarity between device operating states and the diffusion characteristics of abnormal signals to identify potential device anomaly diffusion groups.

[0110] a failure gene atlas evolution reasoning module, which abstracts device historical abnormal data as a "failure gene vector", constructs an atlas, and matches a similar evolution path corresponding to a current potential abnormality based on a graph attention mechanism, to realize intelligent analogy and trend prediction of new abnormalities;

[0111] an expert rule and model collaborative judgment module, which codes node features and propagates risks through constructing a structured knowledge graph and a graph attention network, designs a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model predictions, and fuses logical predicate judgment and classification results to output a comprehensive risk level;

[0112] a multi-level early warning and knowledge backwriting module, which outputs multi-level early warning signals according to the fusion judgment results, and writes the features, evolution trajectories and response strategies in the current round of detection into the system knowledge base in a structured manner to support subsequent model iteration and self-optimization, and realizes a rule backwriting mechanism for the abnormality detection and early warning process.

[0113] Embodiment two, as shown in the figure, the present application proposes a power equipment abnormality detection and early warning method, which is applied to the power equipment abnormality detection and early warning system proposed in embodiment one, and the specific implementation steps are as follows: Figure 2

[0114] S1, the multi-source data acquisition module acquires multi-source perception data of the target power equipment, including but not limited to operating state, environmental parameters, historical maintenance record data, and transmits the multi-source perception data to the edge computing and preprocessing module.

[0115] S2, the edge computing and preprocessing module constructs a power multi-source perception data adaptive denoising method based on semantic graph modeling and multi-scale residual feedback, a three-domain fusion denoising framework based on perception source correlation dynamic measurement + time sequence semantic stability discrimination + graph structure context redundancy suppression, uses a semantic consistency residual regular mechanism to dynamically adjust the noise suppression weight under different data modalities, realizes cross-modal collaborative denoising, introduces a "structure-aware multi-scale consistency function" to feedback and correct the denoising result, so that the denoising process has a self-supervised feedback mechanism, and the specific implementation process is as follows:

[0116] S21, acquire multi-source perception data of the target power equipment, and define data from different perception sources as: , and model them as a three-dimensional tensor: ;

[0117] wherein, 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 real set; D represents a three-dimensional tensor data structure composed of multiple source sensors: perception source x time x index dimension, which has been unified dimension;

[0118] For example, assuming that a certain power site is deployed with 20 sensors (S = 20), each sampling 5 indexes (F = 5, including but not limited to voltage, current, power factor, frequency, temperature and humidity), the data collection period is 1 record per hour, and the continuous collection is 48 hours (T = 48), then: Accordingly: a 3D data tensor is constructed, which includes the complete perception data of the entire site within two days;

[0119] S22, construct a semantic association graph between perception sources, dynamically evaluate the redundancy and synergy degree between each perception source, specifically:

[0120] S2201, adopt attention perception pooling method to extract semantic vector of each perception source at each time:

[0121] ;

[0122] wherein, 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 perception pooling operation; k represents the length of the history window;

[0123] S2202, calculate the semantic similarity of each two perception sources i, j at the current time t: ;

[0124] wherein, represents the semantic similarity weight of the i-th and j-th sensors at time t; represents the transpose operation; represents the Euclidean norm, i.e. the vector length; represents the indicator function, which is 1 if i = j, otherwise 0, used to exclude the influence of self-similarity;

[0125] Accordingly: on the basis of the unified tensor D, high-order semantic representation is abstracted, and a dynamic semantic graph is constructed, which directly affects the subsequent "redundancy suppression mechanism" and "reference data estimation";

[0126] S23, construct a mutation noise enhancement identification mechanism, perform time series fluctuation detection and abnormal enhancement, detect and enhance the influence weight of mutation noise in time series track, specifically:​

[0127] S2301, calculate the sliding stability, calculate the mean volatility of each perception dimension in its sliding window:

[0128] ;

[0129] wherein, represents the short-term fluctuation intensity of the current perception dimension; K represents the size of the sliding window; represents the jth perception index value of the ith perception source at time t;

[0130] S2302 Set dynamic enhancement coefficient : ;

[0131] wherein, represents the dynamic enhancement coefficient, used to adjust the degree of reservation of the value; used to regulate the compression rate of volatility, ; used to prevent excessive suppression, ; represents the natural exponential function;

[0132] S24, for inhibiting "semantic shift type noise", while preventing the damage of high-dimensional characteristics of the real signal in the denoising process, semantic residual regularization denoising is performed, specifically:

[0133] Define the denoising function , generate denoised data ;

[0134] Construct semantic neighborhood reference data : ;

[0135] Define the loss function with semantic residual:

[0136] ;

[0137] wherein, represents the semantic neighborhood reference value of the ith sensor; represents the neighbor node set connected with i in the perception graph; represents the edge weight in the graph structure, that is, the semantic similarity weight of the ith and jth sensors at time t; represents the denoised data, the jth dimension denoised value of each sensor; represents a hyperparameter used to balance the residual loss and semantic consistency loss; represents the mapping function for extracting semantic embedding, that is, attention perception pooling operation ;

[0138] S25, structure multi-scale feedback self-supervised adjustment, that is, further enhance the overall structure consistency, avoid network over-fitting to local interference, specifically:

[0139] Using multi-scale aggregation function (figure convolution / convolution + pooling) to extract multi-layer features:

[0140] ; ;

[0141] Define structure consistency loss: ;

[0142] Comprehensive optimization goal: ;

[0143] Wherein, and Respectively, the original and denoised data structure features extracted by the lth layer; Indicates the multi-scale structure feature extractor; L represents the number of multi-scale feature layers; Indicates the multi-scale structure preservation loss; Indicates the total loss; Indicates the multi-scale consistency regulation factor, that is, the structure preservation importance coefficient;

[0144] S26, output the denoised data sequence of any power equipment i And the denoised data sequence of any power equipment i Is transmitted to the abnormal elasticity clustering construction module.

[0145] S3, the abnormal elasticity clustering construction module constructs the power equipment elasticity association graph construction and adaptive clustering method based on multi-scale semantic reconstruction and abnormal driving mechanism. After denoising of multi-source data, the target is to depict the implicit structure relationship between devices in functional cooperation and abnormal conduction, and to carry out dynamic clustering accordingly. Its specific implementation process is:

[0146] S31, map the multi-dimensional perception state of each power equipment in the time window to a semantic representation vector, including multi-scale behavior semantics, that is, for each device i Denoised data sequence Multi-scale semantic feature extraction e i : ;

[0147] Wherein, k represents F perception features at each time, that is, the sequence length of the denoised data sequence of each device i F represents the number of perception features at each time; Indicates the down-sampled sequence, which is used to capture long-period behavior patterns; And Gated Recurrent Unit, modeling fine-grained and long-term dependency features respectively; Convolution-Attention module, extracting local key timing change patterns; Concatenate operation; i Multi-scale semantic embedding vector of device i;

[0148] S32, combined with multi-scale semantic similarity and abnormal resonance degree, construct an elastic weighted multi-scale similarity graph between devices, dynamically integrate functional association and abnormal interaction, and have a "redundancy suppression mechanism":

[0149] ;

[0150] ; ;

[0151] ;

[0152] Among them, Edge weight between node i and j, i.e. elastic structure strength; Sigmoid function; , , Adjustable weight, used to balance the influence of similarity, resonance degree and redundancy; Multi-scale semantic embedding vector of device i e i Semantic similarity (cosine similarity) between multi-scale semantic embedding vector e j And multi-scale semantic embedding vector e Abnormal resonance degree, i.e. the proportion of two devices abnormal at the same time, statistics in the past k time common abnormal frequency; k represents the length of time window; t represents the current time; Time variable in sliding time window, used for summation iteration; And Indicate whether device i is abnormal at time τ and whether device j is abnormal at time τ respectively; Indicating function, return 1 means the condition is true; Redundancy, correlation of device perception value; And Denoised perception data sequence; Cosine similarity; Logical and operator;

[0153] Accordingly: construct edge weight adjacency matrix A={ }, output elastic association graph between devices ;

[0154] wherein V represents a set of device nodes, i.e. a set of all monitored power device nodes; represents a set of elastic edges, i.e. dynamic association relationships existing between power devices, which are constructed in the embodiment by measuring multi-scale time series embedding similarity (cosine similarity) and abnormal resonance factor;

[0155] S33, abnormal driving graph convolution feature reconstruction, i.e. the constructed device elastic association graph Above, the graph neural network (GNN) is used to propagate structure information and fuse abnormal score signals, realize semantic reconstruction and abnormal difference amplification of devices in embedding space, and specifically:

[0156] S3301, initial semantic feature input: the input feature of each device node i is the embedding representation obtained in the previous step : ;

[0157] wherein d represents the feature dimension; e i represents the multi-scale semantic embedding vector of device i;

[0158] S3302, introduce abnormal disturbance mechanism: for each node, according to its abnormal score score i ∈[0,1], introduce perturbation signal:

[0159] ;

[0160] ;

[0161] wherein, represents the denoising reconstruction error of the i-th device; represents the mean of all device errors; represents a smoothing factor, which controls the nonlinear enhancement amplitude; represents a Gaussian noise disturbance, which simulates structure sensitivity;

[0162] S3303, based on the constructed weighted adjacency matrix A, perform edge-weighted graph convolution propagation: After L layers are executed, the final node feature is output: ;

[0163] wherein, 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 the j-th column in the weighted adjacency matrix, which contains a self-loop (i.e. ), I is the unit matrix; represents a nonlinear activation function; This represents the weight matrix of the l-th layer; Represents the set of adjacent nodes of node i; z represents the set of adjacent nodes containing a self-loop; i The embedding representation of the final device node i is an exceptionally sensitive representation with dimension d; This represents the feature vector of node i at layer L;

[0164] S34. Structure-aware device functional clusters are formed in the device embedding space. The clustering results can proactively respond to anomalous conduction characteristics, adaptively pruning and optimizing the elasticity graph structure between devices to improve structural sparsity and expressive power. Specifically:

[0165] S3401. Initialize the structurally consistent clustering framework: The device obtains the final embedded representation after graph convolutional propagation. The K-Means method was initially used to divide the device into M initial clusters: {C1, C2, ..., C...} m ,…,C M Each cluster center is represented as: ;

[0166] in, This represents the m-th cluster, which contains a group of device nodes with similar structural / abnormal characteristics; This represents the center vector of the m-th cluster;

[0167] S3402. Define a multi-objective optimization function and construct an anomaly-aware clustering objective function: ;

[0168] in, Indicates an indicator function, if i and j belong to the same cluster C. m If the value is 1, then take 1; otherwise, take 0. This indicates the anomaly score for device i; This represents the variance of anomaly scores for all devices in the cluster containing device i. , represents the weights of the structural consistency loss term and the abnormal clustering loss term, respectively; L3 represents the objective function;

[0169] It should be noted that the objective function introduces a structural consistency mechanism (consistent edge connectivity) and anomaly clustering (consistent anomaly trends), actively favoring the grouping of devices that are "functionally similar + have similar anomaly patterns" into one category. This achieves semantic collaborative fusion and clustering of structural anomalies, possessing the ability to be driven by anomalies and interpretable structures. ;

[0170] S3403. To prevent graph redundancy from misleading clustering, define edge pruning constraints:

[0171] ;

[0172] ;

[0173] wherein, denotes the edge weight pruning threshold; denotes the average value of all edge weights in the current graph; denotes the standard deviation of all edge weights in the graph; denotes the edge weight threshold adjustment factor;

[0174] S3404, in each iteration, sequentially perform: updating the embedding z based on the current graph structure i ; performing structure-aware anomaly clustering based on the embedding; correcting the graph structure edge weight based on the clustering result; if the objective function converges or reaches the maximum number of rounds, terminate.

[0175] S4, the fault gene graph evolution reasoning module takes the output results of the inter-device resilient association graph and the anomaly-driven clustering as inputs, proposes an evolution path reasoning method based on the fault gene graph, mines device-level fault behavior sequence templates from historical behavior patterns, combines the current anomaly state, realizes the prediction and deduction of the future potential propagation path, and assists the scheduling control system in identifying the risk chain in advance, and the specific implementation process is:

[0176] S41, a graph-based memory unit is established based on historical fault evolution behaviors, that is, fault sequences are extracted from massive historical fault events, including the order and time interval information of each device anomaly, all historical sequences are abstracted as path information with time labels, and then a fault gene graph G is constructed fault ;

[0177] The edge weight (propagation weight) of the graph contains both frequency factors and time delay factors, ensuring that it reflects the combined influence of historical occurrence probability and propagation efficiency: ;

[0178] wherein, denotes the propagation edge weight from node i to j, which comprehensively reflects the frequency and efficiency of propagation from device i to device j in history; denotes the number of historical fault events from device i to device j; denotes the total number of fault events in which device i served as the starting point of propagation; denotes the average propagation time delay from i to j; denotes the time delay decay factor (constant), which controls the degree of attenuation of the propagation delay on the edge weight, controls the degree of attenuation of the propagation delay on the edge weight, and is empirically set to 10 minutes in this embodiment through experiments;

[0179] S42, project the current detected abnormal equipment state to the fault gene map, dynamically activate the potential propagation path by constructing a path activation mechanism jointly driven by abnormal intensity-structure intensity-propagation similarity, and thus realize the reconstruction and extension of the fault chain;

[0180] The matching score function is: ;

[0181] Wherein, P represents a candidate propagation path; represents the elastic dependence intensity from i to j in the current abnormal graph, which is obtained from the elastic association graph and derived from the cooperative behavior pattern; represents the abnormal state difference: the difference between the abnormal values of devices i and j;

[0182] S43, based on the most possible propagation path set activated in the last step, further estimate the fault propagation probability of other abnormal devices, and construct an evolution graph structure G predict :

[0183] The propagation probability estimation is: ;

[0184] Set the risk threshold θ p , and construct the predictive propagation path: ;

[0185] Generate the final abnormal propagation prediction graph G predict ;

[0186] Wherein, represents the prediction probability of the current abnormal device j in the future; represents the current abnormal device set; represents the time interval from the occurrence of abnormality of device i to the current time; represents the decay time factor, which controls the propagation influence intensity of the recent abnormality and prevents the misleading of the old abnormality; represents the set of future propagation edges obtained by speculation.

[0187] S5, expert rule and model collaborative judgment module integrates expert knowledge rule domain and model algorithm reasoning domain, constructs a "human-machine knowledge" double channel, multi-view joint discrimination mechanism, and through the dynamic game and correction mechanism of "structure constraint + behavior insight", significantly improves the explanation, accuracy and controllability of discrimination, and the specific implementation process is:

[0188] S51, based on the power industry expert experience library and accident diagnosis standard document, abstract the device abnormal mode characteristics and logical rule chain, 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: ;

[0189] where r k represents the kth expert experience rule; represents the abnormal feature associated with the kth rule (including but not limited to temperature too high, current fluctuation, frequency offset); represents the duration of the symptom; represents the risk level indicated by the rule;

[0190] Formalize the rule as a logical predicate function: ;

[0191] Construct a multi-rule driven discriminant vector: ;

[0192] where, represents whether the logic of the kth expert rule is satisfied after applying it to the input x (1 represents satisfaction, 0 represents dissatisfaction); represents the duration of abnormal behavior; represents a 0 / 1 vector composed of the discriminant results of the current device state by all expert rules, i.e., the matching of the expert rule set to the data;

[0193] S52, construct a light graph neural network structure (GAT, Graph Attention Networks), jointly encode the node features, graph structure and propagation risk in the abnormal propagation prediction graph G predict into low-dimensional semantic embedding h i , and perform risk classification based on the discriminant model :

[0194] ;

[0195] ;

[0196] 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 ith device node; represents the downstream classification model (MLP), with parameters θ, which outputs the abnormal risk level of the node; represents the abnormal classification result of the model for the ith device, the risk level predicted by the model;

[0197] S53, dual-domain collaborative game judgment mechanism: a heterogeneous-driven collaborative game model is designed to explain and correct the conflicts between expert rule judgment and model reasoning results, and a trust-weighted collaborative factor is introduced for unified decision-making judgment:

[0198] Collaborative judgment score: ;

[0199] ;

[0200] wherein, represents the sigmoid function normalization expert scoring; represents a dynamic weight factor determined by the degree of abnormality; represents the difference between the model and the expert output, measuring the divergence; represents a hyperparameter that controls the sensitivity of collaboration; represents the final comprehensive risk score of the ith device; represents the expert rule weight vector, i.e., the importance of each rule; represents the expert rule matching result vector of the ith device; represents the fusion judgment score vector of the ith device;

[0201] Output final decision level : ;

[0202] wherein, represents the comprehensive fusion score of device node i being judged as the cth risk level (e.g., normal, warning, severe), which is calculated by the expert rule domain and the model reasoning domain through a game weighting mechanism, reflecting the confidence degree or score level of device i belonging to the cth class in the collaborative judgment of expert rule judgment results and model prediction results.

[0203] S6, multi-level warning and knowledge backwriting module performs hierarchical response processing and knowledge layer adaptive evolution feedback mechanism of abnormal results, triggers precise multi-level warning measures according to the abnormal level, and extracts structured knowledge and automatically generates trusted rules based on evolution trajectory and symptom combination, forming a closed loop of continuous linkage and dynamic update between expert rule library and intelligent model, and the specific implementation process is as follows:

[0204] S61, based on the judgment results, combined with the fault gene map, the propagation influence of each abnormal node is evaluated, forming a multi-dimensional risk score index L i with propagation depth perception, and mapping it to multi-level warning levels: ;

[0205] wherein, L i represents the multi-dimensional propagation influence score of node i; di represents the propagation depth of node i in the fault evolution path graph; p i represents the participation frequency of the node in multiple propagation paths; w1, w2, w3 represent risk score weighting coefficients;

[0206] Warning level mapping: ;

[0207] wherein, Alert i represents the warning level; , , represents the multi-level warning trigger threshold;

[0208] S62, after triggering the warning, the abnormal characteristics of the warning node, the triggering path and the upstream and downstream associated equipment are structurally analyzed, and the standardized knowledge items are automatically generated:

[0209] ;

[0210] wherein, represents the newly generated rule item; Symptoms i represents the abnormal symptom combination (multi-dimensional characteristics); T i represents the abnormal duration; CauseChain i represents the causal path chain triggering the abnormality, which is obtained by fault gene graph evolution path reasoning, describing the whole process of the abnormality from a certain initial node (such as a certain component of a device) in the system to the current node (i.e. the location of the fault);

[0211] Based on the comprehensive evaluation of symptom frequency, path consistency and historical hit performance, a rule confidence index is introduced , and the following mechanism is used to judge whether to write back to the knowledge base: ;

[0212] wherein, represents the confidence score of the newly generated rule; represents the expert rule base before the tth update; represents the updated rule base;

[0213] S63, in order to prevent the expansion and accumulation of redundant rules, a rule confidence dynamic adjustment mechanism based on usage frequency and prediction consistency is designed, and it is used to guide the model parameter collaborative optimization:

[0214] If the newly added rule hits frequently and consistently with the model prediction for a long time, its confidence weight w r is increased;

[0215] If the rule performance is inconsistent or frequently misjudged, the weight reduction process is automatically triggered;

[0216] Accordingly, under this mechanism, a double-domain self-evolution closed-loop structure of model-aided rule generation → rule-driven model constraint optimization → abnormality re-discrimination feedback is formed.

[0217] The embodiments of the present application are described in detail above with reference to the drawings, but the present application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A power equipment anomaly detection and early warning method, characterized in that, The embodiment comprises the following specific implementation steps: S1, collecting the running state, environmental parameters and historical maintenance record data of the target power equipment; S2, through a three-domain fusion denoising framework of data source correlation dynamic measurement, time sequence semantic stability discrimination and graph structure context redundancy suppression, dynamically adjusting the noise suppression weight under different data modalities, performing cross-modal collaborative denoising, constructing a semantic residual loss function based on a semantic neighborhood weighted reference value, and introducing a structure perception multi-scale consistency function to feedback and correct the denoising result; The semantic residual loss function L1 is: ； ； ； wherein, denotes the semantic neighborhood reference value of the i-th perception source; denotes the set of neighbor nodes connected to node i in the perception graph; denotes the edge weight in the graph structure, i.e., the semantic similarity weight of the i-th and j-th perception source at time t; denotes the denoised data, the j-th dimension denoised value of each perception source; denotes the hyperparameters; denotes the mapping function for extracting semantic embeddings, i.e., the attention perception pooling operation ; is a denoising function; denotes the denoised data; D denotes a three-dimensional tensor, which is used to uniformly express the multi-dimensional index data collected by multiple perception nodes within a period of time; S denotes the total number of perception sources, ; F denotes the number of indexes collected by any perception source, ; T denotes the total time series, ; denotes the dynamic enhancement coefficient; denotes the j-th perception index value of the i-th perception source at time t; S3, extracting equipment features through multi-scale semantic embedding, constructing an elastic association graph combined with semantic similarity, abnormal resonance degree and redundancy suppression, reconstructing features by using a graph neural network to fuse abnormal disturbance signals, and realizing dynamic equipment grouping through multi-objective optimization clustering and edge pruning optimization; S4, constructing a fault gene map through historical fault data, calculating edge weights by fusing frequency and time delay factors, activating potential paths combined with the current abnormal state, estimating the future propagation probability of non-abnormal equipment, and generating a final abnormal propagation prediction graph; S5, constructing a structured knowledge graph and a graph attention network to encode node features and propagate risks, designing a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model prediction, and outputting a comprehensive risk level by fusing logical predicate discrimination and classification results; S6, generating a multi-dimensional risk score and mapping an early warning level through propagation influence evaluation, analyzing abnormal features and propagation paths to generate knowledge items after triggering early warning, dynamically updating a rule library based on credibility evaluation, designing an adjustment mechanism driven by rule usage frequency and prediction consistency to optimize model parameters, and forming a dual-domain self-evolution closed loop.

2. The method of claim 1, wherein the method further comprises: The denoising process of the cross-modal collaborative denoising is as follows: S21, the semantic vector of each perception source at each time is extracted by using an attention perception pooling method, and the semantic similarity of each two perception sources i, j at the current time t is calculated: ； ； wherein, denotes the data sequence of the i-th perception source within the time window [t-k, t]; denotes the semantic feature representation of the sampled data of the perception source i at time t; denotes the semantic feature vector of the i-th perception source at time t; denotes the attention perception pooling operation; k denotes the length of the history window; denotes the semantic similarity weight of the i-th and j-th perception source at time t; here, denotes the transpose operation; denotes the Euclidean norm; denotes the indicator function; S22, the average volatility rate in the sliding window of each perception dimension is calculated, and a dynamic enhancement coefficient is set: ； ； wherein, represents the short-term fluctuation intensity of the current perception dimension; K represents the size of the sliding window; represents the jth perception indicator value of the ith perception source at time t; represents the dynamic enhancement coefficient; for regulating the compression rate of fluctuation; for preventing excessive suppression; represents the natural exponential function; S23, denoising data is generated by a denoising function, a semantic residual loss function is constructed based on a semantic neighborhood weighted reference value, and the balance between residual and semantic consistency loss is optimized by combining graph structure edge weight and attention pooling mapping function; S24, multi-layer features of original and denoised data are extracted by a multi-scale aggregation function, a structure consistency loss is defined and integrated into a comprehensive optimization objective, and a regulation factor is used to balance the multi-scale structure preservation strength and suppress local overfitting; S25, the denoised data sequence of any power equipment i is outputted.

3. The method of claim 1, wherein the method further comprises: The grouping process of the dynamic equipment grouping is as follows: S31, for each device i, a denoised data sequence performing multi-scale semantic feature extraction e i : ； wherein, denotes the denoised data sequence of device i, with a time window length of k and F perception features at each time instant; denotes the down-sampled sequence; and denotes the gating recurrent unit; denotes the convolution-attention module; denotes the concatenation operation;e i denotes the multi-scale semantic embedding vector of device i; S32, combine multi-scale semantic similarity and abnormal resonance degree, construct an elastic weighted multi-scale similarity graph between devices, construct an edge weight adjacency matrix A={ },output the elastic correlation graph between devices ; ； ； ； ； wherein, denotes the edge weight between nodes i and j; denotes the Sigmoid function; , , denotes the adjustable weight; denotes the semantic similarity between the multi-scale semantic embedding vector e i of device i and the multi-scale semantic embedding vector e j of device j, where T denotes the vector transpose operation; denotes the abnormal resonance degree; t denotes the current time; denotes the time variable in the sliding time window; and respectively denote whether device i is abnormal at time τ and whether device j is abnormal at time τ; denotes the indicator function; denotes the redundancy; and denotes the denoised perception data sequence; denotes the cosine similarity; denotes the logical AND operator; S33, the multi-scale semantic embedding vector is used as the initial feature, a perturbation signal is introduced based on the abnormal score, the adjacent node information is fused by weighted graph convolution propagation, and the final node feature is outputted after multi-layer iteration; The output process of the final node feature is as follows:

4. The method of claim 3, wherein the step of detecting the abnormality of the power device comprises the steps of: detecting the abnormality of the power device based on the comparison result of the first and second comparison results. The construction process of the abnormal perception clustering objective function is as follows: A1, initial semantic feature input, the input feature of each device node i is the embedding representation obtained in the previous step : ; where e i denotes the multi-scale semantic embedding vector of device i; A2. For each node, according to its anomaly score score i ∈ [0,1], introducing a perturbation signal: ； ； wherein, denotes the denoising reconstruction error of the i-th device; denotes the mean of all device errors; denotes a smoothing factor; denotes a Gaussian noise disturbance; A3, based on the constructed weighted adjacency matrix A, perform band edge weight graph convolution propagation: After L layers are executed, output the final node feature: ; wherein, represents the feature vector of node j at the lth layer; represents the feature vector of node i at the l+1th layer; represents the value in the weighted adjacency matrix of the ith row and jth column, including self-loop, i.e. , 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 including self-loop; z i represents the anomaly-sensitive embedding representation of the final device node i; represents the feature vector of node i at the Lth layer.

5. The method of claim 4, wherein the step of detecting the abnormality of the power device comprises the steps of: detecting the abnormality of the power device based on the comparison result of the comparison between the first and second values of the current. ​ B1. Based on the final embedding representation of the device after graph convolution propagation , the K-Means method is used to divide the devices into M initial clusters: {C1, C2, …, CM}: m ,…,C M , and each cluster center is represented as: ; wherein, denotes the m-th cluster; denotes the center vector of the m-th cluster; B2, build an anomaly perception clustering objective function: ； wherein, represents an indicator function that takes 1 if i and j belong to the same cluster C m and 0 otherwise; 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 weight of the structural consistency loss term and the weight of the anomaly concentration loss term; L3 represents the objective function.​ 6. The method of claim 1, wherein the method further comprises: The final anomaly propagation prediction graph generation process is as follows: S41, based on the historical failure evolution behavior to establish a graph memory unit, that is, to extract the failure sequence from the historical failure events, including the order of abnormal occurrence of each device and the time interval information, all historical sequences will be abstracted as path information with time annotation, and then the failure gene graph G is constructed fault ; determining the edge weights, i.e. the weights, of the graph: ; where, 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 failure events that device i has ever been the origin of; represents the average propagation time delay from i to j; represents the time delay decay factor; S42, project the current detected abnormal equipment state to the fault gene map, dynamically activate the potential propagation path by constructing an anomaly intensity-structure intensity-propagation similarity joint driven path activation mechanism, and realize the reconstruction and extension of the fault chain; Matching score function: ; wherein P represents a candidate propagation path; represents the elasticity dependency strength from i to j in the current anomaly graph; represents the anomaly state difference, i.e. the difference degree between the anomaly values of devices i, j. S43, based on the activated propagation path set {P*}, estimate the failure propagation probability of other abnormal devices that have not yet appeared, and construct the future-oriented evolution graph structure G predict : The propagation probability is estimated as: ; Setting a risk threshold θ p , constructing a predictive propagation path: ; generating a final anomaly propagation prediction graph G predict ; wherein, represents the predicted probability that the current non-faulty device j will have a fault in the future; represents the set of devices for which a fault has been detected so far; represents the time interval from the occurrence of the fault to the current time instant for device i; represents the decay time factor; represents the set of future propagation edges conjectured.

7. The method of claim 6, wherein the step of detecting the abnormality of the power device comprises the steps of: detecting the abnormality of the power device based on the comparison result of the comparison between the first and second values of the current. The output process of the comprehensive risk level is as follows: S51, based on the power industry expert experience library and accident diagnosis standard documents, abstract device abnormal mode characteristics and logical rule chain, and construct structured knowledge graph K exp Each expert rule is represented as a triple: ; wherein r k represents the kth expert experience rule; represents an abnormal feature associated with the kth rule; represents a duration of the abnormal feature; represents a risk level indicated by the rule; Formalize the rules as logical predicate functions: ; Constructing the multi-rule driven discriminant vector: ; wherein, represents whether the logic of the kth expert rule is satisfied after applying it to the input x; represents the duration of the abnormal behavior; represents a 0 / 1 vector composed of the decision results of the current device state by all expert rules, i.e. the matching of the expert rule set to this data; S52, construct a lightweight graph neural network structure, and predict the graph G of abnormal propagation predict Jointly encode the node features, graph structure and propagation risk in the graph G into low-dimensional semantic embeddings h i , and perform risk classification based on a discriminative model : ； ； wherein X represents the original input features of the node; denotes the graph attention network; denotes the embedding representation of the i-th device node; denotes a downstream classification model with parameters θ, which outputs the abnormal risk level of the node; denotes the abnormal classification result of the i-th device by the model, i.e., the risk level predicted by the model; S53, build a dissent-driven collaborative game model to explain and correct the conflicts between expert rule discrimination and model reasoning results, and introduce a trust-weighted collaborative factor for unified decision-making discrimination: Collaborative discrimination score: ； ； wherein, represents a sigmoid function normalization expert score; represents a dynamic weight factor; represents a difference degree of the model and the expert output; represents a hyperparameter for controlling the sensitivity of the synergy; represents a final comprehensive risk score of the ith device; represents an expert rule weight vector; represents an expert rule matching result vector of the ith device; represents a fusion discriminant score vector of the ith device; S54, outputting the final decision level : ; wherein, represents the integrated fusion score that device node i is determined to be of risk level c.

8. The method of claim 2, wherein the method further comprises: The comprehensive optimization objective is: C1. Using a multi-scale aggregation function Extracting multi-layer features: ； ； C2, definition structure consistency loss: ; C3, define the comprehensive optimization goal: ; wherein, and respectively represent the original and denoised data structure features extracted from the lth layer; represents a multi-scale structure feature extractor; L represents the number of multi-scale feature layers; represents a multi-scale structure preservation loss; represents the overall loss; represents a multi-scale consistency regulation factor; represents denoised data; D represents a three-dimensional tensor, that is, a unified expression of multi-dimensional index data collected by multiple perception nodes over a period of time.

9. A power equipment abnormality detection and early warning system for performing the power equipment abnormality detection and early warning method of any one of claims 1-8, characterized in that, Including: A multi-source data acquisition module for acquiring multi-source perception data of the target power equipment; An edge computing and preprocessing module for realizing data preliminary screening and denoising on the field terminal; An anomaly elasticity clustering construction module for dynamically constructing a weighted device relationship graph based on the spatiotemporal similarity between device operating states and the diffusion characteristics of abnormal signals, and performing anomaly-driven clustering analysis to identify potential device abnormal diffusion groups; A fault gene map evolution reasoning module for abstracting device historical abnormal data into fault gene vectors, constructing a map, and matching the similar evolution path corresponding to the current potential anomaly based on a graph attention mechanism to realize intelligent analogy and trend prediction of the anomaly; An expert rule and model collaborative discrimination module for constructing a structured knowledge graph and graph attention network to encode node features and propagation risks, designing a dual-domain collaborative game mechanism to dynamically adjust the weight difference between expert rules and model predictions, and outputting a comprehensive risk level by fusing logical predicate discrimination and classification results; A multi-level early warning and knowledge backwriting module for outputting multi-level early warning signals according to the fusion judgment results, and writing the features, evolution trajectory and response strategy in the current detection into the system knowledge base in a structured way to support subsequent model iteration and self-optimization, and realize the rule backwriting mechanism of the new anomaly detection and early warning process.

Images