Power transformation equipment fault diagnosis method, equipment and medium

Abstract

The invention relates to a power transformation equipment fault diagnosis method, equipment and a medium, and relates to the technical field of power equipment, and the method comprises the following steps: analyzing propagation paths and propagation influences of different features in multi-modal data of power transformation equipment based on a fault feature propagation model so as to construct a knowledge graph; the knowledge graph is optimized by analyzing the propagation dynamic characteristics and the propagation links, and the optimized knowledge graph is used for outputting a power transformation equipment fault diagnosis result. Compared with the prior art, the method can achieve the precise diagnosis of the faults of the power transformation equipment, improves the evaluation accuracy, and reduces the operation and maintenance cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment, in particular to a power transformation equipment fault diagnosis method, device and medium. BACKGROUND

[0002] Power transformation equipment is a key infrastructure in the power system, and its safe and stable operation is crucial for ensuring the reliability of the power grid. With the continuous expansion of the power system and the deepening of the construction of smart grids, the number of power transformation equipment continues to increase, the types of equipment are increasingly diversified, and the operating environment is more complex. Traditional power transformation equipment state evaluation mainly relies on manual inspection, periodic detection and experience judgment, which is difficult to meet the needs of modern power grids for real-time monitoring, accurate evaluation and fault warning of equipment state. In recent years, with the development of artificial intelligence technology, intelligent evaluation methods based on data driving have been gradually applied to the field of power transformation equipment state evaluation, which realizes intelligent evaluation and fault diagnosis of equipment state by collecting equipment operation data and combining machine learning algorithms.

[0003] However, the existing power transformation equipment state evaluation method still has some obvious deficiencies. First, the existing method usually only focuses on single modal data such as temperature, vibration or electrical parameters, lacks comprehensive analysis ability of multi-modal data, and cannot fully reflect the actual operating state of the equipment, resulting in one-sided evaluation results and insufficient accuracy. In view of this, Chinese patent CN120217264A discloses a complex equipment fault diagnosis method and system based on multi-modal knowledge graph, which includes: performing multi-modal feature extraction on the obtained equipment operation data respectively; constructing a knowledge graph based on the extracted multi-modal features; performing multi-modal feature enhancement on the knowledge graph through cross-modal comparison learning; defining a small sample task, and initializing the enhanced knowledge graph; performing subgraph extraction and Mamba enhancement on the feature initialized knowledge graph; and performing fault classification based on the knowledge graph using a graph neural network. Although this method considers multi-modal data, it does not analyze the correlation between fault features in depth, making it difficult to accurately capture the fault propagation path and evolution law, and lacking effective support for equipment state evaluation in complex fault scenarios, especially in the case of multi-device interaction. In addition, it has limitations in knowledge expression and reasoning, making it difficult to effectively combine the experience of field experts and data-driven analysis results, lacking the ability to deeply mine the dynamic characteristics of fault feature propagation, and unable to build a complete fault diagnosis knowledge system, thereby affecting the accuracy and explainability of fault diagnosis. SUMMARY

[0004] The purpose of the present application is to overcome the deficiencies of the prior art and provide a power transformation equipment fault diagnosis method, device and medium.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] According to a first aspect of the present invention, a method for diagnosing faults in substation equipment is provided. This method constructs a knowledge graph by analyzing the propagation paths and influence of different features in the multimodal data of substation equipment based on a fault feature propagation model, and optimizes the knowledge graph by analyzing the propagation dynamic characteristics and propagation links, and outputs the fault diagnosis results of substation equipment using the optimized knowledge graph.

[0007] As a preferred technical solution, the method includes the following steps:

[0008] Acquire multimodal data of substation equipment, analyze the propagation path of different features in the multimodal data based on the fault feature propagation model, calculate the correlation degree between features according to the propagation path, extract feature combinations of the multimodal data based on the correlation degree, and generate an initial feature fusion vector;

[0009] The fault feature propagation model is used to construct a feature importance evaluation system for the initial feature fusion vector. The propagation influence of features is analyzed based on the feature importance evaluation system. Based on the propagation influence, the main influential features are selected to generate an optimized feature fusion vector.

[0010] A fault feature knowledge graph is constructed based on the optimized feature fusion vector. The propagation dynamic characteristics of nodes in the fault feature knowledge graph are analyzed based on the fault feature propagation model. The evolution law of fault features is identified according to the propagation dynamic characteristics. The association strength of the fault feature knowledge graph is updated based on the evolution law to obtain the optimized knowledge graph.

[0011] The fault feature propagation model is used to perform propagation link analysis on the optimized knowledge graph. Based on the results of the propagation link analysis, the main paths of fault propagation are mined. The topology of the fault feature knowledge graph is optimized according to the main paths to obtain the final knowledge graph. The fault diagnosis results of the substation are output based on the final knowledge graph.

[0012] According to a second aspect of the present invention, an electronic device is provided, comprising:

[0013] processor;

[0014] Memory used to store processor-executable instructions;

[0015] The processor is configured to invoke instructions stored in the memory to execute the method.

[0016] According to a third aspect of the present invention, a computer-readable storage medium is provided having computer program instructions stored thereon, which, when executed by a processor, implement the method described thereon.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) This invention realizes the effective fusion and optimization of multimodal data features, which significantly improves the accuracy and reliability of fault diagnosis.

[0019] (2) The present invention constructs a fault feature propagation model and a feature importance assessment system, which can automatically identify and screen the main influential features, reduce redundant information interference, simplify the fault analysis process, and improve diagnostic efficiency.

[0020] (3) By constructing and optimizing the fault feature knowledge graph, this invention achieves accurate identification of the evolution law of fault features and precise mining of the main path of fault propagation, which enhances the interpretability and foresight of the substation condition evaluation and provides a reliable basis for power system operation and maintenance decision-making. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention;

[0022] Figure 2 This is a bar chart comparing the performance of the feature selection method based on propagation influence in embodiments of the present invention.

[0023] Figure 3 This is a logic block diagram for identifying the evolution law based on the fault feature propagation model in an embodiment of the present invention;

[0024] Figure 4 This is a bar chart comparing the performance of fault propagation link analysis based on multi-head attention mechanism in embodiments of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0027] Example 1

[0028] This embodiment provides a method for diagnosing faults in power equipment. The method is based on a fault feature propagation model to analyze the propagation paths and influence of different features in the multimodal data of power equipment to construct a knowledge graph. The knowledge graph is then optimized by analyzing the dynamic characteristics and propagation links of the propagation. Finally, the optimized knowledge graph is used to output the fault diagnosis results of the power equipment.

[0029] like Figure 1 As shown, the method includes the following steps:

[0030] S1, Multimodal data processing and initial feature fusion: Acquire multimodal data of substation equipment, analyze the propagation path of different features in the multimodal data based on the fault feature propagation model, calculate the correlation degree between features according to the propagation path, extract feature combinations of the multimodal data based on the correlation degree, and generate an initial feature fusion vector;

[0031] In one optional implementation, the propagation paths of different features in the multimodal data are analyzed based on a fault feature propagation model. The correlation between features is calculated according to the propagation paths. Based on the correlation, feature combinations of the multimodal data are extracted to generate an initial feature fusion vector, including:

[0032] Construct a fault feature propagation model, which includes a set of feature nodes and a set of propagation edges;

[0033] Calculate the propagation strength between nodes in the feature node set, where the propagation strength characterizes the association relationship between nodes, and determine the propagation path between the feature node sets based on the propagation strength;

[0034] Based on the propagation path, feature sample pairs are constructed. The propagation dynamic characteristics of the nodes in the feature sample pairs are analyzed using the fault feature propagation model. The propagation dynamic characteristics include lag time, response speed, attenuation coefficient, and fluctuation correlation. The evolution law between nodes is identified according to the propagation dynamic characteristics. The evolution law includes linear following type, threshold triggering type, lag accumulation type, and periodic coupling type. The feature correlation degree corresponding to the propagation path is calculated based on the evolution law and propagation intensity.

[0035] Based on the degree of feature association, a hierarchical clustering method is used to cluster the nodes in the feature node set, extract the feature combinations of the multimodal data, calculate the importance and complementarity of features within the feature combinations, obtain the combination score, and generate an initial feature fusion vector based on the feature combinations and their corresponding combination scores.

[0036] S2, Feature Importance Assessment and Optimization: Construct a feature importance assessment system for the initial feature fusion vector using the fault feature propagation model, analyze the propagation influence of features based on the feature importance assessment system, select the main influential features based on the propagation influence, and generate an optimized feature fusion vector;

[0037] In one optional implementation, the feature importance evaluation system of the initial feature fusion vector is constructed using the fault feature propagation model. The propagation influence of features is analyzed based on the feature importance evaluation system. Based on the propagation influence, key influential features are selected to generate an optimized feature fusion vector, including:

[0038] A Bayesian network model is constructed based on the initial feature fusion vector. The Bayesian network model includes a set of conditional nodes and a set of causal edges. The conditional probability and mutual information between adjacent nodes in the set of conditional nodes are calculated using the Bayesian network model. The causal strength between nodes is determined based on the conditional probability and the mutual information.

[0039] A graph neural network is constructed based on the Bayesian network model and the causal strength between nodes. The conditional node set and the causal edge set are mapped to graph structure data. The node features are the corresponding feature values ​​in the initial feature fusion vector, and the edge weights are the calculated causal strength between nodes. The graph neural network is used to extract the topological structure information and the causal strength information between nodes of the conditional node set. The propagation influence of the feature nodes is calculated based on the topological structure information and the causal strength information between nodes.

[0040] The importance score of the feature node is calculated based on the dissemination influence. The mean and standard deviation of the feature distribution are calculated based on the importance score. The feature screening threshold is determined based on the mean and standard deviation. Feature nodes with an importance score greater than the feature screening threshold are selected as the main influence features. Feature fusion is performed on the main influence features to generate an optimized feature fusion vector.

[0041] S3, Fault Feature Knowledge Graph Construction and Optimization: Construct a fault feature knowledge graph based on the optimized feature fusion vector, analyze the propagation dynamic characteristics of nodes in the fault feature knowledge graph based on the fault feature propagation model, identify the evolution law of fault features according to the propagation dynamic characteristics, update the association strength of the fault feature knowledge graph based on the evolution law, and obtain the optimized knowledge graph;

[0042] In one optional implementation, a fault feature knowledge graph is constructed based on the optimized feature fusion vector; the propagation dynamics of nodes in the fault feature knowledge graph are analyzed based on the fault feature propagation model; and the evolution law of fault features is identified according to the propagation dynamics, including:

[0043] A fault feature knowledge graph is constructed based on the optimized feature fusion vector, and a causal discovery algorithm based on the invariance principle is introduced into the fault feature knowledge graph to calculate the causal strength between adjacent nodes in the fault feature knowledge graph. A multi-level node representation of the fault feature knowledge graph is constructed based on the causal strength.

[0044] The multi-level node representation and the causal strength are input into the fault feature propagation model to construct the propagation state vector of the feature node. The cumulative causal influence of the feature node is calculated based on the propagation state vector. The propagation dynamic characteristics of the feature node are analyzed according to the cumulative causal influence and the propagation state vector, and three key indicators, node propagation rate, influence range and stabilization time, are calculated.

[0045] Based on the propagation dynamic characteristics, the temporal evolution characteristics and causal evolution characteristics of the feature nodes are constructed. The temporal evolution characteristics and the causal evolution characteristics are combined to calculate the evolution law of the fault characteristics.

[0046] S4, Propagation Link Analysis and Fault Diagnosis: The fault feature propagation model is used to perform propagation link analysis on the optimized knowledge graph. Based on the results of the propagation link analysis, the main paths of fault propagation are mined. The topology of the fault feature knowledge graph is optimized according to the main paths to obtain the final knowledge graph. The fault diagnosis results of the substation are output according to the final knowledge graph.

[0047] In one optional implementation, the fault feature propagation model is used to perform propagation link analysis on the optimized knowledge graph. Based on the results of the propagation link analysis, the main paths of fault propagation are mined, and the topology of the fault feature knowledge graph is optimized according to the main paths to obtain the final knowledge graph, including:

[0048] The fault feature propagation model is used to analyze the propagation link of the optimized knowledge graph. A multi-head attention mechanism is constructed to calculate the propagation probability between nodes. Based on different temporal features, the dynamic correlation between nodes is extracted according to each attention head of the multi-head attention mechanism. The link importance evaluation function is constructed according to the dynamic correlation and the propagation probability between nodes. The propagation link is preliminarily analyzed based on the output of the link importance evaluation function to obtain the initial propagation link.

[0049] The dynamic correlation degree extracted by the multi-head attention mechanism is used as the weight of the node sequence. The conditional causal entropy of the initial propagation link is calculated based on the weight. The conditional causal entropy is evaluated a second time according to the link importance evaluation function. The output of the link importance evaluation function is compared with an adaptive threshold to select the main path of fault propagation.

[0050] The node topology of the main path and the dynamic correlation are input into the link importance evaluation function. The multi-head attention mechanism is used to dynamically optimize the edge connection relationship of the fault feature knowledge graph. The structure of the fault feature knowledge graph is iteratively optimized to obtain the final knowledge graph.

[0051] Example 2

[0052] This embodiment, based on embodiment 1, provides a further detailed description of step S1: multimodal data processing and initial feature fusion.

[0053] The fault feature propagation model construction begins with multimodal data preprocessing. Taking a 110kV substation GIS equipment as an example, multimodal data including partial discharge signals, SF6 gas density, ambient temperature, vibration signals, and ultrasonic signals are collected. All data types are standardized to ensure their values ​​are uniformly within the range of [0,1]. A sliding window is applied to segment the time-series data, with a window width of 30 minutes and a step size of 5 minutes. Statistical features, such as mean, standard deviation, peak value, and peak-to-average ratio, are extracted from each window, resulting in 83 feature nodes. The feature node set includes 43 time-domain features, 28 frequency-domain features, and 12 time-frequency-domain features, covering the operating status indicators of various components of the equipment.

[0054] The propagation edge set is constructed through feature correlation analysis. The Pearson correlation coefficient between each feature node is calculated, and an initial connection is established when the absolute value of the correlation coefficient exceeds a preset threshold of 0.35. In the GIS equipment case, the correlation coefficient between SF6 gas density and ambient temperature is -0.72, so a connection is established; the correlation coefficient between partial discharge and vibration signal peak value is 0.68, so a connection is established. The initial connections form an undirected graph structure, and the direction of the edges is determined using the PC algorithm, forming a directed propagation network. The improvement lies in introducing temporal sequence constraints to ensure that the direction of the edges is consistent with the temporal sequence of feature changes. For example, if the change in ambient temperature precedes the change in SF6 gas density, the edge direction points from ambient temperature to SF6 gas density.

[0055] The propagation intensity calculation is implemented using a graph neural network framework, constructing a three-layer graph convolutional network. The inputs are the original feature vectors of the feature nodes and the initial adjacency matrix. The first layer performs graph convolution operations, aggregating the feature information of each node and its neighbors, with an output dimension of 64. The second layer introduces an attention mechanism, calculating an attention score based on the node feature similarity, with an output dimension of 32. The third layer generates the latent representation of the nodes, with a dimension of 16. For each pair of connected nodes i and j, the propagation intensity is defined as the degree of influence of node j on node i, calculated as a weighted sum of the attention score and the node representation similarity, with weights of 0.7 and 0.3, respectively. For example, the attention score of the ambient temperature node on the SF6 gas density node is 0.81, and the node representation similarity is 0.75, resulting in a calculated propagation intensity of 0.7 × 0.81 + 0.3 × 0.75 = 0.79, indicating that ambient temperature has a strong influence on SF6 gas density.

[0056] Propagation path determination is based on a propagation intensity threshold screening. A threshold of 0.5 is set, retaining edges with propagation intensities greater than the threshold to form a simplified propagation network. Within this network, all paths from the source node to the target node are searched, with path lengths limited to three hops or less. The total propagation intensity of a path is defined as the product of the propagation intensities of all edges along that path. For example, in the path from ambient temperature to SF6 gas density to partial discharge signal, the propagation intensities of each edge are 0.79 and 0.68, respectively, resulting in a total propagation intensity of 0.79 × 0.68 = 0.54. Paths with a total propagation intensity greater than 0.4 are retained as valid propagation paths. In the GIS equipment case study, 38 main propagation paths involving 45 key feature nodes were selected from 83 nodes and 251 initial edges.

[0057] Feature sample pairs are constructed based on the filtered propagation paths. For each propagation path, the nodes at both ends of the path are extracted as feature sample pairs. Simultaneously, to enhance the correlation analysis between features, intermediate nodes are introduced as auxiliary features. Taking the path of ambient temperature-SF6 gas density-partial discharge signal as an example, two feature sample pairs are constructed: (ambient temperature, SF6 gas density) and (SF6 gas density, partial discharge signal), with SF6 gas density used as an auxiliary feature for both sample pairs. For each feature sample pair, its feature value sequence at different time points is collected, forming a time-series dataset of the feature sample pair. In the GIS equipment case, each sample pair contains data from 720 time points, covering a 15-day monitoring period.

[0058] The propagation dynamics analysis is achieved through time series correlation mining. For each feature sample pair (A, B), it analyzes how changes in A affect changes in B. Specifically, the dynamic time warping algorithm is used to calculate the alignment relationship between two time series and extract the following dynamic characteristics: lag time, response speed, decay coefficient, and fluctuation correlation. Lag time represents the response delay of feature B to changes in feature A, determined through cross-correlation analysis, and is measured in minutes. Response speed represents the rate at which feature B follows changes in feature A, calculated as the ratio of the rate of change of B to the rate of change of A. The decay coefficient represents the degree of weakening of the influence, obtained by fitting an exponential decay curve. Fluctuation correlation represents the consistency of periodic changes, calculated through spectral analysis. Taking (ambient temperature, SF6 gas density) as an example, the analysis yields a lag time of 25 minutes, a response speed of 0.72, a decay coefficient of 0.15, and a fluctuation correlation of 0.83.

[0059] The identification of node evolution patterns is based on the dynamic characteristics of propagation. Pattern mining algorithms are applied to the time-series data of the dynamic characteristics of each feature sample pair to identify typical evolution patterns. The main evolution patterns found in GIS equipment include: linear following (e.g., SF6 gas density follows changes in ambient temperature), threshold-triggered (e.g., partial discharge surges when SF6 gas density falls below a threshold), hysteresis-cumulative (e.g., insulation degradation increases with the cumulative effect of temperature fluctuations), and periodic coupling (e.g., vibration frequency is coupled with periodic temperature changes). For the (ambient temperature, SF6 gas density) sample pair, the identified evolution pattern is linear following, with the following characteristic parameters: following coefficient 0.72, linearity 0.89, and stability 0.94. The linear following pattern indicates that SF6 gas density exhibits a stable linear following relationship with changes in ambient temperature, which is consistent with physical principles.

[0060] The calculation of feature correlation degree comprehensively considers both propagation intensity and evolution law. The correlation degree is defined as the weighted product of propagation intensity and evolution law stability, with the weights dynamically adjusted according to the evolution law type. For linear following type, the propagation intensity weight is 0.6 and the stability weight is 0.4; for threshold triggered type, the weights are 0.5 and 0.5 respectively; for lag accumulation type, the weights are 0.4 and 0.6 respectively; and for periodic coupling type, the weights are 0.55 and 0.45 respectively. Taking (ambient temperature, SF6 gas density) as an example, the propagation intensity is 0.79, the evolution law stability is 0.94, and the correlation degree is calculated as 0.79^0.6 × 0.94^0.4 = 0.85. Similarly, the correlation degree of all feature sample pairs is calculated to construct a feature correlation degree matrix with a dimension of 45×45, corresponding to the correlation degree between 45 key feature nodes.

[0061] Hierarchical clustering is based on the feature correlation matrix, treating correlation as the similarity between features, and employing an agglomerative hierarchical clustering algorithm to group features. The distance metric is the reciprocal of the correlation, and the connectivity criterion is the average connectivity method. The clustering process proceeds bottom-up, initially forming an independent cluster for each feature, then gradually merging the most similar clusters until a preset number of clusters is reached. The clustering results are visualized as a dendrogram, and the optimal number of clusters is determined by cutting the dendrogram. In the GIS equipment case, the optimal number of clusters was determined to be 7 based on the silhouette coefficient and the Davies-Bouldin index. Each cluster contains 6, 8, 7, 5, 9, 4, and 6 features, corresponding to different functional modules or fault types of the equipment.

[0062] Feature combination extraction is based on clustering results. For each feature within a cluster, its similarity to the cluster center is calculated, and the feature with the highest similarity is selected as the representative feature of that cluster. Simultaneously, feature complementarity is considered, and the subset of features with the strongest complementarity is selected from each cluster. Complementarity is evaluated using mutual information between features; lower mutual information indicates stronger complementarity. For example, the first cluster contains six features, including ambient temperature, humidity, and air pressure. Ambient temperature is selected as the representative feature, and ambient temperature, humidity, and air pressure are selected as the complementary feature set. Representative features and complementary feature sets are extracted from each of the seven clusters to form feature combinations. Each feature combination contains one representative feature and two to three complementary features, resulting in seven feature combinations covering 24 features.

[0063] The combination score calculation considers the importance and complementarity of features within the combination. Importance is assessed by the centrality of the feature in the propagation network; higher centrality indicates greater importance. Complementarity is assessed by the average mutual information among features within the combination; lower mutual information indicates stronger complementarity. The combination score is defined as a weighted sum of importance and complementarity, with weights of 0.65 and 0.35, respectively. For example, the importance score of an environment-related feature combination is 0.78, and the complementarity score is 0.83; the combination score is calculated as 0.65 × 0.78 + 0.35 × 0.83 = 0.80. Combination scores are calculated for each of the seven feature combinations, ranked from highest to lowest as follows: environment-related (0.80), SF6 gas-related (0.76), partial discharge-related (0.73), vibration-related (0.69), temperature-related (0.65), pressure-related (0.62), and electrical parameter-related (0.59).

[0064] The initial feature fusion vector is generated based on feature combinations and their scores. Feature engineering is performed on the features within each feature combination, including normalization, dimensionality reduction, and feature transformation. Principal component analysis is used to reduce the dimensionality of the features within the combination, retaining the principal components needed to explain 95% of the variance. For environment-related combinations, the original four features are reduced to two principal components; for SF6 gas-related combinations, the five features are reduced to three principal components. The sum of the feature dimensions after dimensionality reduction for each combination is 16. The features of each combination are weighted and fused according to their combination scores to form the initial feature fusion vector. The fusion process uses nonlinear transformation to enhance the feature expressiveness. The final generated initial feature fusion vector has a dimension of 16 and contains the most representative and complementary feature information from the multimodal data.

[0065] The effectiveness of the initial feature fusion vector was validated through a fault diagnosis task. A support vector machine model was trained using the fusion vector to classify four common faults in GIS equipment: SF6 leakage, insulation degradation, mechanical failure, and poor contact. The model performance was evaluated using 10-fold cross-validation, achieving an accuracy of 92.7% and an F1 score of 0.91, significantly outperforming baseline methods using single-modal features (with a maximum accuracy of 78.4%). Simultaneously, the fusion vector reduced the feature dimensionality from 83 to 16, decreasing computational complexity and storage requirements. In practical applications, this method has been deployed in online monitoring systems of multiple substations, improving the average fault warning lead time from 1.5 days to 3.2 days, providing maintenance personnel with more sufficient response time.

[0066] The fault feature propagation model method effectively integrates multimodal data by mining the propagation relationships and evolution patterns between features, thereby improving the expressive power of features and the accuracy of fault diagnosis. This method is independent of specific equipment types or fault modes, has broad applicability, and can be extended to the field of condition monitoring and fault diagnosis of various equipment within substations.

[0067] Example 3

[0068] This embodiment provides a further detailed description of step S2: feature importance assessment and optimization, based on embodiment 1.

[0069] The initial feature fusion vector is constructed based on multi-source heterogeneous data integration. Taking a fault case of an oil-immersed transformer in a 220kV substation as an example, features are extracted from multi-source data such as vibration sensors, temperature sensors, and dissolved gas analysis. These features include 128 dimensions such as oil temperature, top oil temperature, load current, dissolved hydrogen, dissolved methane, vibration spectrum, and acoustic characteristics. Each feature is normalized, with a value range of [0,1]. All features are concatenated to form an initial feature fusion vector with 128 dimensions, which serves as the basis for subsequent Bayesian network construction.

[0070] The Bayesian network model is constructed starting from the initial feature fusion vector, treating each feature as a node to build a conditional node set of 128 nodes. Initial connections between nodes are determined through correlation analysis; the Pearson correlation coefficient between features is calculated, and an initial connection is established when the absolute value of the correlation coefficient is greater than 0.4. In the transformer case, the correlation coefficient between oil temperature and top-layer oil temperature is 0.92, so a connection is established; the correlation coefficient between oil temperature and vibration spectrum is 0.21, so no connection is established. These initial connections form the skeleton of the directed acyclic graph, and then the direction of the edges is determined using the PC algorithm, forming a causal edge set.

[0071] Conditional probability calculation is a core step in Bayesian network construction. For each pair of adjacent nodes A and B, the conditional probability P(B|A) is calculated, representing the probability that node B will take a certain value given that node A takes a specific value. In implementation, continuous features are discretized into 10 intervals, and a conditional probability table is constructed. Taking the influence of oil temperature (A) on top oil temperature (B) as an example, when the oil temperature is in the interval [60℃, 65℃], the conditional probability that the top oil temperature is in the interval [65℃, 70℃] is 0.83. A complete conditional probability table is generated by statistically analyzing all feature pairs. For transformer data, 2000 samples were collected to ensure the accuracy of the conditional probability estimation.

[0072] Mutual information calculation measures the degree of information sharing between nodes. For a node pair A and B, mutual information I(A;B) represents the amount of information that reduces the uncertainty about B by observing A. The calculation is based on the joint probability distribution and marginal probability distribution in the conditional probability table to obtain the mutual information value. Higher mutual information indicates a stronger correlation between the two nodes. In the transformer case, the mutual information between oil temperature and top-layer oil temperature is 0.78, the mutual information between oil temperature and dissolved hydrogen is 0.42, and the mutual information between oil temperature and vibration spectrum is 0.15, indicating the highest degree of information sharing between oil temperature and top-layer oil temperature.

[0073] The determination of causal strength between nodes comprehensively considers conditional probability and mutual information. Causal strength is defined as the weighted sum of conditional probability and normalized mutual information, with weights of 0.6 and 0.4, respectively. For the causal strength from node A to B, the weighted sum of the average conditional probability and normalized mutual information under all conditions is taken. In the transformer case, the causal strength of oil temperature on top oil temperature is 0.6 × 0.83 + 0.4 × 0.78 = 0.81, and the causal strength of oil temperature on dissolved hydrogen is 0.6 × 0.51 + 0.4 × 0.42 = 0.47. The calculation of causal strength provides edge weight information for the construction of the graph neural network.

[0074] Graph Neural Networks (GNNs) are constructed based on the structure and causal strength of Bayesian networks. They map the conditional node set and causal edge set to graph-structured data. Node features are the corresponding feature values ​​in the initial feature fusion vector, and edge weights are the calculated causal strengths. The GNN employs a three-layer structure, including a graph convolutional layer, a graph attention layer, and an output layer. The first layer's graph convolutional operation aggregates the features of neighboring nodes, with edge weights representing causal strength. The input is 128-dimensional node features, and the output is a 64-dimensional node embedding. The second layer's graph attention mechanism further considers the importance differences between nodes, generating 32-dimensional node representations. The third layer maps these node representations to propagation influence scores.

[0075] Topological structure information extraction is achieved through graph convolution operations. The graph convolutional layer receives node features and an adjacency matrix as input, where each element of the adjacency matrix represents the causal strength between corresponding nodes. The convolution operation weights and aggregates the features of the central node and its neighboring nodes, with the weights proportional to the causal strength. In the transformer example, the first-order neighbors of the oil temperature node include nodes such as the top-level oil temperature and dissolved hydrogen. Through the convolution operation, the features of these neighbors are weighted according to causal strength and combined with the features of the oil temperature node to generate a new representation of the oil temperature node. After graph convolution, the feature values ​​of the oil temperature node change from the initial [0.65, 0.58... 0.72] to [0.71, 0.63... 0.68], representing the node representation after incorporating the information from neighboring nodes.

[0076] The integration of causal strength information between nodes is achieved through a graph attention mechanism. The graph attention layer calculates the attention coefficient between node pairs, taking into account the similarity of node representations and a predefined causal strength. For each node, the importance of its neighboring nodes is determined by the attention coefficient. In the transformer case, the attention coefficients of the oil temperature node to its neighboring nodes, top-level oil temperature and dissolved hydrogen, are 0.65 and 0.35, respectively, indicating that top-level oil temperature has a greater impact on the oil temperature node. The graph attention mechanism enhances the expressive power of the feature propagation model by selectively aggregating neighbor information.

[0077] The propagation influence calculation is based on the output of a graph neural network. Propagation influence is defined as a comprehensive measure of a node's centrality and propagation ability within the graph. The calculation includes three steps: feature propagation simulation, influence range calculation, and stabilization time assessment. Feature propagation simulation uses a random walk algorithm, starting from each node and simulating the propagation process of feature information based on causal strength as the transition probability. Influence range calculation is the proportion of other nodes affected by a given node out of the total number of nodes. Stabilization time assessment is the number of steps required for the propagation process to reach a stable state. Taking the transformer oil temperature node as an example, its influence range is 47 / 128 = 0.37, indicating that oil temperature changes affect 37% of other features; the stabilization time is 12 steps, indicating that the propagation effect tends to stabilize after 12 steps. Combining these indicators, the propagation influence of the oil temperature node is calculated to be 0.58, ranking 8th among all features.

[0078] The importance score of a feature node is calculated based on its propagation influence. The importance score is defined as the weighted sum of propagation influence and the node's own information content, with weights of 0.7 and 0.3, respectively. The node's own information content is calculated using information entropy, reflecting the uncertainty and richness of information of the feature. For the transformer oil temperature node, its propagation influence is 0.58, its information entropy is 0.85 (after normalization), and its importance score is 0.7 × 0.58 + 0.3 × 0.85 = 0.66. Similarly, the importance scores of all 128 features are calculated to obtain the complete feature importance distribution.

[0079] The feature selection threshold is determined based on the statistical distribution of importance scores. The mean μ and standard deviation σ of the importance scores for all features are calculated, and the threshold is set to μ + 0.5σ. In the transformer case, the mean importance score of the 128 features is 0.45, and the standard deviation is 0.18. The selection threshold is calculated as 0.45 + 0.5 × 0.18 = 0.54. Choosing this threshold ensures that a sufficient number of important features are selected while excluding a large number of unimportant features. Features with scores higher than the threshold are considered to be major influence features. In the transformer case, 42 major influence features were selected from the initial 128 features, including key indicators such as oil temperature, top oil temperature, load current, and dissolved hydrogen.

[0080] The selection of key influence features was based on importance scores and threshold comparisons. Features were sorted in descending order of importance scores, and features with scores higher than the threshold were selected as key influence features. In the transformer case, the top 5 features with the highest importance scores were: dissolved hydrogen (0.78), top oil temperature (0.74), dissolved acetylene (0.72), load current (0.69), and dominant frequency of the vibration spectrum (0.67). These features are of significant value for fault identification and play a crucial role in the feature propagation network. Through this selection process, the feature dimension was reduced from 128 to 42, significantly reducing the redundancy of the feature space.

[0081] The optimized feature fusion vector generation process involves reorganizing and enhancing key influential features, standardizing the selected 42 features to ensure consistent value ranges. Then, features are grouped based on correlation and complementarity, forming seven feature groups, including temperature, electrical, and gas features. Principal component analysis (PCA) is applied to the features within each group to reduce dimensionality, retaining the principal components necessary to explain 90% of the variance. For example, the temperature feature group is reduced from 8 features to 3 principal components, and the gas feature group from 10 features to 4 principal components. Finally, the principal components of each group are concatenated to form the optimized feature fusion vector with a dimension of 24. Compared to the initial 128-dimensional vector, the optimized vector not only significantly reduces dimensionality but also retains key information.

[0082] The validation of the optimized feature fusion vector was evaluated through a fault diagnosis task. Fault diagnosis models were trained using both optimized and initial vectors, and their performance differences were compared. Five-fold cross-validation was employed. Results showed that the model using optimized vectors achieved an accuracy of 94.7%, while the model using initial vectors achieved an accuracy of 87.3%. The training time for the optimized vector model was only 28% of that for the initial vector model, significantly improving computational efficiency. The F1 score of the optimized vector model reached 0.93, higher than the 0.85 of the initial vector model. This indicates that feature selection through propagation influence not only reduces dimensionality but also improves model performance. In practical applications, this method has been successfully applied to equipment condition monitoring systems in multiple substations, increasing the average fault warning lead time from 2.1 days to 4.5 days, providing sufficient response time for preventative maintenance.

[0083] The combination of Bayesian networks and graph neural networks provides a theoretical foundation and technical implementation path for feature importance assessment. Bayesian networks capture the conditional independence relationships and causal strength between features, while graph neural networks further integrate topological information, making feature importance assessment more comprehensive and accurate. This method is not dependent on specific fault types or equipment types, has broad applicability, and can be extended to the field of condition monitoring and fault diagnosis of various power equipment.

[0084] Figure 2 This is a bar chart comparing the performance of feature selection methods based on propagation influence in this invention. The chart shows the performance comparison of three different methods in the field of fault diagnosis: the initial feature fusion vector method (traditional benchmark method), the traditional feature selection method (existing technical solution), and the optimized feature fusion vector method (the solution proposed in this embodiment). The chart compares three key indicators: fault diagnosis accuracy, fault warning lead time, and improvement in computational efficiency. In terms of fault diagnosis accuracy, the three methods achieve 87.3%, 91.2%, and 94.7%, respectively, with the optimized feature fusion method performing best. Regarding fault warning lead time, the three methods achieve 2.1 hours, 3.3 hours, and 4.5 hours, respectively, with the optimized method achieving 2.4 hours faster than the traditional benchmark method. In terms of computational efficiency improvement, compared to the traditional benchmark method, the traditional feature selection method improves by 62.8%, while the optimized feature fusion method achieves a 78.2% efficiency improvement. The horizontal comparison shows that the optimized feature fusion vector method proposed in this embodiment outperforms the other two methods in all three key performance indicators, demonstrating its significant advantages in the field of fault diagnosis.

[0085] Example 4

[0086] This embodiment, based on embodiment 1, provides a further detailed explanation of step S3: construction and optimization of the fault feature knowledge graph.

[0087] likeFigure 3 As shown, the method includes:

[0088] The process of constructing a fault feature knowledge graph by optimizing feature fusion vectors relies on features extracted from multi-source equipment monitoring data. Taking a circuit breaker fault case in a 330kV substation as an example, the fusion vector dimension is set to 128, containing multi-dimensional information such as the mechanical features of the circuit breaker operating mechanism, the electrical features of the contact system, and environmental condition features. During knowledge graph construction, each fusion vector is treated as a node, and the edge weights between nodes are determined by calculating the cosine similarity between vectors. Connections are established between node pairs with a similarity higher than 0.45 to form the initial knowledge graph. In the circuit breaker case, a knowledge graph with 96 nodes and 285 edges was constructed, covering the operating status and potential fault types of each component of the circuit breaker.

[0089] A causal discovery algorithm based on the invariance principle calculates the causal strength between adjacent nodes. This algorithm is based on a core assumption: true causal relationships exhibit stability under different environmental conditions. In its implementation, the device operating data is first categorized into eight types according to environmental conditions, including normal load, overload, low temperature, and high temperature. For each environmental type, a conditional probability distribution matrix is ​​constructed, and the conditional entropy between node pairs is calculated. Conditional entropy reflects the predictive ability of one node for another; a lower value indicates stronger predictive ability.

[0090] Taking circuit breaker contact temperature and operating mechanism vibration as two examples, the conditional entropies under eight environmental conditions are 0.28, 0.31, 0.26, 0.33, 0.29, 0.27, 0.30, and 0.32, respectively. The standard deviation of these conditional entropies is 0.025, which is much lower than the standard deviation of random relationships (usually greater than 0.1), indicating a stable relationship between the two. A preliminary causal score is calculated based on the inverse index of the standard deviation, and then weighted according to topological features (such as node centrality and connection strength) to obtain the final causal strength. In this circuit breaker case, the causal strength of contact temperature on operating mechanism vibration is 0.37, while the causal strength of operating mechanism vibration on contact temperature is 0.72, indicating that the operating mechanism problem is the cause of the abnormal contact temperature.

[0091] The multi-level node representation is constructed using a graph convolutional network, which consists of three layers. The update rule of each layer considers the node's own features, the features of its neighboring nodes, and their causal strength. The first layer takes the original fused vector as input and outputs a dimension of 96; the second layer takes the output of the first layer as input and outputs a dimension of 64; and the third layer outputs a dimension of 32.

[0092] Taking the circuit breaker operating mechanism node as an example, its initial fusion vector, after processing in the first layer, is fused with information from neighboring nodes (such as the contact system and control circuit). The feature value distribution changes from [-0.9, 0.8] to [-0.7, 0.6], narrowing the numerical range but making the features more focused. The second layer further integrates second-order neighbor information, such as the insulation system state. The third layer captures global structural information, forming the final multi-level representation. Experiments show that compared with the single-level representation, the multi-level representation improves the accuracy of fault classification tasks by 14.2%, reaching 93.5%.

[0093] A fault feature propagation model analyzes the dynamic characteristics of node propagation. This model is based on the Susceptible-Infected-Recovery (SIR) propagation framework, treating equipment state characteristics as propagable information. The parameters in the propagation model are obtained through training with historical fault data, such as the propagation rate β and recovery rate γ. For the propagation of abnormal vibration in the circuit breaker operating mechanism, β is set to 0.28 and γ to 0.13, indicating that the vibration abnormality has moderate propagation capability and low recovery capability.

[0094] During the propagation state vector construction process, the multi-level representation vector of each node is combined with its current propagation state (susceptible state, infected state, or recovering state) to form an enhanced state vector. For vibration nodes of the operating mechanism, they are initially set to the infected state (value 1), and connected nodes are set to the susceptible state (value 0). After calculation by the propagation model, the state values ​​of adjacent nodes, such as the contact system and control loop, become 0.23 and 0.18, respectively, indicating the possibility of vibration anomalies propagating to these components.

[0095] The cumulative causal impact calculation comprehensively considers both direct and indirect effects. For each feature node, its cumulative impact is calculated iteratively. During the iteration process, the impact value of the node is updated in each round, considering the impact of directly connected nodes and the causal strength. After 10 iterations, the impact value tends to stabilize, yielding the final cumulative impact value. In the circuit breaker case, the cumulative impact value of the vibration node on the contact temperature is 0.81, which is much higher than the impact value of 0.42 on the control circuit, indicating that the vibration anomaly has a greater impact on the contact system. Eigenvalue analysis of the cumulative impact matrix reveals the main direction and intensity of the fault impact. The largest eigenvalue is 3.17, and the corresponding eigenvector points to the propagation path from the operating mechanism to the contact system, indicating that this is the most important fault propagation path.

[0096] The analysis of propagation dynamics combines propagation state vectors and cumulative causal effects. The propagation state vectors reflect short-term dynamics, while the cumulative causal effects reflect long-term trends. During the analysis, three key indicators are calculated: node propagation rate, influence range, and settling time. The propagation rate is defined as the growth rate of the influence value per unit time; the influence range is the ratio of the number of affected nodes to the total number of nodes; and the settling time is the time step required for the propagation influence to reach a stable state.

[0097] For the abnormal vibration of the operating mechanism, the propagation rate was 0.15 / hour, the affected area was 34 / 96 = 0.35, and the stabilization time was 18 hours. This indicates that the vibration abnormality propagated at a moderate speed and had a limited affected area, but required a relatively long time to stabilize. In contrast, the abnormal contact temperature had a propagation rate of 0.28 / hour, an affected area of ​​0.52, and a stabilization time of 11 hours, exhibiting a faster propagation speed and a wider affected area.

[0098] The temporal evolution features are constructed based on node state time series analysis, collecting state data for each node over a continuous 72-hour period to extract temporal features such as trends, periodicity, and abrupt changes. Trend features are extracted using the moving average method, periodicity features using Fourier analysis, and abrupt changes are identified using a change point detection algorithm. Taking the circuit breaker contact temperature as an example, its 72-hour evolution shows an upward trend (slope of 0.27℃ / hour), comprising 4 cycles (approximately 18 hours per cycle), with 2 significant abrupt changes (at the 13th and 47th hours). The temporal feature vector is set to 16 dimensions, including 4 trend parameters, 8 periodic parameters, and 4 abrupt change parameters. Different fault types exhibit different temporal characteristics; for example, operating mechanism faults typically show significant abrupt changes, while insulation aging shows a stable upward trend.

[0099] The causal evolution feature construction focuses on the dynamic changes in causal relationships between nodes. Causal evolution features are obtained by calculating the rate of change of causal intensity within different time windows. Specifically, 72 hours of data are divided into six 12-hour windows, and the change in causal intensity between adjacent windows is calculated. Taking the impact of operating mechanism vibration on contact temperature as an example, the causal intensities for the six windows are 0.65, 0.69, 0.72, 0.76, 0.80, and 0.83, respectively, showing a stable upward trend with an average rate of change of 0.036 per window. The causal evolution feature vector has a dimension of 12, including the rate of change and volatility of each major causal relationship. Causal evolution features can reflect the intrinsic mechanism of fault development, such as the process where vibration gradually leads to poor contact at the contact point, which in turn causes a temperature increase.

[0100] The evolution patterns of fault characteristics are obtained by combining temporal and causal evolution features. A weighted fusion method is used, with a weight of 0.6 for temporal features and 0.4 for causal features. The fused feature vector has 28 dimensions, with the first 16 dimensions derived from temporal features and the last 12 dimensions from causal features. Based on the fused features, a clustering algorithm is used to identify typical evolution patterns. In the circuit breaker case, five typical evolution patterns were identified: sudden (e.g., mechanical jamming), gradual (e.g., insulation aging), fluctuating (e.g., poor contact), periodic (e.g., problems caused by temperature fluctuations), and composite (multi-factor coupling).

[0101] For abnormal vibration of the operating mechanism, its evolution pattern was identified as gradual, with a confidence level of 0.87. This type of fault usually begins with internal mechanical wear of the operating mechanism, gradually affecting operating accuracy, leading to poor contact of the contacts, and eventually causing electrical faults. The typical development time is 3-7 days, with early manifestations of prolonged operating time (an increase of 3-5 ms), mid-term manifestations of abnormal vibration (an increase of 30-50%), and late-term manifestations of contact temperature rise (an increase of 15-25℃).

[0102] The practical application of evolutionary patterns is reflected in the formulation of predictive maintenance strategies. Based on the identified evolutionary patterns, differentiated detection frequencies and early warning thresholds are set. For progressive faults such as abnormal vibration of the operating mechanism, the detection frequency is set to once every 8 hours, and the early warning threshold is set to 1.3 times the normal value. When the detected vibration amplitude exceeds the threshold and shows a stable upward trend, combined with causal evolution characteristics showing an increased impact on contact temperature, the early warning mechanism is triggered. In a practical application at a 330kV substation, this method successfully predicted three circuit breaker operating mechanism faults, issuing early warnings an average of 72 hours in advance, 48 hours earlier than traditional early warning methods, providing maintenance personnel with sufficient response time. Practice shows that the early warning accuracy based on evolutionary patterns reaches 91.2%, and the false alarm rate is reduced to 8.3%, a significant improvement compared to traditional early warning methods (accuracy 78.5%, false alarm rate 21.4%).

[0103] Example 5

[0104] This embodiment, based on embodiment 3, provides a specific implementation method for constructing a multi-level node representation of a fault feature knowledge graph, including:

[0105] Construct a nonlinear feature transformation matrix and an adaptive bias term, and use the nonlinear feature transformation matrix and the adaptive bias term to perform a nonlinear transformation on the optimized feature fusion vector. Construct a fault feature knowledge graph based on the optimized feature fusion vector after the nonlinear transformation.

[0106] Based on the fault feature knowledge graph, an environment-specific conditional entropy matrix is ​​constructed. The causal effect between nodes is calculated according to the conditional entropy matrix. The causal effect between nodes is weighted and combined using the fault feature knowledge graph to obtain the causal strength between adjacent nodes.

[0107] A dynamic causal strength threshold is set for the nodes in the fault feature knowledge graph. The associated neighbors of the nodes are filtered according to the dynamic causal strength threshold. An attention aggregation function is constructed using the feature representation of the associated neighbors and the causal strength. The node features are then fused at multiple levels through the attention aggregation function to obtain the multi-level node representation of the fault feature knowledge graph.

[0108] Nonlinear feature transformation is achieved through a feature transformation matrix and an adaptive bias term. The feature transformation matrix is ​​a square matrix of dimension d×d, where d is the dimension of the optimized feature fusion vector, typically set to 64 or 128. Taking a 500kV substation GIS equipment fault case as an example, the optimized feature fusion vector has a dimension of 128, and the corresponding feature transformation matrix has a dimension of 128×128. The initial value of this matrix is ​​generated using an orthogonal initialization method to ensure the preservation of feature information during the transformation process. The adaptive bias term is a vector of length d, initially set to zero and gradually adjusted during training.

[0109] When applying a nonlinear transformation to the optimized feature fusion vector, it is first multiplied by the feature transformation matrix, then an adaptive bias term is added, and finally a nonlinear mapping is performed using the ReLU activation function. In the GIS equipment case, after the nonlinear transformation, the distribution range of relevant feature values ​​of the optimized feature fusion vector of an SF6 gas density sensor was adjusted from [-0.8, 0.9] to [0, 0.95], making subsequent causal relationship analysis more stable and reliable.

[0110] The fault feature knowledge graph is constructed based on feature vectors obtained through nonlinear transformation. Nodes in the graph represent equipment state characteristics, such as SF6 gas density and insulator surface contamination, with each node associated with a feature vector. Edges represent potential relationships between nodes, and initial edge weights are determined by calculating the cosine similarity of the node feature vectors. In the GIS equipment case study, a knowledge graph with 78 nodes and 312 edges was constructed, covering multiple aspects such as the equipment's physical state, electrical characteristics, and environmental conditions.

[0111] During the construction process, the node set is first defined, including all monitoring points and fault types; then, the feature similarity between node pairs is calculated; finally, the connection relationship of the edges is determined according to the similarity threshold (set to 0.45). For a certain pair of nodes, such as the SF6 gas density and temperature sensor, its feature vector cosine similarity is 0.63, which is higher than the threshold, so a connection is established in the graph.

[0112] Constructing an environment-specific conditional entropy matrix is ​​a core step in causal discovery. This matrix reflects the conditional dependencies between nodes under different environmental conditions. During construction, the device operating environment is first categorized into several types based on historical data, such as normal load, heavy load, and high ambient temperature. For GIS devices, a total of six environmental conditions were defined.

[0113] For each environmental type, the conditional entropy between node pairs is calculated, forming a conditional entropy submatrix. Conditional entropy calculation is based on the probability distribution of node features, achieved through feature discretization and frequency statistics. For example, under high ambient temperature conditions, the conditional entropy values ​​of SF6 gas density with other nodes range from [0.25, 0.92], with the lowest conditional entropy (0.25) with the temperature sensor node, indicating a strong conditional dependency between them. The conditional entropy submatrices for each environmental condition are merged to form a complete conditional entropy matrix with dimensions 78×78×6, corresponding to the conditional entropy relationships of 78 nodes under 6 environmental conditions.

[0114] The calculation of causal effects between nodes is based on the principle of invariance. The principle of invariance states that if node A is a cause of node B, then the mechanism by which A influences B should remain stable under different environments. Therefore, causal relationships are assessed by comparing the degree of change in conditional entropy under different environments. Specifically, for each pair of adjacent nodes, the standard deviation of their conditional entropy under different environments is calculated; the smaller the standard deviation, the more stable the relationship and the higher the probability of causality.

[0115] In the GIS equipment case study, the standard deviation of the conditional entropy of the SF6 gas density and temperature sensor nodes under six environmental conditions was 0.04, significantly lower than the standard deviation of other nodes (average 0.21), indicating a causal relationship between them. A preliminary causal effect value was calculated based on the standard deviation by taking the reciprocal of the standard deviation and normalizing it so that the sum of all effect values ​​equaled 1. For the SF6 gas density and temperature sensor node pair, the preliminary causal effect value was calculated to be 0.85, indicating a strong causal relationship between them.

[0116] The weighted combination of causal strengths takes into account the topological structure of the knowledge graph. For each pair of adjacent nodes, the final causal strength is calculated by combining the initial causal effect value and the node's connectivity in the graph. Connectivity is quantified using indicators such as degree centrality and path centrality. Degree centrality represents the number of direct connections between nodes, while path centrality represents the transitive effect of a node in the graph. In GIS equipment, the degree centrality of the SF6 gas density node is 0.47, the path centrality is 0.54, and the overall topological index is 0.6.

[0117] In the weighted combination, the initial causal effect value weight is set to 0.7, and the topological index weight is set to 0.3. Therefore, the final causal strength of the SF6 gas density and temperature sensor node is 0.85×0.7+0.6×0.3=0.77, indicating that temperature is an important cause of SF6 gas density changes. Similarly, the causal strength of all adjacent node pairs in the knowledge graph is calculated to form a causal strength matrix, providing a foundation for the construction of multi-level node representations.

[0118] The dynamic causal strength threshold is adaptively adjusted based on the node's connectivity. For nodes with high connectivity, a higher threshold is set to filter out weak causal relationships; for nodes with low connectivity, a lower threshold is used to retain more information. The threshold calculation is proportional to the node's degree, with a baseline threshold set to 0.5 and an adjustment range of [0.3, 0.7]. The specific formula is: the node threshold equals the baseline threshold plus the difference between the node's degree and the average degree multiplied by an adjustment coefficient (0.05).

[0119] Taking the SF6 gas density node as an example, its connectivity is 8, which is higher than the average connectivity of 5.2. Therefore, the set causality strength threshold is 0.5 + (8 - 5.2) × 0.05 = 0.58. Based on this threshold, 5 related neighbors are selected from the 8 neighboring nodes, including the temperature sensor (causality strength 0.77) and the gas pressure sensor (causality strength 0.72). The feature representations of these related neighbors will be used for subsequent multi-level node representation construction.

[0120] The attention aggregation function is constructed using causal strength as the attention weight. This function employs a multi-head attention mechanism, typically using 4 or 8 attention heads. Each attention head independently calculates its attention score, and the results are then combined. The attention score calculation combines node feature similarity and causal strength, with the former weighted at 0.4 and the latter at 0.6. Feature similarity is calculated using the cosine similarity of the feature vectors, while causal strength is directly derived from the aforementioned calculation.

[0121] In the GIS equipment case, for the SF6 gas density node, its feature similarity with the temperature sensor node is 0.65, and its causal strength is 0.77. The calculated attention score is 0.65 × 0.4 + 0.77 × 0.6 = 0.722. The attention scores of all associated neighbors are normalized using softmax to obtain the final aggregation weight. For example, the normalized weight of the temperature sensor node is 0.31, and the normalized weight of the gas pressure sensor node is 0.28. Feature aggregation uses a weighted summation method, concatenating the node's own features with the weighted neighbor features. After linear transformation and nonlinear activation, the node representation for this layer is obtained.

[0122] Multi-level node representation is achieved by stacking multiple layers of attention aggregation functions. Typically, 2-3 layers are used, with the output of each layer serving as the input to the next. Different layers capture different ranges of structural information: the first layer primarily aggregates information from directly adjacent nodes, the second layer further integrates information from second-order neighbors, and the third layer captures a broader range of relationships. In the GIS device example, a 3-layer structure was used, with output dimensions of 96, 64, and 32, respectively. Each layer contains an attention aggregation function, a residual connection, and a layer normalization component.

[0123] Residual connections, by adding input features to aggregated features, help deep networks more effectively propagate gradients. Layer normalization stabilizes the feature distribution and accelerates the training process. Taking the SF6 gas density node as an example, after the first layer of aggregation, its representation vector dimension is 96, containing information directly related to factors such as temperature and pressure; after the second layer of aggregation, the representation vector dimension is reduced to 64, further incorporating second-order relationships, such as fluctuations in other parameters caused by temperature changes; after the third layer of aggregation, the representation vector dimension is 32, effectively capturing more complex multi-level relationships. Compared to the initial representation, the multi-level representation improves the accuracy of downstream fault diagnosis tasks by 18.3%, reaching 94.2%.

[0124] The evaluation and application of multi-level node representation employed various technical indicators. Through visualization analysis, the t-SNE algorithm was used to map the 32-dimensional representation vector to a two-dimensional space. This revealed a clear spatial clustering of nodes with similar fault types; for example, nodes related to SF6 leakage formed a dense cluster with center coordinates approximately (-15.3, 24.7) and a cluster radius of approximately 5.2. Classification task evaluation showed that the multi-level representation achieved a fault classification accuracy of 94.2%, significantly higher than the baseline method using only original features (75.9%). In practical predictive maintenance of GIS equipment, combining multi-level node representation with a time-series prediction model improved the average early warning time for SF6 leakage faults by 2.7 days, from 1.3 days to 4.0 days, providing maintenance personnel with more sufficient response time.

[0125] Especially under complex environmental conditions, such as extreme temperatures or load fluctuations, the false alarm rate of traditional diagnostic methods is 18.6%, while the false alarm rate of the method based on multi-level node representation is only 5.2%, significantly improving the reliability of early warning and reducing unnecessary inspection costs. Furthermore, this method demonstrates good generalization ability across different manufacturers and types of GIS equipment, with an average accuracy rate of 91.7%, indicating its broad applicability.

[0126] Example 6

[0127] This embodiment provides a further detailed description of step S4: propagation link analysis and fault diagnosis, based on embodiment 1.

[0128] The construction of the multi-head attention mechanism begins with a node feature matrix, which contains features extracted from multimodal data of the power equipment. For a 500kV substation main transformer case, features are extracted from data obtained from sensors such as temperature, vibration, and sound, forming a node feature matrix with dimensions 156×64. Each node represents a specific equipment state feature, such as abnormal oil temperature or abnormal winding temperature. The multi-head attention mechanism consists of eight attention heads, each responsible for processing associated features at different time scales. During construction, three parameter matrices are first defined: a query matrix, a key matrix, and a value matrix, each with dimensions 64×8. Multiplying the feature matrix by these three parameter matrices yields three projection matrices: query, key, and value.

[0129] Propagation probability calculation is a core step in the multi-head attention mechanism. The original attention score is obtained by multiplying the query matrix by the transpose of the key matrix and dividing by a scaling factor of 8. These scores are then normalized using the softmax function to obtain the attention weights, i.e., the propagation probabilities between nodes. For example, in the case of abnormal oil temperature, the propagation probability from the oil temperature rise node to the winding insulation deterioration node is calculated to be 0.83, indicating a strong propagation relationship between the two. Similarly, the propagation probabilities between all node pairs in the knowledge graph are calculated to form a propagation probability matrix, providing a foundation for subsequent analysis.

[0130] Dynamic correlation extraction is achieved through a multi-head attention mechanism using different heads. Heads 1 and 2 are responsible for extracting abrupt change features at short timescales (1-5 minutes); heads 3, 4, and 5 are responsible for extracting fluctuation features at medium timescales (10-60 minutes); and heads 6, 7, and 8 are responsible for extracting trend features at long timescales (6-24 hours). For each timescale, the corresponding node correlation is calculated. Taking oil temperature and winding temperature as an example, the short-term correlation is 0.65, indicating a moderate abrupt change correlation; the medium-term correlation is 0.78, indicating a strong fluctuation correlation; and the long-term correlation is 0.92, indicating a high long-term trend correlation.

[0131] The weighted average method is used to combine the correlations across multiple time scales. Considering that fault propagation usually exhibits a long-term trend, different weights are assigned to different time scales: 0.5 for long-term scales, 0.3 for medium-term scales, and 0.2 for short-term scales. For each pair of nodes, their correlations across each time scale are combined according to their weights to obtain the final dynamic correlation. Taking oil temperature and winding temperature as an example, the final dynamic correlation is 0.65×0.2+0.78×0.3+0.92×0.5=0.84, indicating that the two have a strong overall correlation.

[0132] The link importance assessment function integrates two indicators: propagation probability and dynamic correlation. Considering that propagation probability more directly reflects the likelihood of fault propagation, while dynamic correlation reflects the strength of the correlation, a higher weight (0.6) is assigned to propagation probability, and a weight (0.4) to dynamic correlation. For each potential propagation link, its combined value is calculated; for example, the combined value of the link from abnormal oil temperature to insulation degradation is 0.83 × 0.6 + 0.84 × 0.4 = 0.836. To make the assessment results more distinctive, a nonlinear mapping function is applied to convert the combined value into an importance score. This mapping function is similar to the sigmoid function, but the parameters are adjusted to make the intermediate region more sensitive. The calculated importance score for the above link is 0.78.

[0133] The initial propagation link selection is based on an importance score and a preset threshold, which is determined statistically from historical fault data and set to 0.65 in this example. For all links in the knowledge graph, an importance score is calculated, and links with scores higher than the threshold are selected as initial propagation links. In the transformer case, 35 initial propagation links were selected from 128 links, significantly reducing the complexity of subsequent analysis while preserving the main fault propagation relationships.

[0134] Conditional causal entropy calculation is a crucial step in assessing the causal determinism of a link. For each initial propagation link, a probability distribution matrix of the node sequence is first constructed, with each element representing the propagation probability between the corresponding nodes. Then, the dynamic correlation degree extracted using a multi-head attention mechanism is used as weights to adjust the probability distribution matrix. Based on the adjusted probability distribution, the conditional entropy of the node sequence is calculated. Taking the oil temperature anomaly-insulation degradation-partial discharge link as an example, the dynamic correlation degrees between nodes are 0.84 and 0.79, and the propagation probabilities are 0.83 and 0.75, respectively. The calculated conditional causal entropy is 0.37. The lower the conditional causal entropy, the more certain the causal relationship in the link, and the higher the probability of fault propagation.

[0135] The secondary assessment of link importance integrates importance score and conditional causal entropy. Since conditional causal entropy is an inverse indicator (lower values ​​are better), its inverse value (1 - entropy value) is taken and weighted with the importance score. Considering that conditional causal entropy better reflects the causal certainty of a link, a higher weight of 0.55 is assigned to the inverse entropy value, and a weight of 0.45 is assigned to the importance score. Taking the aforementioned link as an example, the conditional causal entropy is 0.37, the importance score is 0.78, and the secondary assessment score is 0.55 × (1 - 0.37) + 0.45 × 0.78 = 0.83. This score comprehensively considers the propagation probability, dynamic correlation, and causal certainty of the link, providing a more comprehensive assessment of the link's importance.

[0136] The selection of primary paths is achieved through an adaptive threshold mechanism. This mechanism is based on the statistical distribution of historical equipment failure data and uses a percentile method to calculate the threshold. Specifically, secondary evaluation scores of historical equipment failure links are collected, and their 85th percentile value is calculated as the threshold, which is 0.75 in this example. Links with evaluation scores higher than the threshold are identified as primary failure propagation paths. After this step, the initial 35 propagation links are reduced to 12 primary paths. These paths form the backbone of failure propagation, providing a basis for subsequent knowledge graph optimization.

[0137] The knowledge graph topology optimization is implemented iteratively. The selected main paths are organized into a directed graph, where nodes represent device state characteristics, edges represent fault propagation relationships, and the initial weights of the edges are set to their corresponding dynamic correlation degrees. Then, a multi-head attention mechanism is used to iteratively update the weights of each edge in the graph. In each iteration, the attention scores of other nodes connected to each node are calculated. Based on these scores and the current edge weight, a new weight is calculated using the formula: New weight = 0.3 × Attention score + 0.7 × Current weight. The iteration process continues until the weight change between two adjacent iterations is less than 0.01. In the transformer case, after 5 iterations, a final knowledge graph with stable edge weights was obtained.

[0138] The critical fault links in the final knowledge graph have higher reliability. For example, the weights of each edge in the link of abnormal oil temperature – insulation degradation – partial discharge – winding deformation are optimized to 0.91, 0.94, and 0.90, respectively, with a comprehensive weight (product of the three) of 0.77, indicating that this link has a high probability of fault propagation. Similarly, the link of abnormal oil temperature – decreased cooling system efficiency – hotspot formation has a comprehensive weight of 0.82, which is also an important fault propagation path. A total of 8 critical links constitute the main network of transformer fault propagation, providing a reliable basis for fault diagnosis.

[0139] The fault diagnosis results are output based on the analysis of the final knowledge graph, identifying the nodes and links with the highest activation levels to determine the fault type and location. Then, based on the comprehensive link weight and node activation level, the fault severity is assessed, categorized into three levels: minor (0-0.4), moderate (0.4-0.8), and severe (0.8-1.0). Next, by comparing the current knowledge graph with historical graphs, the fault development trend is analyzed, categorized as stable, slowly deteriorating, and rapidly deteriorating. Finally, a diagnostic report is generated, including fault details, severity, development trend, and maintenance recommendations.

[0140] In the case of abnormal transformer oil temperature, the diagnostic results showed that there was insulation degradation caused by the abnormal oil temperature, with a moderate severity (score 0.76) and a slow deterioration trend. Further inspection and preventative maintenance were recommended within two weeks. This method has been practically applied in several 500kV substations, and compared with traditional diagnostic methods, the accuracy rate has increased by 15.7%, reaching 92.3%, providing strong support for equipment condition assessment and preventative maintenance.

[0141] Figure 4 This is a bar chart comparing the performance of fault propagation link analysis based on multi-head attention mechanism in this invention. The chart shows the performance comparison of three different methods across four key fault diagnosis indicators: this technical solution (a diagnostic method based on temporal feature extraction and graph optimization), the traditional graph analysis method (a diagnostic method based on traditional graph feature extraction), and basic link tracing (a traditional fault link tracing method). In terms of fault propagation path identification accuracy, the three methods achieve 92.7%, 76.3%, and 65.2%, respectively; key node identification accuracy is 89.5%, 71.8%, and 58.7%, respectively; propagation temporal prediction accuracy is 87.3%, 69.5%, and 54.3%, respectively; and diagnostic efficiency after graph optimization reaches 94.2%, 78.9%, and 67.1%, respectively. The comparison across the four dimensions shows that this technical solution significantly outperforms the other two methods in all evaluation indicators, especially in terms of diagnostic efficiency after graph optimization, which is 15.3 percentage points higher than the traditional graph analysis method and 27.1 percentage points higher than the basic link tracing method. This result fully verifies the comprehensive advantages of this technical solution in the field of fault diagnosis.

[0142] Example 7

[0143] This embodiment, based on embodiment 6, provides a specific implementation method for filtering out the main path of fault propagation, including:

[0144] The conditional causal entropy is evaluated a second time according to the link importance evaluation function. The link importance evaluation function uses a gated loop unit to extract the temporal features of the conditional causal entropy. The gated loop unit extracts the long and short-term temporal features of the conditional causal entropy through a reset gate and an update gate.

[0145] An adaptive threshold is constructed based on the hidden state of the gated loop unit. The long and short-term time series features are compared with the adaptive threshold to filter out the main path of fault propagation.

[0146] Faults often propagate, with one fault triggering a chain reaction of failures in other parts of the system. Preliminary importance assessments of each link have been obtained through conditional causal entropy calculations; further secondary evaluations are needed to identify the true primary paths of fault propagation.

[0147] The link importance evaluation function uses a gated recurrent unit (GRU) to extract temporal features from conditional causal entropy. In its implementation, the conditional causal entropy values ​​of each link within a historical time window are input as temporal data into the GRU network. The GRU network contains two key components: a reset gate and an update gate, used to learn the temporal variation characteristics of the conditional causal entropy.

[0148] The reset gate is used to determine how much previous state information to retain. For each time step t, the reset gate receives the current input x. t (i.e., the conditional causal entropy value at the current time step) and the hidden state h from the previous time step. {t-1} By calculating the weighted sum of the two inputs and applying the sigmoid activation function, a vector with values ​​ranging from 0 to 1 is generated. When the value of the reset gate is close to 0, it indicates that the historical information has been "reset" and the current hidden state mainly depends on the current input; when the value of the reset gate is close to 1, the historical information is completely preserved.

[0149] The update gate controls the proportion of information from the previous state passed to the current state. The update gate also receives the current input x. t The hidden state h of the previous time step {t-1} By calculating a weighted sum and applying the sigmoid activation function, a vector with values ​​ranging from 0 to 1 is generated. The update gate determines how much old information is retained and how much new information is introduced into the current state.

[0150] By leveraging the synergistic effect of reset and update gates, GRU can effectively capture the long- and short-term temporal variations of conditional causal entropy. Long-term characteristics reflect the stability importance of links over a longer period, while short-term characteristics reflect recent dynamic changes in the importance of links.

[0151] A time series analysis was performed on the conditional causal entropy of three critical links A-B, B-C, and A-C in an industrial system. Assuming 10 time steps of historical data were collected, the conditional causal entropy sequence for link A-B is [0.85, 0.83, 0.88, 0.90, 0.87, 0.89, 0.92, 0.91, 0.94, 0.95], the conditional causal entropy sequence for link B-C is [0.45, 0.48, 0.44, 0.46, 0.45, 0.47, 0.46, 0.48, 0.47, 0.49], and the conditional causal entropy sequence for link A-C is [0.65, 0.67, 0.64, 0.66, 0.68, 0.69, 0.67, 0.71, 0.73, 0.75].

[0152] These sequences are input into a GRU network with 64 hidden units, an input layer dimension of 1 (conditional causal entropy value of a single link), and an output layer dimension of 1 (link importance score). The input to the GRU network is a sequence of shape (10,1) (10 time steps, 1 feature per step).

[0153] For link A-B, when processing the 5th time step (value 0.87), it is assumed that the hidden state h4 of the previous time step contains information from the previous 4 time steps [0.85, 0.83, 0.88, 0.90]. The reset gate calculation result is 0.82, indicating that 82% of the historical information is retained; the update gate calculation result is 0.75, indicating that the new state integrates 75% of the candidate hidden states and 25% of the previous hidden state. In this way, GRU can effectively capture the upward trend of the conditional causal entropy of this link, that is, the characteristic that the importance of this link gradually increases over time.

[0154] The GRU network processes the conditional causal entropy sequences of links B-C and A-C, capturing their temporal variation characteristics. After processing all time steps, the GRU network outputs the final hidden state, which contains comprehensive information about the entire sequence.

[0155] An adaptive threshold is constructed based on the final hidden state of the GRU network. This adaptive threshold is dynamically adjusted according to the current system state, avoiding the limitations of a fixed threshold. Specifically, the final hidden state of the GRU network is mapped to a threshold coefficient through a fully connected layer. This coefficient is multiplied by the mean of the conditional causal entropy of all links to obtain the final adaptive threshold.

[0156] In the above case, assuming that after the GRU network processes the conditional causal entropy sequence of the three links, the resulting hidden state is mapped to generate a threshold coefficient of 1.2, and the mean of the conditional causal entropy of all links is 0.6, then the adaptive threshold is 1.2 × 0.6 = 0.72.

[0157] The GRU outputs for each link were compared with adaptive thresholds. The GRU output for link A-B was 0.85, greater than the threshold of 0.72; the GRU output for link B-C was 0.47, less than the threshold of 0.72; and the GRU output for link A-C was 0.74, greater than the threshold of 0.72. Therefore, the main fault propagation paths identified were A-B and A-C.

[0158] This link importance assessment method based on gated cyclic units (GRUs) fully considers the temporal variation characteristics of conditional causal entropy, and can more accurately identify critical fault propagation paths in the system compared to static analysis methods. In practical applications, the parameters of the GRU network, such as the number of hidden layer units and the number of layers, can be adjusted according to the system size and complexity to obtain the best assessment results. The selected main fault propagation paths can provide important basis for predictive maintenance and fault diagnosis of industrial systems.

[0159] Example 8

[0160] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0161] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0162] The processing unit executes the various methods and processes described above, such as methods S1 to S4. For example, in some embodiments, methods S1 to S4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1 to S4 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S4 by any other suitable means (e.g., by means of firmware).

[0163] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0164] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0165] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for diagnosing faults in power equipment, characterized in that, This method analyzes the propagation paths and influence of different features in the multimodal data of substations based on the fault feature propagation model to construct a knowledge graph. It then optimizes the knowledge graph by analyzing the dynamic characteristics and propagation links, and uses the optimized knowledge graph to output the fault diagnosis results of the substations.

2. The method for diagnosing faults in power equipment according to claim 1, characterized in that, The method includes the following steps: Acquire multimodal data of substation equipment, analyze the propagation path of different features in the multimodal data based on the fault feature propagation model, calculate the correlation degree between features according to the propagation path, extract feature combinations of the multimodal data based on the correlation degree, and generate an initial feature fusion vector; The fault feature propagation model is used to construct a feature importance evaluation system for the initial feature fusion vector. The propagation influence of features is analyzed based on the feature importance evaluation system. Based on the propagation influence, the main influential features are selected to generate an optimized feature fusion vector. A fault feature knowledge graph is constructed based on the optimized feature fusion vector. The propagation dynamic characteristics of nodes in the fault feature knowledge graph are analyzed based on the fault feature propagation model. The evolution law of fault features is identified according to the propagation dynamic characteristics. The association strength of the fault feature knowledge graph is updated based on the evolution law to obtain the optimized knowledge graph. The fault feature propagation model is used to perform propagation link analysis on the optimized knowledge graph. Based on the results of the propagation link analysis, the main paths of fault propagation are mined. The topology of the fault feature knowledge graph is optimized according to the main paths to obtain the final knowledge graph. The fault diagnosis results of the substation are output based on the final knowledge graph.

3. The method for diagnosing faults in power equipment according to claim 2, characterized in that, The method described above analyzes the propagation paths of different features in the multimodal data based on the fault feature propagation model, calculates the correlation between features according to the propagation paths, extracts feature combinations from the multimodal data based on the correlation, and generates an initial feature fusion vector, including: Construct a fault feature propagation model, which includes a set of feature nodes and a set of propagation edges; Calculate the propagation strength between nodes in the feature node set, where the propagation strength characterizes the association relationship between nodes, and determine the propagation path between the feature node sets based on the propagation strength; Based on the propagation path, feature sample pairs are constructed. The propagation dynamic characteristics of the nodes in the feature sample pairs are analyzed using the fault feature propagation model. The propagation dynamic characteristics include lag time, response speed, attenuation coefficient, and fluctuation correlation. The evolution law between nodes is identified according to the propagation dynamic characteristics. The evolution law includes linear following type, threshold triggering type, lag accumulation type, and periodic coupling type. The feature correlation degree corresponding to the propagation path is calculated based on the evolution law and propagation intensity. Based on the degree of feature association, a hierarchical clustering method is used to cluster the nodes in the feature node set, extract the feature combinations of the multimodal data, calculate the importance and complementarity of features within the feature combinations, obtain the combination score, and generate an initial feature fusion vector based on the feature combinations and their corresponding combination scores.

4. The method for diagnosing faults in power equipment according to claim 2, characterized in that, The aforementioned method involves constructing a feature importance evaluation system for the initial feature fusion vector using the fault feature propagation model, analyzing the propagation influence of features based on the feature importance evaluation system, selecting key influential features based on the propagation influence, and generating an optimized feature fusion vector, including: A Bayesian network model is constructed based on the initial feature fusion vector. The Bayesian network model includes a set of conditional nodes and a set of causal edges. The conditional probability and mutual information between adjacent nodes in the set of conditional nodes are calculated using the Bayesian network model. The causal strength between nodes is determined based on the conditional probability and the mutual information. A graph neural network is constructed based on the Bayesian network model and the causal strength between nodes. The conditional node set and the causal edge set are mapped to graph structure data. The node features are the corresponding feature values ​​in the initial feature fusion vector, and the edge weights are the calculated causal strength between nodes. The graph neural network is used to extract the topological structure information and the causal strength information between nodes of the conditional node set. The propagation influence of the feature nodes is calculated based on the topological structure information and the causal strength information between nodes. The importance score of the feature node is calculated based on the dissemination influence. The mean and standard deviation of the feature distribution are calculated based on the importance score. The feature screening threshold is determined based on the mean and standard deviation. Feature nodes with an importance score greater than the feature screening threshold are selected as the main influence features. Feature fusion is performed on the main influence features to generate an optimized feature fusion vector.

5. The method for diagnosing faults in power equipment according to claim 2, characterized in that, The method of constructing a fault feature knowledge graph based on the optimized feature fusion vector, analyzing the propagation dynamics of nodes in the fault feature knowledge graph based on the fault feature propagation model, and identifying the evolution law of fault features according to the propagation dynamics include: A fault feature knowledge graph is constructed based on the optimized feature fusion vector, and a causal discovery algorithm based on the invariance principle is introduced into the fault feature knowledge graph to calculate the causal strength between adjacent nodes in the fault feature knowledge graph. A multi-level node representation of the fault feature knowledge graph is constructed based on the causal strength. The multi-level node representation and the causal strength are input into the fault feature propagation model to construct the propagation state vector of the feature node. The cumulative causal influence of the feature node is calculated based on the propagation state vector. The propagation dynamic characteristics of the feature node are analyzed according to the cumulative causal influence and the propagation state vector, and three key indicators, node propagation rate, influence range and stabilization time, are calculated. Based on the propagation dynamic characteristics, the temporal evolution characteristics and causal evolution characteristics of the feature nodes are constructed. The temporal evolution characteristics and the causal evolution characteristics are combined to calculate the evolution law of the fault characteristics.

6. The method for diagnosing faults in power equipment according to claim 5, characterized in that, The method described above involves constructing a fault feature knowledge graph based on the optimized feature fusion vector, introducing a causal discovery algorithm based on the invariance principle into the fault feature knowledge graph, calculating the causal strength between adjacent nodes in the fault feature knowledge graph, and constructing a multi-level node representation of the fault feature knowledge graph based on the causal strength, including: Construct a nonlinear feature transformation matrix and an adaptive bias term, and use the nonlinear feature transformation matrix and the adaptive bias term to perform a nonlinear transformation on the optimized feature fusion vector. Construct a fault feature knowledge graph based on the optimized feature fusion vector after the nonlinear transformation. Based on the fault feature knowledge graph, an environment-specific conditional entropy matrix is ​​constructed. The causal effect between nodes is calculated according to the conditional entropy matrix. The causal effect between nodes is weighted and combined using the fault feature knowledge graph to obtain the causal strength between adjacent nodes. A dynamic causal strength threshold is set for the nodes in the fault feature knowledge graph. The associated neighbors of the nodes are filtered according to the dynamic causal strength threshold. An attention aggregation function is constructed using the feature representation of the associated neighbors and the causal strength. The node features are then fused at multiple levels through the attention aggregation function to obtain the multi-level node representation of the fault feature knowledge graph.

7. The method for diagnosing faults in power equipment according to claim 2, characterized in that, The process of using the fault feature propagation model to perform propagation link analysis on the optimized knowledge graph, mining the main paths of fault propagation based on the results of the propagation link analysis, and optimizing the topology of the fault feature knowledge graph according to the main paths to obtain the final knowledge graph includes: The fault feature propagation model is used to analyze the propagation link of the optimized knowledge graph. A multi-head attention mechanism is constructed to calculate the propagation probability between nodes. Based on different temporal features, the dynamic correlation between nodes is extracted according to each attention head of the multi-head attention mechanism. The link importance evaluation function is constructed according to the dynamic correlation and the propagation probability between nodes. The propagation link is preliminarily analyzed based on the output of the link importance evaluation function to obtain the initial propagation link. The dynamic correlation degree extracted by the multi-head attention mechanism is used as the weight of the node sequence. The conditional causal entropy of the initial propagation link is calculated based on the weight. The conditional causal entropy is evaluated a second time according to the link importance evaluation function. The output of the link importance evaluation function is compared with an adaptive threshold to select the main path of fault propagation. The node topology of the main path and the dynamic correlation are input into the link importance evaluation function. The multi-head attention mechanism is used to dynamically optimize the edge connection relationship of the fault feature knowledge graph. The structure of the fault feature knowledge graph is iteratively optimized to obtain the final knowledge graph.

8. The method for diagnosing faults in power equipment according to claim 7, characterized in that, The process of performing a secondary evaluation of the conditional causal entropy based on the link importance evaluation function, and comparing the output of the link importance evaluation function with an adaptive threshold to filter out the main paths of fault propagation, includes: The conditional causal entropy is evaluated a second time according to the link importance evaluation function. The link importance evaluation function uses a gated loop unit to extract the temporal features of the conditional causal entropy. The gated loop unit extracts the long and short-term temporal features of the conditional causal entropy through a reset gate and an update gate. An adaptive threshold is constructed based on the hidden state of the gated loop unit. The long and short-term time series features are compared with the adaptive threshold to filter out the main path of fault propagation.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method of any one of claims 1 to 8.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method of any one of claims 1 to 8.

Images