Power equipment fault prediction and diagnosis system based on artificial intelligence algorithm

Abstract

The application relates to the technical field of artificial intelligence. By providing an electric power equipment fault prediction and diagnosis system based on an artificial intelligence algorithm, the following are achieved: multi-modal data acquisition and preprocessing of vibration signals, temperature signals, electromagnetic field signals and partial discharge signals of electric power equipment to generate preprocessed multi-modal feature vectors; dynamic time benchmark axis construction processing of the preprocessed multi-modal feature vectors to generate time-calibrated multi-modal data; semantic enhancement feature fusion processing according to the multi-modal data to generate global semantic consistency features; fault prediction and diagnosis model training processing to generate fault type classification results and root cause tracing results; visual interactive interface generation processing based on the fault type classification results and the root cause tracing results to output a decision view including a fault probability heat zone distribution map and an economic treatment strategy, so as to solve the problems of multi-modal data time sequence misalignment and missing physical correlation in the related art.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms. Background Technology

[0002] With the improvement of the intelligence level of power equipment, fault prediction and diagnosis technology for key equipment such as transformers, circuit breakers, and cables has become a core link in ensuring the safe operation of the power grid.

[0003] However, due to the lack of a dynamic time reference axis for multimodal data, the related technologies have resulted in a misalignment between the causal events and the evolution of physical states. At the same time, cross-modal feature fusion ignores the embedding of physical mechanisms such as heat conduction equations and electromagnetic field coupling, which leads to poor interpretability of fault prediction models and increases the rate of missed detection of early symptoms. Summary of the Invention

[0004] Therefore, it is necessary to provide a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms to address the above-mentioned technical problems, so as to solve the problems of temporal misalignment and lack of physical correlation of multimodal data in related technologies, improve the time calibration accuracy of fault prediction models, enhance the physical interpretability of cross-modal features and reduce the early symptoms missed rate.

[0005] This application provides a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms, the system comprising:

[0006] The multimodal acquisition and preprocessing module is used to acquire and preprocess multimodal data of vibration signals, temperature signals, electromagnetic field signals and partial discharge signals of power equipment, and generate preprocessed multimodal feature vectors.

[0007] The dynamic time calibration module is used to construct a dynamic time reference axis for the preprocessed multimodal feature vectors based on the device physical model and causal event spectrum, and generate time-calibrated multimodal data.

[0008] The semantic feature fusion module is used to perform semantic enhancement feature fusion processing on time-calibrated multimodal data according to physical association rules to generate global semantic consistency features;

[0009] The fault prediction and diagnosis module is used to train fault prediction and diagnosis models based on global semantic consistency features, and generate fault type classification results and root cause tracing results.

[0010] The visualization decision generation module is used to generate a visualization interface based on the failure type classification results and root cause tracing results, and outputs a decision view including a failure probability hot zone distribution map and economical handling strategies.

[0011] Furthermore, based on the failure type classification results and root cause tracing results, a visual interactive interface is generated, outputting a decision view including a failure probability heatmap and economical handling strategies, including:

[0012] Using the following formula, based on the fault type classification results and root cause tracing results, a three-dimensional spatial mapping of the equipment is performed to generate a fault probability heat map:

[0013]

[0014] Among them, G i (x,y,z) represents the three-dimensional Gaussian distribution value of the i-th fault sample at the spatial point (x,y,z), where X represents the spatial coordinate vector [x,y,z], and μ i Σ represents the spatial location vector of the i-th fault sample. i Let |Σ| represent the covariance matrix of the i-th fault sample. i | represents the determinant of the covariance matrix;

[0015] Based on the fault probability hot zone distribution map, risk level labeling is performed through the fault propagation path to generate a color gradient visualization view.

[0016] Based on a color gradient visualization view, and combined with real-time electricity price data from the electricity market, an economic optimization process is performed to generate a decision view.

[0017] Furthermore, based on the fault probability hotspot distribution map, risk level labeling is performed through the fault propagation path to generate a color gradient visualization view, including:

[0018] Based on the fault probability hot zone distribution map, a fault propagation path topology map is generated through dynamic tracking and processing of fault propagation paths.

[0019] Using the following formula, based on the fault propagation path topology graph, node risk weight allocation is performed to generate a risk level weight distribution table:

[0020]

[0021] RW i =α×R i +β×CI i +γ×S i

[0022] Among them, R i Let ω represent the risk weight of node i, n represent the total number of nodes connected to node i, and ω represent the risk weight of node i. j P represents the inherent risk weight of node j. ij CI represents the probability of fault propagation from node j to node i. iLet d represent the centrality index of node i, m represent the total number of nodes in the topology graph, and d represent the centrality index of node i. ij f represents the shortest path distance from node i to node j. j RW represents the failure frequency of node j. i Let S represent the overall risk weight of node i, α represent the weight coefficient of the risk weight, β represent the weight coefficient of the centrality index, and γ represent the weight coefficient of the node importance, and α + β + γ = 1. i This represents the importance score of node i;

[0023] Based on the risk level weight distribution table, gradient mapping is performed using color coding rules to generate a color gradient visualization view.

[0024] Furthermore, based on the color gradient visualization view, and combined with real-time electricity market price data, economic optimization is performed to generate a decision view, including:

[0025] Based on the color gradient visualization view, risk nodes are prioritized and a list of key risk nodes is generated.

[0026] Based on the list of key risk nodes, and combined with real-time electricity price data from the electricity market, a time-period price correlation processing is performed to generate a cost impact assessment matrix.

[0027] Based on the cost impact assessment matrix, a multi-objective optimization algorithm is used to perform strategy trade-offs and generate a decision view.

[0028] Furthermore, based on the fault type classification results and root cause tracing results, a three-dimensional spatial mapping process is performed on the equipment to generate a fault probability heat map, including:

[0029] Based on the fault type classification results and root cause tracing results, a three-dimensional fault coordinate set is generated by performing spatial coordinate mapping processing through the three-dimensional physical model of the equipment.

[0030] Based on a three-dimensional fault coordinate set, a dynamic weight distribution matrix is ​​generated through dynamic weight allocation processing of fault types.

[0031] Based on the dynamic weight distribution matrix, a probability density calculation is performed using a hot zone aggregation algorithm to generate a fault probability hot zone distribution map.

[0032] Furthermore, semantic enhancement feature fusion processing is performed on the time-calibrated multimodal data according to physical association rules to generate global semantic consistency features, including:

[0033] Based on physical association rules, a multimodal physical knowledge graph is generated through the construction and processing of the device physical knowledge graph.

[0034] Based on time-calibrated multimodal data and multimodal physical knowledge graphs, graph structure transformation processing is performed to generate graph data including node features and physical association edges;

[0035] Based on graph data, a cross-modal attention mechanism is used to dynamically aggregate features and generate globally semantically consistent features.

[0036] Furthermore, based on the time-calibrated multimodal data and multimodal physical knowledge graph, graph structure transformation processing is performed to generate graph data including node features and physical association edges, including:

[0037] Based on time-calibrated multimodal data, node feature vectors are generated through physical attribute encoding.

[0038] Based on a multimodal physical knowledge graph, a physical association edge weight matrix is ​​generated through dynamic relationship modeling.

[0039] Based on node feature vectors and physical connection edge weight matrices, graph topology generation processing is performed to generate graph data containing node features and physical connection edges.

[0040] Furthermore, based on the multimodal physical knowledge graph, a physical association edge weight matrix is ​​generated through dynamic relationship modeling, including:

[0041] Based on a multimodal physical knowledge graph, a physical relationship subgraph associated with the operating status of equipment is generated through dynamic parsing and processing of physical relationships.

[0042] Based on the physical relationship subgraph, a dynamic weight calculation rule is used to perform real-time device status matching and generate a dynamic association weight parameter set.

[0043] Based on the dynamic association weight parameter set, the weight matrix is ​​reconstructed to generate the physical association edge weight matrix.

[0044] Furthermore, based on the device physical model and causal event graph, the preprocessed multimodal feature vectors are subjected to dynamic time reference axis construction to generate time-calibrated multimodal data, including:

[0045] Based on the physical model of the equipment, a theoretical time axis is generated by processing the physical state evolution equation.

[0046] Based on the causal event graph, causal event node identification processing is performed on the preprocessed multimodal feature vector to generate an event-driven temporal dependency graph.

[0047] Based on the theoretical state evolution time reference axis and time series dependency graph, multimodal data calibration is performed through dynamic timestamp remapping rules to generate time-calibrated multimodal data.

[0048] Furthermore, fault prediction and diagnosis model training is performed based on global semantic consistency features to generate fault type classification results and root cause tracing results, including:

[0049] Based on global semantic consistency features, a hybrid training model is constructed to generate supervised and unsupervised learning branches.

[0050] Fault type classification is performed through a supervised learning branch, generating classification results including over-temperature and insulation breakdown.

[0051] Anomaly pattern detection is performed through an unsupervised learning branch to generate anomalous features that deviate from the normal baseline.

[0052] Based on the classification results and abnormal characteristics, the fault propagation path is inverted through a graph-structured root cause tracing network to generate fault type classification results and root cause tracing results.

[0053] The technical solution provided in this application includes the following technical effects: By providing a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms, the system includes: a multimodal acquisition and preprocessing module, used to acquire and preprocess multimodal data of vibration signals, temperature signals, electromagnetic field signals and partial discharge signals of power equipment, and generate preprocessed multimodal feature vectors; a dynamic time calibration module, used to construct a dynamic time reference axis for the preprocessed multimodal feature vectors based on the equipment physical model and causal event spectrum, and generate time-calibrated multimodal data; and a semantic feature fusion module, used to process the time-calibrated multimodal data according to physical association rules. The system performs semantic enhancement feature fusion processing to generate global semantic consistency features; the fault prediction and diagnosis module is used to train fault prediction and diagnosis models based on global semantic consistency features, generating fault type classification results and root cause tracing results; the visualization decision generation module is used to generate a visualization interactive interface based on fault type classification results and root cause tracing results, outputting a decision view including fault probability heat map and economic disposal strategy, in order to solve the problems of temporal misalignment and lack of physical correlation of multimodal data in related technologies, thereby improving the time calibration accuracy of fault prediction models, enhancing the physical interpretability of cross-modal features, and reducing the early symptom missed rate. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a structural diagram of a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms in one embodiment of the present invention;

[0056] Figure 2 This is a flowchart of a decision view generated by combining a color gradient-based visualization view with real-time electricity market price data for economic optimization in one embodiment of the present invention.

[0057] Figure 3 This is a flowchart illustrating how, in one embodiment of the present invention, a dynamic time reference axis is constructed from preprocessed multimodal feature vectors based on a device physical model and a causal event graph to generate time-calibrated multimodal data. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific implementation methods of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0059] like Figure 1 As shown, this application provides a power equipment fault prediction and diagnosis system 100 based on artificial intelligence algorithms. The system 100 includes:

[0060] The multimodal acquisition and preprocessing module 101 is used to acquire and preprocess multimodal data of vibration signals, temperature signals, electromagnetic field signals and partial discharge signals of power equipment, and generate preprocessed multimodal feature vectors.

[0061] Specifically, data are collected from vibration, temperature, electromagnetic field, and partial discharge signals of the power equipment. A sensor network is used to acquire raw data from different signal sources, ensuring data integrity and accuracy. Subsequently, the collected raw data undergoes preliminary processing, including filtering, noise reduction, and normalization, to improve data quality and usability.

[0062] Subsequently, feature extraction is performed on different types of signals. Frequency domain features are extracted for vibration signals, time-varying features for temperature signals, amplitude and phase features for electromagnetic field signals, and discharge intensity and frequency features for partial discharge signals. The extracted signal features are then fused and combined into a multimodal feature vector according to certain rules and formats, providing fundamental data support for subsequent fault prediction and diagnosis.

[0063] The dynamic time calibration module 102 is used to construct a dynamic time reference axis for the preprocessed multimodal feature vector based on the device physical model and causal event spectrum, and generate time-calibrated multimodal data.

[0064] Specifically, based on the physical model of the device, the time reference axis of its theoretical state evolution is determined; then, the causal event graph is used to identify the causal event nodes in the multimodal feature vectors and generate a time-series dependency graph; then, through the dynamic timestamp remapping rule, the preprocessed multimodal feature vectors are mapped onto the theoretical state evolution time reference axis to achieve time calibration of multimodal data.

[0065] The semantic feature fusion module 103 is used to perform semantic enhancement feature fusion processing on the time-calibrated multimodal data according to the physical association rules to generate global semantic consistency features.

[0066] Specifically, based on the physical characteristics and operating mechanisms of the equipment, a corresponding physical knowledge graph is constructed to clarify the association rules between various physical quantities. Then, time-calibrated multimodal data is mapped onto this physical knowledge graph. Through graph structure transformation, the data from different modalities is converted into graph data containing node features and physical association edges. Node features represent the key characteristics of each modal data, while physical association edges reflect the inherent physical connections between them.

[0067] Subsequently, a cross-modal attention mechanism is used to dynamically aggregate features of the graph data. In this process, features of different modalities are assigned different weights and fused according to their importance and relevance in physical association, thereby generating features that can comprehensively reflect the equipment status and have global semantic consistency, i.e. global semantic consistency features, which provide a more physically interpretable and comprehensive feature foundation for subsequent fault prediction and diagnosis.

[0068] The fault prediction and diagnosis module 104 is used to train and process fault prediction and diagnosis models based on global semantic consistency features, and generate fault type classification results and root cause tracing results.

[0069] Specifically, a hybrid training model is constructed, comprising a supervised learning branch and an unsupervised learning branch. In the supervised learning branch, the model is trained using labeled global semantic consistency feature samples. Through feature-fault type association learning, a fault type classification model is generated, achieving more accurate classification of fault types such as over-temperature and insulation breakdown. In the unsupervised learning branch, the model is trained using unlabeled global semantic consistency feature samples. Through cluster analysis and pattern recognition, abnormal patterns in equipment operation are detected, identifying features deviating from the normal baseline.

[0070] Subsequently, by combining the outputs of supervised and unsupervised learning, a graph-structured root cause tracing network is used to invert the fault propagation path. By analyzing the correlation between fault types and abnormal patterns, fault type classification results and root cause tracing results are generated, thereby achieving a more comprehensive prediction and diagnosis of faults.

[0071] The visualization decision generation module 105 is used to generate a visualization interactive interface based on the failure type classification results and root cause tracing results, and outputs a decision view including a failure probability hot zone distribution map and an economical handling strategy.

[0072] Specifically, based on the fault type classification results and root cause tracing results, a 3D fault coordinate set containing fault location information is generated by spatial coordinate mapping processing through the equipment's 3D physical model. Then, according to the fault type and root cause tracing information, the 3D fault coordinate set is weighted to generate a dynamic weight distribution matrix, which reflects the probability and importance of fault occurrence at different locations. Finally, a hot zone aggregation algorithm is used to calculate the probability density of the dynamic weight distribution matrix, generating a fault probability hot zone distribution map, which visually displays the fault risk level of each area of ​​the equipment.

[0073] Simultaneously, by combining real-time electricity price data and fault information from the power market, an economic assessment is conducted to analyze the costs and benefits under different handling strategies, generating a decision view for economical handling strategies. The fault probability hotspot distribution map and the economical handling strategy decision view are integrated into a visual interactive interface. This interface is used for processing and outputting a decision view including the fault probability hotspot distribution map and economical handling strategies, providing users with information for fault handling decisions.

[0074] One embodiment of this application provides a power equipment fault prediction and diagnosis system based on artificial intelligence algorithms. The system includes: a multimodal acquisition and preprocessing module for acquiring and preprocessing multimodal data from vibration signals, temperature signals, electromagnetic field signals, and partial discharge signals of the power equipment to generate preprocessed multimodal feature vectors; a dynamic time calibration module for constructing a dynamic time reference axis on the preprocessed multimodal feature vectors based on the equipment's physical model and causal event graph to generate time-calibrated multimodal data; and a semantic feature fusion module for semantic enhancement of the time-calibrated multimodal data according to physical association rules. Feature fusion processing generates globally semantically consistent features; the fault prediction and diagnosis module is used to train fault prediction and diagnosis models based on globally semantically consistent features, generating fault type classification results and root cause tracing results; the visualization decision generation module is used to generate a visualization interactive interface based on fault type classification results and root cause tracing results, outputting a decision view including a fault probability heat map and economic handling strategies, in order to solve the problems of temporal misalignment and lack of physical correlation of multimodal data in related technologies, thereby improving the time calibration accuracy of fault prediction models, enhancing the physical interpretability of cross-modal features, and reducing the early symptom missed rate.

[0075] Furthermore, based on the failure type classification results and root cause tracing results, a visual interactive interface is generated, outputting a decision view including a failure probability heatmap and economical handling strategies, including:

[0076] Using the following formula, based on the fault type classification results and root cause tracing results, a three-dimensional spatial mapping of the equipment is performed to generate a fault probability heat map:

[0077]

[0078] Among them, G i (x,y,z) represents the three-dimensional Gaussian distribution value of the i-th fault sample at the spatial point (x,y,z), where X represents the spatial coordinate vector [x,y,z], and μ i Σ represents the spatial location vector of the i-th fault sample. i Let |Σ| represent the covariance matrix of the i-th fault sample. i | represents the determinant of the covariance matrix;

[0079] Based on the fault probability hot zone distribution map, risk level labeling is performed through the fault propagation path to generate a color gradient visualization view.

[0080] Based on a color gradient visualization view, and combined with real-time electricity price data from the electricity market, an economic optimization process is performed to generate a decision view.

[0081] Specifically, based on the fault type classification results and root cause tracing results, a spatial coordinate mapping process is performed using the equipment's 3D physical model to generate a 3D fault coordinate set including fault location information. Then, based on the fault type and root cause tracing information, a dynamic weight allocation process is applied to the 3D fault coordinate set to generate a dynamic weight distribution matrix. This matrix reflects the probability and importance of fault occurrence at different locations. Finally, a hotspot aggregation algorithm is used to calculate the probability density of the dynamic weight distribution matrix, generating a fault probability hotspot distribution map that visually displays the fault risk level in each area of ​​the equipment.

[0082] After generating the fault probability hotspot distribution map, the map is annotated with risk levels by analyzing fault propagation paths. Color gradient technology is then used to visually represent different risk levels through color changes, generating a color gradient visualization view. Subsequently, combined with real-time electricity market price data, the color gradient visualization view undergoes economic optimization, evaluating the costs and benefits of different handling strategies. This results in a decision view that includes the fault probability hotspot distribution map and economic handling strategies, presented to users through a visual interactive interface to assist them in making fault handling decisions.

[0083] Furthermore, based on the fault probability hotspot distribution map, risk level labeling is performed through the fault propagation path to generate a color gradient visualization view, including:

[0084] Based on the fault probability hot zone distribution map, a fault propagation path topology map is generated through dynamic tracking and processing of fault propagation paths.

[0085] Using the following formula, based on the fault propagation path topology graph, node risk weight allocation is performed to generate a risk level weight distribution table:

[0086]

[0087] RW i =α×R i +β×CI i +γ×S i

[0088] Among them, R i Let ω represent the risk weight of node i, n represent the total number of nodes connected to node i, and ω represent the risk weight of node i. j P represents the inherent risk weight of node j. ij CI represents the probability of fault propagation from node j to node i. i Let d represent the centrality index of node i, m represent the total number of nodes in the topology graph, and d represent the centrality index of node i. ij f represents the shortest path distance from node i to node j. j RW represents the failure frequency of node j. iLet S represent the overall risk weight of node i, α represent the weight coefficient of the risk weight, β represent the weight coefficient of the centrality index, and γ represent the weight coefficient of the node importance, and α + β + γ = 1. i This represents the importance score of node i;

[0089] Based on the risk level weight distribution table, gradient mapping is performed using color coding rules to generate a color gradient visualization view.

[0090] Specifically, using a fault probability heatmap as a foundation, combined with the equipment's physical architecture and operating characteristics, fault propagation paths are dynamically tracked. This step aims to identify and depict possible propagation routes of faults within the equipment or system, thereby generating a fault propagation path topology map. This topology map not only shows the possible paths of fault propagation but also reflects the connections between nodes and the direction of propagation.

[0091] Next, node risk weight allocation is performed. This process comprehensively considers multiple risk factors, such as the inherent risk attributes of nodes, the probability of fault propagation between nodes, the centrality of nodes in the network, the shortest path distance between nodes, and the frequency of node failures. Through systematic analysis of these factors, the risk weight of each node is calculated, thereby generating a risk level weight distribution table. This distribution table assigns a risk weight value to each node, reflecting the potential risk level of the node in fault propagation.

[0092] Next, based on the risk level weight distribution table, gradient mapping is performed using color coding rules. This includes assigning corresponding colors to different risk levels according to their risk weights, generating a color gradient. Through this visualization process, a color gradient visualization view is generated, presenting the fault risk level in an intuitive color-changing format, facilitating users to quickly identify high-risk areas and critical paths of fault propagation. This view not only improves the intuitiveness of fault diagnosis but also provides important visual references for subsequent maintenance decisions.

[0093] like Figure 2 As shown, based on a color gradient visualization view, and combined with real-time electricity market price data, economic optimization is performed to generate a decision view, including:

[0094] S201: Based on the color gradient visualization view, perform risk node priority sorting and generate a list of key risk nodes;

[0095] S202: Based on the list of key risk nodes, and combined with real-time electricity price data from the electricity market, time-period price correlation processing is performed to generate a cost impact assessment matrix;

[0096] S203: Based on the cost impact assessment matrix, a multi-objective optimization algorithm is used to perform strategy trade-offs and generate a decision view.

[0097] Specifically, in step S201, the color gradient visualization view is analyzed to identify the risk levels represented by different color gradients, and each risk node is prioritized according to the risk levels to generate a list of key risk nodes, thus clarifying the key objects for subsequent economic analysis.

[0098] Next, in step S202, the obtained list of key risk nodes is combined with real-time electricity price data from the electricity market to perform time-of-use price correlation processing. This process involves considering electricity price fluctuations over different time periods and analyzing the impact of price changes on the maintenance or repair costs of each key risk node. A cost impact assessment matrix is ​​then generated, showing the cost changes resulting from handling faults at different risk nodes during different time periods.

[0099] Then, in step S203, based on the cost impact assessment matrix, a multi-objective optimization algorithm is used to weigh various possible fault handling strategies. The multi-objective optimization algorithm comprehensively considers multiple objectives such as minimizing costs and maximizing fault handling effectiveness, seeking a balance among numerous strategies to generate a decision view. This view clearly presents the advantages and disadvantages of each strategy under different constraints, providing decision-makers with a scientific and reasonable basis for decision-making, and helping them to formulate economical, effective, and reliable power equipment fault handling solutions.

[0100] Furthermore, based on the fault type classification results and root cause tracing results, a three-dimensional spatial mapping process is performed on the equipment to generate a fault probability heat map, including:

[0101] Based on the fault type classification results and root cause tracing results, a three-dimensional fault coordinate set is generated by performing spatial coordinate mapping processing through the three-dimensional physical model of the equipment.

[0102] Based on a three-dimensional fault coordinate set, a dynamic weight distribution matrix is ​​generated through dynamic weight allocation processing of fault types.

[0103] Based on the dynamic weight distribution matrix, a probability density calculation is performed using a hot zone aggregation algorithm to generate a fault probability hot zone distribution map.

[0104] Specifically, the fault type classification results are combined with the root cause tracing results, and the fault information is mapped to spatial coordinates using the equipment's three-dimensional physical model. This step correlates the abstract fault data with the actual physical location of the equipment, generating a three-dimensional fault coordinate set containing fault location information, laying the foundation for subsequent spatial analysis.

[0105] Subsequently, based on the fault type and detailed root cause tracing information, dynamic weights are assigned to each coordinate point in the three-dimensional fault coordinate set. This process considers factors such as fault type severity, frequency of occurrence, and the influence of the root cause. High-risk areas are highlighted through weight allocation, generating a dynamic weight distribution matrix that quantitatively describes the degree of fault risk at different locations on the equipment.

[0106] Next, a hot zone aggregation algorithm is used to process the dynamic weight distribution matrix. This algorithm analyzes the weight distribution to calculate the fault probability density of each area of ​​the equipment, integrating the scattered fault data into intuitive visual information to generate a fault probability hot zone distribution map. This distribution map visually displays the fault probability of different parts of the equipment using variations in color depth or brightness, helping maintenance personnel to quickly locate high-risk areas, thereby achieving visualized management of power equipment fault risks.

[0107] Furthermore, semantic enhancement feature fusion processing is performed on the time-calibrated multimodal data according to physical association rules to generate global semantic consistency features, including:

[0108] Based on physical association rules, a multimodal physical knowledge graph is generated through the construction and processing of the device physical knowledge graph.

[0109] Based on time-calibrated multimodal data and multimodal physical knowledge graphs, graph structure transformation processing is performed to generate graph data including node features and physical association edges;

[0110] Based on graph data, a cross-modal attention mechanism is used to dynamically aggregate features and generate globally semantically consistent features.

[0111] Specifically, a physical knowledge graph of the devices is constructed based on physical association rules. This step generates a multimodal physical knowledge graph by integrating the physical characteristics and operating mechanisms of the devices, providing a structured knowledge framework for subsequent data fusion.

[0112] Next, the time-calibrated multimodal data is mapped onto a multimodal physical knowledge graph, and graph structure transformation is performed. The above steps transform data from different modalities into graph data, where node features represent the key characteristics of each modality, and physical association edges reflect the inherent physical connections between them.

[0113] Subsequently, a cross-modal attention mechanism is used to dynamically aggregate features from the graph data. This mechanism dynamically adjusts feature weights and fuses them based on the importance and relevance of each modal feature in the physical association, thereby generating a globally semantically consistent feature. This feature comprehensively reflects the device status, exhibiting stronger physical interpretability and consistency, thus providing a foundation for subsequent fault prediction and diagnosis.

[0114] Furthermore, based on the time-calibrated multimodal data and multimodal physical knowledge graph, graph structure transformation processing is performed to generate graph data including node features and physical association edges, including:

[0115] Based on time-calibrated multimodal data, node feature vectors are generated through physical attribute encoding.

[0116] Based on a multimodal physical knowledge graph, a physical association edge weight matrix is ​​generated through dynamic relationship modeling.

[0117] Based on node feature vectors and physical connection edge weight matrices, graph topology generation processing is performed to generate graph data containing node features and physical connection edges.

[0118] Specifically, physical attribute encoding is performed on the time-calibrated multimodal data. This process involves encoding the physical attributes of different modal data, mapping physical quantities such as vibration, temperature, and electromagnetic fields into feature vectors, thereby generating node feature vectors that can characterize the key features of each modal data, laying the foundation for subsequent graph data construction.

[0119] Next, dynamic relationship modeling is carried out based on the multimodal physical knowledge graph. This step focuses on constructing a physical association edge weight matrix. Through dynamic analysis of each physical entity and relationship in the knowledge graph, the association strength between different physical quantities is quantified, and a physical association edge weight matrix is ​​generated. This matrix intuitively reflects the correlation and influence between each physical quantity.

[0120] Next, the graph topology is generated by combining the node feature vectors and the physical association edge weight matrix. By integrating the node feature vectors and the physical association edge weight matrix, a complete graph data structure is constructed, where nodes represent key features of multimodal data and edges represent the relationships and weights between physical quantities. This generates graph data containing node features and physical association edges, providing structured input for subsequent graph neural network analysis or other graph data-based processing.

[0121] Furthermore, based on the multimodal physical knowledge graph, a physical association edge weight matrix is ​​generated through dynamic relationship modeling, including:

[0122] Based on a multimodal physical knowledge graph, a physical relationship subgraph associated with the operating status of equipment is generated through dynamic parsing and processing of physical relationships.

[0123] Based on the physical relationship subgraph, a dynamic weight calculation rule is used to perform real-time device status matching and generate a dynamic association weight parameter set.

[0124] Based on the dynamic association weight parameter set, the weight matrix is ​​reconstructed to generate the physical association edge weight matrix.

[0125] Specifically, through dynamic parsing of physical relationships, physical relationships in the multimodal physical knowledge graph are analyzed and extracted in real time. This process aims to identify and construct physical relationship subgraphs closely related to the current operating state of the equipment, which include the correlation between different physical quantities, such as the interaction between vibration and temperature, electromagnetic field and partial discharge.

[0126] Next, dynamic weight calculation rules are used to perform real-time device state matching processing on the physical relationships in the physical relationship subgraph. This step dynamically calculates the weight parameters of each physical relationship by analyzing the changes in physical quantities under the current operating state of the equipment, generating a dynamic association weight parameter set. The dynamic weight calculation rules take into account the changing trends of physical quantities, historical data, and contextual information of the equipment's operating state, thereby quantifying the importance of each physical relationship.

[0127] Next, based on the dynamic association weight parameter set, a weight matrix reconstruction process is performed to generate a physical association edge weight matrix. This step integrates the weight parameters from the dynamic association weight parameter set into a single matrix structure to generate the physical association edge weight matrix. Each element in this matrix represents the association strength between different physical quantities, providing a foundation for subsequent graph data processing and analysis, and also providing a quantitative basis for physical associations in cross-modal feature fusion.

[0128] like Figure 3 As shown, based on the device physical model and causal event graph, the preprocessed multimodal feature vectors are dynamically constructed using a time reference axis to generate time-calibrated multimodal data, including:

[0129] S301: Based on the equipment physical model, the theoretical time axis is generated through the physical state evolution equation to generate the theoretical state evolution time reference axis;

[0130] S302: Based on the causal event graph, perform causal event node identification processing on the preprocessed multimodal feature vector to generate an event-driven temporal dependency graph;

[0131] S303: Based on the theoretical state evolution time reference axis and time series dependency map, multimodal data calibration is performed through dynamic timestamp remapping rules to generate time-calibrated multimodal data.

[0132] Specifically, in step S301, based on the physical model of the equipment, the theoretical operating state of the equipment is simulated and predicted using the physical state evolution equation, and a theoretical state evolution time reference axis is generated. This reference axis provides a reference for the physical state of the equipment changing over time under ideal conditions.

[0133] Subsequently, in step S302, based on the causal event graph, a deep analysis is performed on the preprocessed multimodal feature vectors to identify the causal event nodes contained therein, and an event-driven temporal dependency graph is constructed. This graph depicts the sequence and causal relationship between different events, reflecting the time sequence relationship of key events during equipment operation.

[0134] Subsequently, in step S303, the multimodal data is calibrated using a dynamic timestamp remapping rule, combining the theoretical state evolution time reference axis and the time-series dependency graph. By comparing the time difference between the actual data and the theoretical reference, the timestamps of the multimodal feature vectors are adjusted to achieve more precise alignment of the data in different modalities in the time dimension. This ensures that each modal data can more accurately reflect the true state of the equipment at the same point in time, thereby generating time-calibrated multimodal data, providing time-consistent data support for subsequent fault prediction and diagnosis.

[0135] Furthermore, fault prediction and diagnosis model training is performed based on global semantic consistency features to generate fault type classification results and root cause tracing results, including:

[0136] Based on global semantic consistency features, a hybrid training model is constructed to generate supervised and unsupervised learning branches.

[0137] Fault type classification is performed through a supervised learning branch, generating classification results including over-temperature and insulation breakdown.

[0138] Anomaly pattern detection is performed through an unsupervised learning branch to generate anomalous features that deviate from the normal baseline.

[0139] Based on the classification results and abnormal characteristics, the fault propagation path is inverted through a graph-structured root cause tracing network to generate fault type classification results and root cause tracing results.

[0140] Specifically, a hybrid training model is constructed using global semantic consistency features. This step involves building two branches: a supervised learning branch and an unsupervised learning branch. The supervised learning branch focuses on learning from labeled data, enabling the identification and classification of specific fault types, such as overheating and insulation breakdown; the unsupervised learning branch processes unlabeled data, aiming to discover hidden patterns and anomalous features in the data, and identify abnormal situations that deviate from normal operating conditions.

[0141] Next, a supervised learning branch is used to classify the fault types. This branch is trained using labeled data of known fault types to learn the feature patterns of different fault types, generate corresponding classification results, and clearly indicate the specific fault types that the equipment may encounter.

[0142] Simultaneously, the unsupervised learning branch performs anomaly pattern detection. Without predefined fault labels, this branch analyzes the inherent structure and distribution of the data to identify anomalous features that significantly differ from normal operating conditions, thereby discovering potential unknown fault modes.

[0143] Subsequently, combining the classification results of the supervised learning branch and the abnormal features of the unsupervised learning branch, a graph-structured root cause tracing network is used to invert the fault propagation path. This network analyzes the causal relationships and propagation paths between fault types and abnormal features, tracing back to the root cause of the fault and generating comprehensive fault type classification and root cause tracing results, providing more accurate guidance for equipment maintenance and fault repair.

[0144] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0145] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0146] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A power equipment fault prediction and diagnosis system based on artificial intelligence algorithms, characterized in that, The system includes: The multimodal acquisition and preprocessing module is used to acquire and preprocess multimodal data of vibration signals, temperature signals, electromagnetic field signals and partial discharge signals of power equipment, and generate preprocessed multimodal feature vectors. The dynamic time calibration module is used to construct a dynamic time reference axis for the preprocessed multimodal feature vector based on the device physical model and causal event spectrum, and generate time-calibrated multimodal data. The semantic feature fusion module is used to perform semantic enhancement feature fusion processing on the time-calibrated multimodal data according to the physical association rules to generate global semantic consistency features; The fault prediction and diagnosis module is used to train the fault prediction and diagnosis model based on the global semantic consistency features, and generate fault type classification results and root cause tracing results. The visualization decision generation module is used to generate a visualization interactive interface based on the fault type classification results and root cause tracing results, and outputs a decision view including a fault probability hot zone distribution map and an economical handling strategy. The process of generating a visual interactive interface based on the fault type classification results and root cause tracing results outputs a decision view including a fault probability heatmap and economical handling strategies, including: Using the following formula, based on the fault type classification results and root cause tracing results, a three-dimensional spatial mapping process is performed on the equipment to generate a fault probability heat map: Among them, G i (x,y,z) represents the three-dimensional Gaussian distribution value of the i-th fault sample at the spatial point (x,y,z), where X represents the spatial coordinate vector [x,y,z], and μ i Σ represents the spatial location vector of the i-th fault sample. i Let |Σ| represent the covariance matrix of the i-th fault sample. i | represents the determinant of the covariance matrix; Based on the fault probability hot zone distribution map, risk level labeling is performed through the fault propagation path to generate a color gradient visualization view. Based on the color gradient visualization view, economic optimization is performed by combining real-time electricity price data from the electricity market to generate the decision view.

2. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithms according to claim 1, characterized in that, The step of generating a color gradient visualization view by labeling risk levels based on the fault probability hotspot distribution map and fault propagation paths includes: Based on the aforementioned fault probability hotspot distribution map, a fault propagation path topology map is generated through dynamic tracking of fault propagation paths. Using the following formula, based on the fault propagation path topology graph, node risk weight allocation is performed to generate a risk level weight distribution table: RW i =α×R i +β×CI i +γ×S i Among them, R i Let ω represent the risk weight of node i, n represent the total number of nodes connected to node i, and ω represent the risk weight of node i. j P represents the inherent risk weight of node j. ij CI represents the probability of fault propagation from node j to node i. i Let d represent the centrality index of node i, m represent the total number of nodes in the topology graph, and d represent the centrality index of node i. ij f represents the shortest path distance from node i to node j. j RW represents the failure frequency of node j. i Let S represent the overall risk weight of node i, α represent the weight coefficient of the risk weight, β represent the weight coefficient of the centrality index, and γ represent the weight coefficient of the node importance, and α + β + γ = 1. i This represents the importance score of node i; Based on the risk level weight distribution table, gradient mapping is performed using color encoding rules to generate the color gradient visualization view.

3. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithms according to claim 1, characterized in that, The process of generating the decision view by combining the color gradient visualization view with real-time electricity market price data for economic optimization includes: Based on the color gradient visualization view, risk nodes are prioritized and sorted to generate a list of key risk nodes. Based on the list of key risk nodes, and combined with real-time electricity price data from the electricity market, time-period price correlation processing is performed to generate a cost impact assessment matrix. Based on the cost impact assessment matrix, a multi-objective optimization algorithm is used to perform strategy trade-offs and generate the decision view.

4. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithm according to claim 1, characterized in that, The step of performing three-dimensional spatial mapping processing on the equipment based on the fault type classification results and root cause tracing results to generate a fault probability heat map includes: Based on the fault type classification results and root cause tracing results, a three-dimensional fault coordinate set is generated by performing spatial coordinate mapping processing through the three-dimensional physical model of the equipment. Based on the three-dimensional fault coordinate set, a dynamic weight distribution matrix is ​​generated through dynamic weight allocation processing of fault types. Based on the dynamic weight distribution matrix, the probability density is calculated using a hot zone aggregation algorithm to generate the fault probability hot zone distribution map.

5. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithms according to claim 1, characterized in that, The step of performing semantic enhancement feature fusion processing on the time-calibrated multimodal data according to physical association rules to generate global semantic consistency features includes: Based on the physical association rules, a multimodal physical knowledge graph is generated through the construction and processing of the device physical knowledge graph. Based on the time-calibrated multimodal data and the multimodal physical knowledge graph, graph structure transformation processing is performed to generate graph data including node features and physical association edges; Based on the graph data, the global semantic consistency features are generated by performing dynamic feature aggregation through a cross-modal attention mechanism.

6. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithms according to claim 5, characterized in that, The multimodal data based on the time-calibrated data and the multimodal physical knowledge graph are subjected to graph structure transformation processing to generate graph data including node features and physical association edges, including: Based on the time-calibrated multimodal data, node feature vectors are generated through physical attribute encoding processing. Based on the multimodal physical knowledge graph, a physical association edge weight matrix is ​​generated through dynamic relationship modeling. Based on the node feature vectors and the weight matrix of the physical association edges, graph topology generation processing is performed to generate the graph data containing node features and physical association edges.

7. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithms according to claim 6, characterized in that, The process of generating a physical association edge weight matrix based on the multimodal physical knowledge graph through dynamic relationship modeling includes: Based on the multimodal physical knowledge graph, a physical relationship subgraph associated with the device's operating status is generated through dynamic parsing of physical relationships. Based on the physical relationship subgraph, real-time device status matching is performed through dynamic weight calculation rules to generate a dynamic association weight parameter set. Based on the dynamic association weight parameter set, a weight matrix reconstruction process is performed to generate the physical association edge weight matrix.

8. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithm according to claim 1, characterized in that, The process of constructing a dynamic time reference axis for the preprocessed multimodal feature vectors based on the device physical model and causal event graph to generate time-calibrated multimodal data includes: Based on the physical model of the device, a theoretical time axis is generated by processing the physical state evolution equation; a theoretical state evolution time reference axis is generated. Based on the causal event graph, the preprocessed multimodal feature vector is subjected to causal event node identification processing to generate an event-driven temporal dependency graph. Based on the theoretical state evolution time reference axis and the time-dependent graph, multimodal data calibration is performed using dynamic timestamp remapping rules to generate the time-calibrated multimodal data.

9. The power equipment fault prediction and diagnosis system based on artificial intelligence algorithm according to claim 1, characterized in that, The training process for the fault prediction and diagnosis model based on the global semantic consistency features generates fault type classification results and root cause tracing results, including: Based on the global semantic consistency feature, a hybrid training model is constructed to generate supervised learning branches and unsupervised learning branches. The supervised learning branch is used to classify fault types, generating classification results including over-temperature and insulation breakdown. The unsupervised learning branch is used to perform abnormal pattern detection processing to generate abnormal features that deviate from the normal state baseline; Based on the classification results and the abnormal features, the fault propagation path is inverted using a graph-structured root cause tracing network to generate the fault type classification results and root cause tracing results.

Images