Power failure intelligent monitoring method and system based on knowledge graph

Abstract

The invention discloses a power fault intelligent monitoring method and system based on a knowledge graph, and relates to the technical field of power fault intelligent monitoring, and the method comprises the steps: obtaining a fault neural network according to a historical fault library, constructing a fault knowledge graph based on the power equipment topology and the historical fault library, and generating a component-level association rule; and based on a result of fault neural network reasoning, outputting a probabilistic traceability report containing fault component positioning and an evolution path. According to the invention, intelligent monitoring and accurate positioning of power transformer faults are realized by combining multi-modal sensor data with a time mark alignment technology, a learnable gating attention mechanism and a knowledge graph. The fault picture set is expanded by using the conditional generative adversarial network, and the fault neural network is constructed based on the historical fault library, so that the scarce fault type can be effectively identified and the accuracy of fault diagnosis can be improved. Meanwhile, the timeliness and the reliability of fault detection are remarkably improved through a probabilistic traceability report, and the power failure time and the maintenance cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent power fault monitoring technology, specifically to an intelligent power fault monitoring method and system based on knowledge graphs. Background Technology

[0002] As a core component of the power grid system, the operational reliability of power transformers directly affects the safety and stability of the entire power system. A fault in a transformer can lead to anything from localized power outages to cascading failures causing widespread blackouts and significant economic losses.

[0003] Traditional monitoring methods widely used in the industry have significant technical limitations, resulting in severely delayed fault response, high false alarm rates, and the inability to detect sudden faults due to their limited availability. Regular offline monitoring projects, such as frequency response analysis, are typically conducted only once every 1-2 years. Existing online monitoring systems generally employ fixed threshold alarm mechanisms, leading to high false alarm rates under complex operating conditions and failing to deeply analyze the internal working state and potential fault modes of equipment. This is particularly true for critical equipment like power transformers, which have complex structures and a wide variety of fault types. Traditional monitoring methods often only provide surface information, making early warning and precise fault source location difficult. Furthermore, the scarcity of samples for certain fault types leads to insufficient training of deep learning models, affecting the comprehensiveness and robustness of fault identification. The spatiotemporal asynchrony of multimodal sensors makes data fusion difficult, limiting the accuracy of fault tracing. Therefore, this invention proposes a knowledge graph-based intelligent power fault monitoring method to address these problems and improve the intelligence level of power equipment fault monitoring. Summary of the Invention

[0004] To address the aforementioned technical challenges, this paper presents a knowledge graph-based intelligent monitoring method and system for power faults. This solution combines multimodal sensor data with time-stamp alignment, a learnable gated attention mechanism, and a knowledge graph to achieve intelligent monitoring and precise location of power transformer faults. By expanding the fault image set using conditional generative adversarial networks and constructing a fault neural network based on a historical fault database, it effectively identifies rare fault types and improves the accuracy of fault diagnosis. Furthermore, by evaluating the probability of fault occurrence through component-level association rules and decision tree algorithms, it outputs a probabilistic tracing report containing the location and evolution path of the faulty component, significantly improving the timeliness and reliability of fault detection, reducing power outage time and maintenance costs, and providing strong technical support for the stable operation of the power system.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A knowledge graph-based intelligent monitoring method for power faults includes:

[0007] Obtain power transformer structural information, which includes power equipment topology, historical fault database, and multimodal sensor data corresponding to the power transformer;

[0008] The multimodal sensor data is time-stamped using a high-precision time synchronization module;

[0009] Based on dual-stream feature data acquired by multimodal sensors, and using a learnable gating attention mechanism, the confidence levels of static and dynamic feature stream data are obtained.

[0010] Obtain the faulty neural network based on the historical fault database;

[0011] A fault knowledge graph is constructed based on the power equipment topology and historical fault database, and component-level association rules are generated.

[0012] Based on the confidence levels of multimodal sensor data, static feature stream data, and dynamic feature stream data, a decision tree is constructed using component-level association rules and cases from the historical fault database to obtain the probability of power transformer failure.

[0013] If the probability of a power transformer malfunctioning is greater than a preset threshold, the power transformer characteristic data information is acquired and uploaded to the fault neural network for inference. The power transformer characteristic data information includes the appearance information, humidity distribution information, and temperature distribution information of the power transformer components.

[0014] Based on the results of fault neural network inference, a probabilistic tracing report containing the location of faulty components and their evolution paths is output.

[0015] Preferably, obtaining the fault neural network based on the historical fault database specifically includes:

[0016] Obtain a set of fault images of power transformers;

[0017] Based on the fault images, obtain the number of samples of each fault type in the image dataset, and filter out the rare fault categories.

[0018] Based on the scarce fault categories, an expanded fault image set is obtained using a conditional generative adversarial network.

[0019] Based on the historical fault database, the images in the expanded fault image set are labeled, and the labels are used to delineate the fault areas and fault types of power transformers.

[0020] Based on the annotated and expanded fault image set, obtain the fault neural network.

[0021] Preferably, the step of obtaining an expanded fault image set based on a conditional generative adversarial network according to scarce fault categories specifically includes:

[0022] Based on a set of images with scarce fault categories, a synthetic image with a specified fault category is obtained using a generative adversarial network.

[0023] Based on image preprocessing, the synthesized image is standardized, including scaling the image region corresponding to the faulty component and normalizing the illumination.

[0024] Based on the image set of rare fault categories, and based on image preprocessing, the image region corresponding to the faulty component in the real fault image is obtained. The real fault image is an image of the faulty component in the historical fault database.

[0025] Based on the image regions of faulty components in real and synthetic images, obtain structural similarity index, distortion index, and spatial similarity.

[0026] The structural similarity index is specifically as follows:

[0027]

[0028] In the formula, SSIM is the structural similarity index between real fault images and synthetic images, μ x μ y These are the average values ​​of real fault images and synthetic images, respectively. σ represents the variance of the real fault image and the synthetic image, respectively. xy The covariance between real fault images and synthetic images;

[0029] The distortion index is as follows:

[0030]

[0031] In the formula, MSE is the transition value for calculating the distortion index, DDIN is the distortion index, x is the real faulty image, y is the composite image, L is the maximum pixel value, M is the length of the image, and N is the width of the image.

[0032] Spatial similarity is specifically defined as:

[0033] Sim = BERT(n x ,n y );

[0034] In the formula, Sim represents spatial similarity, BERT represents the cosine similarity function, and n... x ,n y These are the feature vectors extracted from real fault images and synthetic images by convolutional neural networks, respectively.

[0035] The accuracy of the synthesized image is obtained based on the structural similarity index, the distortion index, and the spatial similarity.

[0036] Specifically, the accuracy of the synthesized image is as follows:

[0037] Q=0.1*SSIM+0.1*DDIN+0.8*Sim;

[0038] In the formula, Q represents the accuracy of the synthesized image;

[0039] If the accuracy of the synthesized image is within a preset threshold range, then the synthesized image will be added to the expanded set of fault images.

[0040] Preferably, the result of the fault neural network inference outputs a probabilistic tracing report containing the location and evolution path of the faulty component, specifically including:

[0041] Based on the fault neural network, fault information of the power transformer is obtained, which includes the faulty component of the power transformer, the fault type, and the first probability of the faulty component failing.

[0042] Based on multimodal sensor data, the fault type of the power transformer is determined. The fault type includes independent faults and coupled faults. The coupled fault is a compound fault with a causal relationship caused by the interaction between two or more power transformer component faults.

[0043] If the fault type is a coupled fault, then based on the fault information of the power transformer and the actual fault image of the component, the second probability of the power transformer component fault is obtained. The second probability of the fault is the accuracy of the fault component area in the fault information of the power transformer and the fault component area in the actual fault image.

[0044] Based on the first probability and the second probability of failure, a probabilistic tracing report of component failure is obtained.

[0045] Preferably, determining the fault type of the power transformer based on multimodal sensor data specifically includes:

[0046] Based on multimodal sensor data, abnormal data is obtained, which is data that deviates from the normal operating range of the power transformer;

[0047] Based on component-level association rules, multi-dimensional auxiliary data is obtained, which is data that has a logical association with the abnormal data in the component-level association rules.

[0048] Based on the knowledge graph, a fault conflict matrix is ​​obtained, which is a three-dimensional matrix containing power transformer components, fault types, and fault characteristics.

[0049] Based on the fault conflict matrix, fault-related components are obtained. These fault-related components are those that may fail, which are found in the component-level association rules based on the abnormal data.

[0050] Based on the faulty component and its associated components, spatial overlap and temporal correlation are obtained using multimodal sensor data.

[0051] Specifically, the spatial overlap is as follows:

[0052]

[0053] In the formula, O s For spatial overlap, Area(Sensor) A Area(Sensor) B These are the sensor monitoring areas for the faulty component and the related faulty component, respectively.

[0054] The specific time-series correlation is as follows:

[0055]

[0056] In the formula, C t For time-series correlation, max is the maximum value function, T A T B These are the timing sequences of the faulty component and the associated components, respectively. and The sensor readings of the faulty component and the fault-related component at the i-th time point are respectively, where Arver is the averaging function and n is the length of the time series;

[0057] If the spatial overlap is greater than the first threshold and the temporal correlation is greater than the second threshold, it is a coupled fault; otherwise, it is an independent fault.

[0058] Preferably, obtaining the probabilistic tracing report of component failure based on the first probability of failure and the second probability of failure specifically includes:

[0059] Based on decision trees and multimodal sensor data, the causes and probabilities of failure of independent faulty components are obtained;

[0060] Based on component-level association rules, obtain the path of coupled faulty components;

[0061] Based on real fault images and multimodal sensor data, the confidence level of the multimodal sensor is obtained based on the accuracy.

[0062] Specifically, the confidence level of the multimodal sensor is as follows:

[0063]

[0064] Using 24 hours as a time window, obtain the time decay factor within 7 time windows;

[0065] The time decay factor is specifically:

[0066] TF = e -0.2*(t-k) ;

[0067] In the formula, TF is the time decay factor, t is the current time, and k is the historical time, ranging from t-6 to t;

[0068] Based on the confidence level of the multimodal sensor and the time decay factor, an adjustment factor for the second probability of fault occurrence is obtained;

[0069] The adjustment factor for the second probability of failure is specifically as follows:

[0070] w k =TF*Con;

[0071] Based on the adjustment factor of the second probability of fault occurrence and multimodal sensor data, obtain the rate of change of the second probability of fault occurrence;

[0072] Specifically, the rate of change of the second probability of the fault occurrence is as follows:

[0073]

[0074] In the formula, P w Let P(k) be the rate of change of the second probability of failure, P(k) be the second probability of failure of the faulty component at time k, and P(k-1) be the second probability of failure of the faulty component at time k-1.

[0075] The second probability of failure occurrence is adjusted based on the rate of change of the second probability of failure occurrence;

[0076] Specifically, the adjusted second probability of fault occurrence is as follows:

[0077] γ=0.5*(1+tanh(10*(C t -0.5)));

[0078] P2(t)=P(t)+γ*P w *O s ;

[0079] In the formula, γ is the gain coefficient, tanh is the hyperbolic tangent function, and C t Let Pt be the mean of the temporal correlation of each faulty component along the path of the coupled faulty component at time t, P2(t) be the adjusted second probability of fault occurrence, and P(t) be the second probability of fault occurrence of the faulty component at time t. w O is the rate of change of the second probability of the fault occurring. sLet be the mean of the spatial overlap of each faulty component on the path of the coupled faulty component at time t;

[0080] Based on the first probability of fault occurrence and the adjusted second probability of fault occurrence, and using Bayes' theorem, the probability of each power transformer component on the coupled fault component path is obtained.

[0081] Based on the probability of each power transformer component failing along the coupled faulty component path, a probabilistic tracing report of the faulty power transformer component is obtained.

[0082] Furthermore, a knowledge graph-based intelligent power fault monitoring system is proposed to implement the aforementioned knowledge graph-based intelligent power fault monitoring method, including:

[0083] The main control module is used to time-align multimodal sensor data through a high-precision time synchronization module, acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gating attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on power equipment topology and the historical fault database, generate component-level association rules, construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of power transformer failure based on component-level association rules and cases in the historical fault database, acquire power transformer feature data information, upload the power transformer feature data information to the fault neural network for inference, and output a probabilistic tracing report containing fault component location and evolution path based on the inference results of the fault neural network.

[0084] The neural network module is used to acquire a fault image set of power transformers, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, acquire an expanded fault image set based on a conditional generative adversarial network based on the rare fault categories, label the images in the expanded fault image set according to the historical fault database, and acquire a fault neural network based on the labeled expanded fault image set.

[0085] The output module is used to obtain fault information of the power transformer based on the fault neural network, determine the fault type of the power transformer based on multimodal sensor data, and if the fault type is a coupled fault, obtain the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component, and obtain the probabilistic tracing report of the component fault based on the first probability of the fault and the second probability of the fault.

[0086] Optionally, the main control module specifically includes:

[0087] A time-stamp alignment unit is used to perform time-stamp alignment on multimodal sensor data through a high-precision time synchronization module;

[0088] The fault probability unit is used to acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gated attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on the power equipment topology and the historical fault database, generate component-level association rules, and construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of a power transformer failure based on the component-level association rules and cases in the historical fault database.

[0089] The probabilistic path unit is used to acquire characteristic data information of the power transformer and upload the characteristic data information of the power transformer to the fault neural network for inference. Based on the inference result of the fault neural network, a probabilistic tracing report containing the location and evolution path of the faulty component is output.

[0090] Optionally, the neural network module specifically includes:

[0091] An expansion unit is used to acquire a fault image set of a power transformer, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, and acquire an expanded fault image set based on a conditional generative adversarial network according to the rare fault categories.

[0092] The training unit is used to annotate images in the expanded fault image set according to the historical fault database, and to obtain a fault neural network based on the annotated expanded fault image set.

[0093] Optionally, the output module specifically includes:

[0094] The acquisition unit is used to acquire fault information of the power transformer according to the fault neural network, determine the fault type of the power transformer according to the multimodal sensor data, and if the fault type is a coupled fault, acquire the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component.

[0095] The fault location unit is used to obtain a probabilistic source tracing report of component faults based on a first probability of fault occurrence and a second probability of fault occurrence.

[0096] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0097] This invention proposes a knowledge graph-based intelligent monitoring method and system for power faults. By combining multimodal sensor data with time-stamp alignment technology, a learnable gated attention mechanism, and a knowledge graph, it achieves intelligent monitoring and precise location of power transformer faults. The fault image set is expanded using conditional generative adversarial networks, and a fault neural network is constructed based on a historical fault database, effectively identifying rare fault types and improving the accuracy of fault diagnosis. Simultaneously, the probability of fault occurrence is evaluated through component-level association rules and decision tree algorithms, outputting a probabilistic tracing report containing the location of the faulty component and its evolution path. This significantly improves the timeliness and reliability of fault detection, reduces power outage time and maintenance costs, and provides strong technical support for the stable operation of power systems. Attached Figure Description

[0098] Figure 1 This is a flowchart of a knowledge graph-based intelligent power fault monitoring method proposed in this invention.

[0099] Figure 2 This is a flowchart of the fault neural network acquisition process in this invention;

[0100] Figure 3 This is a flowchart illustrating the probabilistic tracing report acquisition process that includes fault component location and evolution path in this invention.

[0101] Figure 4 This is a block diagram of a knowledge graph-based intelligent power fault monitoring system proposed in this invention. Detailed Implementation

[0102] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0103] Reference Figure 1 - Figure 3 As shown in the figure, an intelligent power fault monitoring method based on knowledge graphs in an embodiment of the present invention includes:

[0104] Obtain power transformer structural information, which includes power equipment topology, historical fault database, and multimodal sensor data corresponding to the power transformer;

[0105] The multimodal sensor data is time-stamped using a high-precision time synchronization module;

[0106] Based on dual-stream feature data acquired by multimodal sensors, and using a learnable gating attention mechanism, the confidence levels of static and dynamic feature stream data are obtained.

[0107] Obtain the faulty neural network based on the historical fault database;

[0108] Specifically, based on a historical fault database, a fault neural network is obtained, including:

[0109] Obtain a set of fault images of power transformers;

[0110] Based on the fault images, obtain the number of samples of each fault type in the image dataset, and filter out the rare fault categories.

[0111] Based on the scarce fault categories, an expanded fault image set is obtained using a conditional generative adversarial network.

[0112] Based on the historical fault database, the images in the expanded fault image set are labeled, and the labels are used to delineate the fault areas and fault types of power transformers.

[0113] Based on the annotated and expanded fault image set, obtain the fault neural network;

[0114] Specifically, the step of obtaining an expanded fault image set based on a conditional generative adversarial network according to scarce fault categories includes:

[0115] Based on a set of images with scarce fault categories, a synthetic image with a specified fault category is obtained using a generative adversarial network.

[0116] Based on image preprocessing, the synthesized image is standardized, including scaling the image region corresponding to the faulty component and normalizing the illumination.

[0117] Based on the image set of rare fault categories, and based on image preprocessing, the image region corresponding to the faulty component in the real fault image is obtained. The real fault image is an image of the faulty component in the historical fault database.

[0118] Based on the image regions of faulty components in real and synthetic images, obtain structural similarity index, distortion index, and spatial similarity.

[0119] The structural similarity index is specifically as follows:

[0120]

[0121] In the formula, SSIM is the structural similarity index between real fault images and synthetic images, μ x μ y These are the average values ​​of real fault images and synthetic images, respectively. σ represents the variance of the real fault image and the synthetic image, respectively. xy The covariance between real fault images and synthetic images;

[0122] The distortion index is as follows:

[0123]

[0124] In the formula, MSE is the transition value for calculating the distortion index, DDIN is the distortion index, x is the real faulty image, y is the composite image, L is the maximum pixel value, M is the length of the image, and N is the width of the image.

[0125] Spatial similarity is specifically defined as:

[0126] Sim = BERT(n x ,n y );

[0127] In the formula, Sim represents spatial similarity, BERT represents the cosine similarity function, and n... x ,n y These are the feature vectors extracted from real fault images and synthetic images by convolutional neural networks, respectively.

[0128] The accuracy of the synthesized image is obtained based on the structural similarity index, the distortion index, and the spatial similarity.

[0129] Specifically, the accuracy of the synthesized image is as follows:

[0130] Q=0.1*SSIM+0.1*DDIN+0.8*Sim;

[0131] In the formula, Q represents the accuracy of the synthesized image;

[0132] If the accuracy of the synthesized image is within a preset threshold range, then the synthesized image will be added to the expanded set of fault images.

[0133] In this scheme, the number of samples of each fault type in the image dataset is obtained through the image set of power transformers, and rare fault categories are filtered out. Based on the rare fault categories, an expanded fault image set is obtained based on a conditional generative adversarial network. The images in the expanded fault image set are labeled according to the historical fault database. Based on the labeled expanded fault image set, a fault neural network is obtained. The conditional adversarial network is used to solve the problem of data scarcity. At the same time, a multi-dimensional evaluation system is proposed to ensure data quality and model performance.

[0134] It is understandable that training deep learning network models with historical fault images can improve the identification ability of power transformer faults. However, in practical applications, training deep learning network models solely based on historical fault images can lead to overfitting because deep learning network models cannot learn the distribution characteristics of the data. For example, faults such as multi-point grounding of the iron core, loose silicon steel sheets, and bushing explosions can have serious consequences, but these are relatively rare faults in power transformers. Therefore, there are not enough fault images for deep learning neural networks to fully learn the characteristics of these faults. By generating images of rare fault categories through conditional generative adversarial networks, and further filtering the generated images of rare fault categories by calculating the structural similarity index, distortion index, and spatial similarity of the image regions of faulty components in real fault images and rare fault category images, the quantity and quality of the training dataset are improved. This makes the model more robust and adaptable, enhances its ability to identify various faults, and makes the monitoring results of power transformer faults more accurate.

[0135] A fault knowledge graph is constructed based on the power equipment topology and historical fault database, and component-level association rules are generated.

[0136] Based on the confidence levels of multimodal sensor data, static feature stream data, and dynamic feature stream data, a decision tree is constructed using component-level association rules and cases from the historical fault database to obtain the probability of power transformer failure.

[0137] If the probability of a power transformer malfunctioning is greater than a preset threshold, the power transformer characteristic data information is acquired and uploaded to the fault neural network for inference. The power transformer characteristic data information includes the appearance information, humidity distribution information, and temperature distribution information of the power transformer components.

[0138] Based on the results of fault neural network inference, a probabilistic tracing report containing the location of faulty components and their evolution paths is output.

[0139] Specifically, the result of the fault neural network inference outputs a probabilistic tracing report containing the location of the faulty component and its evolution path, including:

[0140] Based on the fault neural network, fault information of the power transformer is obtained, which includes the faulty component of the power transformer, the fault type, and the first probability of the faulty component failing.

[0141] Based on multimodal sensor data, the fault type of the power transformer is determined. The fault type includes independent faults and coupled faults. The coupled fault is a compound fault with a causal relationship caused by the interaction between two or more power transformer component faults.

[0142] If the fault type is a coupled fault, then based on the fault information of the power transformer and the actual fault image of the component, the second probability of the power transformer component fault is obtained. The second probability of the fault is the accuracy of the fault component area in the fault information of the power transformer and the fault component area in the actual fault image.

[0143] Based on the first probability and the second probability of failure, obtain a probabilistic tracing report of component failure;

[0144] Specifically, determining the fault type of the power transformer based on multimodal sensor data includes:

[0145] Based on multimodal sensor data, abnormal data is obtained, which is data that deviates from the normal operating range of the power transformer;

[0146] Based on component-level association rules, multi-dimensional auxiliary data is obtained, which is data that has a logical association with the abnormal data in the component-level association rules.

[0147] Based on the knowledge graph, a fault conflict matrix is ​​obtained, which is a three-dimensional matrix containing power transformer components, fault types, and fault characteristics.

[0148] The following is a partial fault conflict matrix:

[0149]

[0150] Based on the fault conflict matrix, fault-related components are obtained. These fault-related components are those that may fail, which are found in the component-level association rules based on the abnormal data.

[0151] Based on the faulty component and its associated components, spatial overlap and temporal correlation are obtained using multimodal sensor data.

[0152] Specifically, the spatial overlap is as follows:

[0153]

[0154] In the formula, O s For spatial overlap, Area(Sensor) A Area(Sensor) B These are the sensor monitoring areas for the faulty component and the related faulty component, respectively.

[0155] The specific time-series correlation is as follows:

[0156]

[0157] In the formula, Ct For time-series correlation, max is the maximum value function, T A T B These are the timing sequences of the faulty component and the associated components, respectively. and The sensor readings of the faulty component and the fault-related component at the i-th time point are respectively, where Arver is the averaging function and n is the length of the time series;

[0158] If the spatial overlap is greater than the first threshold and the temporal correlation is greater than the second threshold, it is a coupled fault; otherwise, it is an independent fault.

[0159] Specifically, obtaining the probabilistic tracing report of component failure based on the first probability of failure and the second probability of failure includes:

[0160] Based on decision trees and multimodal sensor data, the causes and probabilities of failure of independent faulty components are obtained;

[0161] Based on component-level association rules, obtain the path of coupled faulty components;

[0162] For example:

[0163] Based on the abnormal data, the fault-related components are identified as components A, B, C, D, and E, with the probability of failure for components A, B, C, D, and E being 0.8, 0.2, 0.6, 0.5, and 0.4, respectively.

[0164] Based on spatial overlap and temporal correlation, components A and B, A and C, C and D, and C and E are identified as coupling faults, with a first threshold of 0.5 and a second threshold of 0.8.

[0165] The probability of failure from component A to component B is calculated as: 0.8 * 0.2 * 2 = 0.32;

[0166] The probability of failure from component A to component C to component D is calculated as: 0.8 * 0.6 * 0.5 * 3 = 0.72;

[0167] The probability of failure from component A to component C to component E is calculated as: 0.8 * 0.6 * 0.4 * 3 = 0.576;

[0168] Comparing the failure probabilities of the three paths mentioned above, the path from component A to component C to component D is taken as the path of coupled failure components.

[0169] Based on real fault images and multimodal sensor data, the confidence level of the multimodal sensor is obtained based on the accuracy.

[0170] Specifically, the confidence level of the multimodal sensor is as follows:

[0171]

[0172] Using 24 hours as a time window, obtain the time decay factor within 7 time windows;

[0173] The time decay factor is specifically:

[0174] TF = e -0.2*(t-k) ;

[0175] In the formula, TF is the time decay factor, t is the current time, and k is the historical time, ranging from t-6 to t;

[0176] Based on the confidence level of the multimodal sensor and the time decay factor, an adjustment factor for the second probability of fault occurrence is obtained;

[0177] The adjustment factor for the second probability of failure is specifically as follows:

[0178] w k =TF*Con;

[0179] Based on the adjustment factor of the second probability of fault occurrence and multimodal sensor data, obtain the rate of change of the second probability of fault occurrence;

[0180] Specifically, the rate of change of the second probability of the fault occurrence is as follows:

[0181]

[0182] In the formula, P w Let P(k) be the rate of change of the second probability of failure, P(k) be the second probability of failure of the faulty component at time k, and P(k-1) be the second probability of failure of the faulty component at time k-1.

[0183] The second probability of failure occurrence is adjusted based on the rate of change of the second probability of failure occurrence;

[0184] Specifically, the adjusted second probability of fault occurrence is as follows:

[0185] γ=0.5*(1+tanh(10*(C t -0.5)));

[0186] P2(t)=P(t)+γ*P w *O s ;

[0187] In the formula, γ is the gain coefficient, tanh is the hyperbolic tangent function, and C tLet Pt be the mean of the temporal correlation of each faulty component along the path of the coupled faulty component at time t, P2(t) be the adjusted second probability of fault occurrence, and P(t) be the second probability of fault occurrence of the faulty component at time t. w O is the rate of change of the second probability of the fault occurring. s Let be the mean of the spatial overlap of each faulty component on the path of the coupled faulty component at time t;

[0188] Based on the first probability of fault occurrence and the adjusted second probability of fault occurrence, and using Bayes' theorem, the probability of each power transformer component on the coupled fault component path is obtained.

[0189] Based on the probability of each power transformer component failing along the coupled faulty component path, a probabilistic tracing report of the faulty power transformer component is obtained.

[0190] In this scheme, the gain coefficient is determined by the time-series correlation C. t The gain coefficient can be calculated to quantify the temporal causal strength of fault propagation. When C t When γ > 0.8, γ ≈ 1, allowing the rate of change to fully adjust the second probability of fault occurrence, when C t When γ is less than 0.3, γ≈0, completely suppressing the adjustment and ensuring that probabilistic corrections are only made for faults with clear temporal evolution patterns; P w By calculating the time decay factor and sensor confidence level within the sliding window, the trend of smooth probability change is calculated, filtering out instantaneous noise (such as sensor false alarms) while amplifying continuously changing signals (such as slowly rising temperature), providing a stable and reliable indication of the direction of fault evolution; spatial overlap is O s As a quantitative indicator of spatial coupling, it automatically shields interference signals from spatial isolation components; through the combined effect of the three, it achieves precise adjustment of the second probability of fault occurrence, improving accuracy while enhancing the efficiency of tracing through parameter interpretability.

[0191] Reference Figure 4 As shown, further, combining the above-mentioned knowledge graph-based intelligent power fault monitoring method, a knowledge graph-based intelligent power fault monitoring system is proposed, including:

[0192] The main control module is used to time-align multimodal sensor data through a high-precision time synchronization module, acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gating attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on power equipment topology and the historical fault database, generate component-level association rules, construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of power transformer failure based on component-level association rules and cases in the historical fault database, acquire power transformer feature data information, upload the power transformer feature data information to the fault neural network for inference, and output a probabilistic tracing report containing fault component location and evolution path based on the inference results of the fault neural network.

[0193] The neural network module is used to acquire a fault image set of power transformers, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, acquire an expanded fault image set based on a conditional generative adversarial network based on the rare fault categories, label the images in the expanded fault image set according to the historical fault database, and acquire a fault neural network based on the labeled expanded fault image set.

[0194] The output module is used to obtain fault information of the power transformer based on the fault neural network, determine the fault type of the power transformer based on multimodal sensor data, and if the fault type is a coupled fault, obtain the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component, and obtain the probabilistic tracing report of the component fault based on the first probability of the fault and the second probability of the fault.

[0195] The main control module specifically includes:

[0196] A time-stamp alignment unit is used to perform time-stamp alignment on multimodal sensor data through a high-precision time synchronization module;

[0197] The fault probability unit is used to acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gated attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on the power equipment topology and the historical fault database, generate component-level association rules, and construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of a power transformer failure based on the component-level association rules and cases in the historical fault database.

[0198] The probabilistic path unit is used to acquire characteristic data information of the power transformer and upload the characteristic data information of the power transformer to the fault neural network for inference. Based on the inference result of the fault neural network, a probabilistic tracing report containing the location and evolution path of the faulty component is output.

[0199] The neural network module specifically includes:

[0200] An expansion unit is used to acquire a fault image set of a power transformer, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, and acquire an expanded fault image set based on a conditional generative adversarial network according to the rare fault categories.

[0201] The training unit is used to annotate images in the expanded fault image set according to the historical fault database, and to obtain a fault neural network based on the annotated expanded fault image set.

[0202] The output module specifically includes:

[0203] The acquisition unit is used to acquire fault information of the power transformer according to the fault neural network, determine the fault type of the power transformer according to the multimodal sensor data, and if the fault type is a coupled fault, acquire the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component.

[0204] The fault location unit is used to obtain a probabilistic source tracing report of component faults based on a first probability of fault occurrence and a second probability of fault occurrence.

[0205] In summary, the advantages of this invention are as follows: By combining multimodal sensor data with time-stamp alignment technology, a learnable gated attention mechanism, and a knowledge graph, intelligent monitoring and precise location of power transformer faults are achieved. Expanding the fault image set using conditional generative adversarial networks and constructing a fault neural network based on a historical fault database effectively identifies rare fault types and improves the accuracy of fault diagnosis. Simultaneously, by evaluating the probability of fault occurrence through component-level association rules and decision tree algorithms, a probabilistic source tracing report containing the location and evolution path of the faulty component is output, significantly improving the timeliness and reliability of fault detection, reducing power outage time and maintenance costs, and providing strong technical support for the stable operation of the power system.

[0206] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A knowledge graph-based intelligent monitoring method for power faults, characterized in that, include: Obtain power transformer structural information, which includes power equipment topology, historical fault database, and multimodal sensor data corresponding to the power transformer; The multimodal sensor data is time-stamped using a high-precision time synchronization module; Based on dual-stream feature data acquired by multimodal sensors, and using a learnable gating attention mechanism, the confidence levels of static and dynamic feature stream data are obtained. Obtain the faulty neural network based on the historical fault database; A fault knowledge graph is constructed based on the power equipment topology and historical fault database, and component-level association rules are generated. Based on the confidence levels of multimodal sensor data, static feature stream data, and dynamic feature stream data, a decision tree is constructed using component-level association rules and cases from the historical fault database to obtain the probability of power transformer failure. If the probability of a power transformer malfunctioning is greater than a preset threshold, the power transformer characteristic data information is acquired and uploaded to the fault neural network for inference. The power transformer characteristic data information includes the appearance information, humidity distribution information, and temperature distribution information of the power transformer components. Based on the results of fault neural network inference, a probabilistic tracing report is output, which includes the location of the faulty component and its evolution path.

2. The intelligent power fault monitoring method based on knowledge graph according to claim 1, characterized in that, The step of obtaining the fault neural network based on the historical fault database specifically includes: Obtain a set of fault images of power transformers; Based on the fault images, obtain the number of samples of each fault type in the image dataset, and filter out the rare fault categories. Based on the scarce fault categories, an expanded fault image set is obtained using a conditional generative adversarial network. Based on the historical fault database, the images in the expanded fault image set are labeled, and the labels are used to delineate the fault areas and fault types of power transformers. Based on the annotated and expanded fault image set, obtain the fault neural network.

3. The intelligent power fault monitoring method based on knowledge graph according to claim 2, characterized in that, The step of obtaining an expanded fault image set based on a conditional generative adversarial network according to scarce fault categories specifically includes: Based on a set of images with scarce fault categories, a synthetic image with a specified fault category is obtained using a generative adversarial network. Based on image preprocessing, the synthesized image is standardized, including scaling the image region corresponding to the faulty component and normalizing the illumination. Based on the image set of rare fault categories, and based on image preprocessing, the image region corresponding to the faulty component in the real fault image is obtained. The real fault image is an image of the faulty component in the historical fault database. Based on the image regions of faulty components in real and synthetic images, obtain structural similarity index, distortion index, and spatial similarity. The structural similarity index is specifically as follows: In the formula, SSIM is the structural similarity index between real fault images and synthetic images, μ x μ y These are the average values ​​of real fault images and synthetic images, respectively. σ represents the variance of the real fault image and the synthetic image, respectively. xy The covariance between real fault images and synthetic images; The distortion index is as follows: In the formula, MSE is the transition value for calculating the distortion index, DDIN is the distortion index, x is the real faulty image, y is the composite image, L is the maximum pixel value, M is the length of the image, and N is the width of the image. Spatial similarity is specifically defined as: Sim=BERT(n x ,n y ); In the formula, Sim represents spatial similarity, BERT represents the cosine similarity function, and n... x ,n y These are the feature vectors extracted from real fault images and synthetic images by convolutional neural networks, respectively. The accuracy of the synthesized image is obtained based on the structural similarity index, the distortion index, and the spatial similarity. Specifically, the accuracy of the synthesized image is as follows: Q=0.1*SSIM+0.1*DDIN+0.8*Sim; In the formula, Q represents the accuracy of the synthesized image; If the accuracy of the synthesized image is within a preset threshold range, then the synthesized image will be added to the expanded set of fault images.

4. The intelligent power fault monitoring method based on knowledge graph according to claim 1, characterized in that, The results of the fault neural network inference output a probabilistic source tracing report containing the location and evolution path of the faulty component, specifically including: Based on the fault neural network, fault information of the power transformer is obtained, which includes the faulty component of the power transformer, the fault type, and the first probability of the faulty component failing. Based on multimodal sensor data, the fault type of the power transformer is determined. The fault type includes independent faults and coupled faults. The coupled fault is a compound fault with a causal relationship caused by the interaction between two or more power transformer component faults. If the fault type is a coupled fault, then based on the fault information of the power transformer and the actual fault image of the component, the second probability of the power transformer component fault is obtained. The second probability of the fault is the accuracy of the fault component area in the fault information of the power transformer and the fault component area in the actual fault image. Based on the first probability and the second probability of failure, a probabilistic tracing report of component failure is obtained.

5. The intelligent power fault monitoring method based on knowledge graph according to claim 4, characterized in that, The method of determining the fault type of the power transformer based on multimodal sensor data specifically includes: Based on multimodal sensor data, abnormal data is obtained, which is data that deviates from the normal operating range of the power transformer; Based on component-level association rules, multi-dimensional auxiliary data is obtained, which is data that has a logical association with the abnormal data in the component-level association rules. Based on the knowledge graph, a fault conflict matrix is ​​obtained, which is a three-dimensional matrix containing power transformer components, fault types, and fault characteristics. Based on the fault conflict matrix, fault-related components are obtained. These fault-related components are those that may fail, which are found in the component-level association rules based on the abnormal data. Based on the faulty component and its associated components, spatial overlap and temporal correlation are obtained using multimodal sensor data. Specifically, the spatial overlap is as follows: In the formula, O s For spatial overlap, Area(Sevsor) A Area(Sensor) B These are the sensor monitoring areas for the faulty component and the related faulty component, respectively. The specific time-series correlation is as follows: In the formula, C t For time-series correlation, max is the maximum value function, T A T B These are the timing sequences of the faulty component and the associated components, respectively. and The sensor readings of the faulty component and the fault-related component at the i-th time point are respectively, where Arver is the averaging function and n is the length of the time series; If the spatial overlap is greater than the first threshold and the temporal correlation is greater than the second threshold, it is a coupled fault; otherwise, it is an independent fault.

6. The intelligent power fault monitoring method based on knowledge graph according to claim 4, characterized in that, The process of obtaining a probabilistic tracing report of component failure based on a first probability of failure and a second probability of failure specifically includes: Based on decision trees and multimodal sensor data, the causes and probabilities of failure of independent faulty components are obtained; Based on component-level association rules, obtain the path of coupled faulty components; Based on real fault images and multimodal sensor data, the confidence level of the multimodal sensor is obtained based on the accuracy. Specifically, the confidence level of the multimodal sensor is as follows: Using 24 hours as a time window, obtain the time decay factor within 7 time windows; The time decay factor is specifically: TF=e -0.2*(t-k) ; In the formula, TF is the time decay factor, t is the current time, and k is the historical time, ranging from t-6 to t; Based on the confidence level of the multimodal sensor and the time decay factor, an adjustment factor for the second probability of fault occurrence is obtained; The adjustment factor for the second probability of failure is specifically as follows: w k =TF*Con; Based on the adjustment factor of the second probability of fault occurrence and multimodal sensor data, obtain the rate of change of the second probability of fault occurrence; Specifically, the rate of change of the second probability of the fault occurrence is as follows: In the formula, P w Let P(k) be the rate of change of the second probability of failure, P(k) be the second probability of failure of the faulty component at time k, and P(k-1) be the second probability of failure of the faulty component at time k-1. The second probability of failure occurrence is adjusted based on the rate of change of the second probability of failure occurrence; Specifically, the adjusted second probability of fault occurrence is as follows: γ=0.5*(1+tanh(10*(C t -0.5))); P2(t)=P(t)+γ*P w *O s ; In the formula, γ is the gain coefficient, tanh is the hyperbolic tangent function, and C t Let Pt be the mean of the temporal correlation of each faulty component along the path of the coupled faulty component at time t, P2(t) be the adjusted second probability of fault occurrence, and P(t) be the second probability of fault occurrence of the faulty component at time t. w O is the rate of change of the second probability of the fault occurring. s Let be the mean of the spatial overlap of each faulty component on the path of the coupled faulty component at time t; Based on the first probability of fault occurrence and the adjusted second probability of fault occurrence, and using Bayes' theorem, the probability of each power transformer component on the coupled fault component path is obtained. Based on the probability of each power transformer component failing along the coupled faulty component path, a probabilistic tracing report of the faulty power transformer component is obtained.

7. A knowledge graph-based intelligent power fault monitoring system, used to implement the knowledge graph-based intelligent power fault monitoring method as described in any one of claims 1-6, characterized in that, include: The main control module is used to time-align multimodal sensor data through a high-precision time synchronization module, acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gating attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on power equipment topology and the historical fault database, generate component-level association rules, construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of power transformer failure based on component-level association rules and cases in the historical fault database, acquire power transformer feature data information, upload the power transformer feature data information to the fault neural network for inference, and output a probabilistic tracing report containing fault component location and evolution path based on the inference results of the fault neural network. The neural network module is used to acquire a fault image set of power transformers, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, acquire an expanded fault image set based on a conditional generative adversarial network based on the rare fault categories, label the images in the expanded fault image set according to the historical fault database, and acquire a fault neural network based on the labeled expanded fault image set. The output module is used to obtain fault information of the power transformer based on the fault neural network, determine the fault type of the power transformer based on multimodal sensor data, and if the fault type is a coupled fault, obtain the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component, and obtain the probabilistic tracing report of the component fault based on the first probability of the fault and the second probability of the fault.

8. The intelligent power fault monitoring system based on knowledge graph according to claim 7, characterized in that, The main control module specifically includes: A time-stamp alignment unit is used to perform time-stamp alignment on multimodal sensor data through a high-precision time synchronization module; The fault probability unit is used to acquire dual-stream feature data from multimodal sensors, obtain confidence scores for static and dynamic feature stream data based on a learnable gated attention mechanism, acquire a fault neural network based on a historical fault database, construct a fault knowledge graph based on the power equipment topology and the historical fault database, generate component-level association rules, and construct a decision tree based on the confidence scores of multimodal sensor data, static feature stream data, and dynamic feature stream data, and obtain the probability of a power transformer failure based on the component-level association rules and cases in the historical fault database. The probabilistic path unit is used to acquire characteristic data information of the power transformer and upload the characteristic data information of the power transformer to the fault neural network for inference. Based on the inference result of the fault neural network, a probabilistic tracing report containing the location and evolution path of the faulty component is output.

9. A knowledge graph-based intelligent power fault monitoring system according to claim 7, characterized in that, The neural network module specifically includes: An expansion unit is used to acquire a fault image set of a power transformer, acquire the number of samples of each fault type in the image dataset based on the fault images, filter out rare fault categories, and acquire an expanded fault image set based on a conditional generative adversarial network according to the rare fault categories. The training unit is used to annotate images in the expanded fault image set according to the historical fault database, and to obtain a fault neural network based on the annotated expanded fault image set.

10. A knowledge graph-based intelligent power fault monitoring system according to claim 7, characterized in that, The output module specifically includes: The acquisition unit is used to acquire fault information of the power transformer according to the fault neural network, determine the fault type of the power transformer according to the multimodal sensor data, and if the fault type is a coupled fault, acquire the second probability of the power transformer component fault based on the fault information of the power transformer and the actual fault image of the component. The fault location unit is used to obtain a probabilistic source tracing report of component faults based on a first probability of fault occurrence and a second probability of fault occurrence.

Images