A method and system for analyzing operation faults of GIS equipment based on 3D modeling

Abstract

The present invention relates to the technical field of data processing, and discloses a method and system for analyzing operation faults of GIS devices based on three-dimensional modeling, which are used to improve the accuracy of analyzing operation faults of GIS devices based on three-dimensional modeling. The method includes: performing data cleaning processing on the collected operation status data, environmental data, and historical fault data of GIS devices to obtain a cleaned data set; performing data denoising on the cleaned data set to obtain a denoised data set; filling in missing values of the denoised data set through a variational autoencoder to obtain a filled data set; performing three-dimensional modeling based on the filled data set to obtain a three-dimensional model of GIS devices; collecting real-time operation data of GIS devices, and extracting fault features from the three-dimensional model of GIS devices through the real-time operation data to obtain a fault feature vector; inputting the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method and system for analyzing operation faults of GIS devices based on three-dimensional modeling. Background Art

[0002] Gas-insulated switchgear (GIS) devices are widely used in power systems, and their operation stability and reliability are crucial for the safety of power systems. At present, the fault detection and analysis methods for GIS devices mainly include infrared imaging, ultrasonic detection, and partial discharge detection, etc. These methods rely on manual experience and traditional instruments to monitor the operation state of devices and diagnose faults, and to a certain extent, they can detect device faults and provide maintenance references.

[0003] Disadvantages: Traditional fault detection methods have certain limitations. First, the detection accuracy of these methods for complex faults is not high, and it is difficult to comprehensively capture various fault characteristics generated during the operation of GIS devices. Second, the existing technologies perform poorly in terms of real-time performance and cannot monitor faults in real time and respond quickly. In addition, these methods cannot effectively utilize a large amount of operation data for fault trend analysis and prediction, resulting in low information utilization rate and difficulty in meeting the requirements of modern power systems for efficient and accurate fault detection. Summary of the Invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention is proposed.

[0005] Therefore, the prior art problem to be solved by the present invention is that the detection accuracy for complex faults is not high, and it is difficult to comprehensively capture various fault characteristics generated during the operation of GIS devices

[0006] To solve the above technical problems, the present invention provides the following technical solution: A method for analyzing operation faults of GIS devices based on three-dimensional modeling, which includes performing data cleaning processing on the collected operation state data, environmental data, and historical fault data of GIS devices to obtain a cleaned data set;

[0007] Performing data denoising on the cleaned data set to obtain a denoised data set;

[0008] Performing missing value filling on the denoised data set through a variational autoencoder to obtain a filled data set;

[0009] Performing three-dimensional modeling according to the filled data set to obtain a three-dimensional model of the GIS device;

[0010] Collecting the real-time operation data of the GIS device, and extracting fault characteristics from the three-dimensional model of the GIS device through the real-time operation data to obtain a fault feature vector;

[0011] Input the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data.

[0012] As a preferred embodiment of the method for analyzing the operating faults of GIS equipment based on 3D modeling according to the present invention, wherein: the step of performing data cleaning on the collected operating state data, environmental data, and historical fault data of the GIS equipment to obtain a cleaned data set includes:

[0013] Perform outlier detection and processing on the operating state data of the GIS equipment to obtain processed operating state data;

[0014] Perform data fusion based on the processed operating state data and environmental data to obtain fusion data;

[0015] Perform feature engineering on the fusion data to extract key feature vectors and obtain a feature vector data set;

[0016] Construct a label data set based on the feature vector data set and historical fault data;

[0017] Perform data splitting on the label data set to obtain a training data set and a validation data set;

[0018] Train a classification model based on the training data set to obtain a trained classification model;

[0019] Evaluate the trained classification model using the validation data set to obtain an evaluation result;

[0020] Optimize the trained classification model according to the evaluation result to obtain an optimized classification model;

[0021] Apply the optimized classification model to the feature vector data set to classify the data and obtain classification data;

[0022] Perform data cleaning on the classification data to obtain a cleaned data set.

[0023] As a preferred embodiment of the method for analyzing the operating faults of GIS equipment based on 3D modeling according to the present invention, wherein: the step of performing data denoising on the cleaned data set to obtain a denoised data set includes:

[0024] Construct a Gaussian mixture model based on the cleaned data set to obtain Gaussian mixture model parameters;

[0025] Perform outlier detection on each data point in the cleaned data set, calculate its outlier score according to the Gaussian mixture model parameters, and obtain an outlier score set;

[0026] Set an anomaly threshold according to the set of anomaly scores, and mark the data points with anomaly scores higher than the anomaly threshold as anomaly points to obtain the marked data set;

[0027] Perform principal component analysis on the marked data set to extract the principal component load matrix and obtain the principal component load matrix;

[0028] Perform projection transformation on the marked data set according to the principal component load matrix to obtain the projected data set;

[0029] Perform kernel discriminant analysis on the projected data set to construct a discriminant model and obtain the discriminant model;

[0030] Use the discriminant model to discriminate the anomaly points in the projected data set to obtain the anomaly point discrimination result;

[0031] According to the anomaly point discrimination result, remove the data discriminated as anomaly points from the projected data set to obtain the denoised projected data set;

[0032] Perform back-projection transformation on the denoised projected data set to obtain the back-projected data set;

[0033] Perform data format conversion on the back-projected data set to obtain the denoised data set.

[0034] As a preferred solution of the method for analyzing the operation faults of GIS devices based on 3D modeling according to the present invention, wherein: the step of filling the missing values of the denoised data set through a variational autoencoder includes:

[0035] Perform standardization processing on the denoised data set to obtain the standardized data set;

[0036] Construct a variational autoencoder model according to the standardized data set, including an encoder and a decoder network, to obtain the variational autoencoder model;

[0037] Apply the encoder network in the variational autoencoder model to the standardized data set to obtain the latent space distribution;

[0038] Sample latent variables from the latent space distribution to obtain a set of latent variables;

[0039] Input the set of latent variables into the decoder network of the variational autoencoder model to obtain the reconstructed data set;

[0040] Calculate the reconstruction loss according to the standardized data set and the reconstructed data set;

[0041] Calculate the KL divergence loss according to the latent space distribution;

[0042] Add the reconstructed loss and the KL divergence loss to obtain the total loss;

[0043] Update the parameters of the variational autoencoder model based on the total loss to obtain an updated variational autoencoder model;

[0044] Use the updated variational autoencoder model to fill in the missing values in the denoised data set to obtain the filled data set.

[0045] As a preferred solution of the method for analyzing the operation faults of GIS devices based on 3D modeling according to the present invention, wherein: the step of performing 3D modeling according to the filled data set to obtain a 3D model of the GIS device includes:

[0046] Extract features from the filled data set to obtain a feature data set;

[0047] Construct 3D point cloud data according to the feature data set to obtain the original point cloud data;

[0048] Perform point cloud filtering on the original point cloud data to remove outliers and obtain filtered point cloud data;

[0049] Perform point cloud segmentation on the filtered point cloud data to extract key point cloud clusters and obtain segmented point cloud data;

[0050] Perform 3D reconstruction on the segmented point cloud data to construct a 3D mesh model and obtain a rough 3D model;

[0051] Perform mesh optimization on the rough 3D model, including vertex refinement and patch subdivision, to obtain an optimized 3D model;

[0052] Perform texture mapping on the optimized 3D model and assign real textures to the model according to the feature data set to obtain a textured 3D model;

[0053] Perform lighting settings on the textured 3D model, including setting ambient light and point light sources, to obtain an illuminated 3D model;

[0054] Generate a rendered image according to the illuminated 3D model to obtain a 3D rendered image of the GIS device;

[0055] Combine the 3D rendered image of the GIS device with the feature data set to obtain the 3D model of the GIS device.

[0056] As a preferred solution of the method for analyzing the operation faults of GIS devices based on 3D modeling according to the present invention, wherein: the step of collecting the real-time operation data of the GIS device and extracting fault features from the 3D model of the GIS device through the real-time operation data to obtain a fault feature vector includes:

[0057] Preprocess the real-time operation data, including data standardization and feature scaling, to obtain the processed real-time data;

[0058] Map the processed real-time data onto the 3D model of the GIS device to obtain a 3D model with data;

[0059] Perform model segmentation on the 3D model with data to obtain a segmented model dataset;

[0060] Construct a model graph based on the segmented model dataset to obtain model graph data;

[0061] Perform graph convolution operations on the model graph data to extract graph features and obtain graph feature data;

[0062] Input the graph feature data into a graph attention network to obtain attention-weighted graph features;

[0063] Perform pooling operations on the attention-weighted graph features to obtain a pooled feature tensor;

[0064] Input the pooled feature tensor into a fully connected layer for feature transformation to obtain a potential fault feature vector;

[0065] Perform normalization on the potential fault feature vector to obtain a normalized fault feature vector;

[0066] Input the normalized fault feature vector into a classification layer to obtain the fault feature vector.

[0067] As a preferred solution of the method for analyzing the operation faults of GIS devices based on 3D modeling according to the present invention, wherein:

[0068] The step of inputting the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data includes:

[0069] Construct an input tensor based on the fault feature vector to obtain input tensor data;

[0070] Input the input tensor data into a convolutional neural network model to obtain a convolutional feature map;

[0071] Perform max pooling operations on the convolutional feature map to obtain a pooled feature map;

[0072] Flatten the pooled feature map into a one-dimensional vector to obtain a one-dimensional feature vector;

[0073] Input the one-dimensional feature vector into a long short-term memory network to obtain temporal sequence feature data;

[0074] Input the timing feature data into the attention layer for feature weighting to obtain weighted feature data;

[0075] Input the weighted feature data into the fully connected layer for non - linear transformation to obtain the fault probability distribution;

[0076] Classify according to the fault probability distribution to obtain the classification result;

[0077] Map the classification result to a fault code to obtain fault code data;

[0078] Combine the fault code data with the corresponding fault description information to obtain the fault analysis data.

[0079] A GIS device operation fault analysis system based on 3D modeling, which includes:

[0080] An acquisition module, which is used to perform data cleaning on the acquired GIS device operation status data, environmental data, and historical fault data to obtain a cleaned data set;

[0081] A denoising module, which is used to perform data denoising on the cleaned data set to obtain a denoised data set;

[0082] A filling module, which is used to fill in the missing values of the denoised data set through a variational auto - encoder to obtain a filled data set;

[0083] A modeling module, which is used to perform 3D modeling according to the filled data set to obtain a 3D model of the GIS device;

[0084] An acquisition module, which is used to acquire the real - time operation data of the GIS device and extract fault features from the 3D model of the GIS device through the real - time operation data to obtain a fault feature vector;

[0085] An analysis module, which is used to input the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data.

[0086] A preferred solution of the GIS device operation fault analysis system based on 3D modeling, where: the acquisition module further includes:

[0087] Used to perform data cleaning on the acquired GIS device operation status data, environmental data, and historical fault data to obtain a cleaned data set;

[0088] Used to perform data denoising on the cleaned data set to obtain a denoised data set;

[0089] Used to fill in the missing values of the denoised data set through a variational auto - encoder to obtain a filled data set;

[0090] For performing three-dimensional modeling based on the filled dataset to obtain a three-dimensional model of the GIS device;

[0091] For collecting real-time operation data of the GIS device and extracting fault features from the three-dimensional model of the GIS device through the real-time operation data to obtain a fault feature vector;

[0092] For inputting the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data.

[0093] A GIS device operation fault analysis system based on three-dimensional modeling, characterized in that the analysis module further includes:

[0094] For constructing an input tensor based on the fault feature vector to obtain input tensor data;

[0095] For inputting the input tensor data into a convolutional neural network model to obtain a convolutional feature map;

[0096] For performing a max pooling operation on the convolutional feature map to obtain a pooled feature map;

[0097] For flattening the pooled feature map into a one-dimensional vector to obtain a one-dimensional feature vector;

[0098] For inputting the one-dimensional feature vector into a long short-term memory network to obtain temporal feature data;

[0099] For inputting the temporal feature data into an attention layer to perform feature weighting to obtain weighted feature data;

[0100] For inputting the weighted feature data into a fully connected layer to perform a non-linear transformation to obtain a fault probability distribution;

[0101] For classifying according to the fault probability distribution to obtain a classification result;

[0102] For mapping the classification result to a fault code to obtain fault code data;

[0103] For combining the fault code data with corresponding fault description information to obtain fault analysis data.

[0104] Beneficial effects of the present invention: In the technical solution provided by the present invention, by performing data cleaning on the collected GIS equipment operating status data, environmental data and historical fault data, the noise and outliers in the data are effectively removed, ensuring the quality and reliability of the data. This step includes sub-steps such as outlier detection and processing, data fusion, feature engineering, label data set construction, data segmentation, classification model training and optimization, and each sub-step is carefully designed to maximize the effect and efficiency of data cleaning. Next, the data is denoised by a Gaussian mixture model to effectively remove outliers in the data, and the purity of the data is further improved by principal component analysis and kernel discriminant analysis to ensure that the data processed subsequently is more accurate and reliable. The multi-level processing method of this step enables the denoised data set to not only retain the main features of the data, but also effectively reduce the interference of noise on the analysis results. Subsequently, the missing values ​​of the denoised data set are filled by a variational autoencoder, further enhancing the integrity and accuracy of the data. The variational autoencoder maps the standardized data set to the latent space through the dual processing of the encoder and decoder networks, and reconstructs the data by sampling latent variables, ultimately achieving the filling of missing values. The comprehensive calculation and optimization of reconstruction loss and KL divergence loss ensure the quality of the filled data set, so that the data still maintains a high degree of consistency and integrity in the filled state. On this basis, three-dimensional modeling is carried out according to the filled data set, and a high-precision three-dimensional model of GIS equipment is constructed through steps such as feature extraction, point cloud data construction, point cloud filtering, point cloud segmentation, three-dimensional reconstruction, mesh optimization, texture mapping and lighting setting. This model not only has real physical properties, but also can intuitively display the structure and status of the equipment through rendered images, providing a solid foundation for fault analysis. The collection and processing of real-time operation data is an important part of fault analysis. By preprocessing the real-time operation data, it is mapped to the three-dimensional model, and model segmentation, graph convolution operation, graph attention network weighting, pooling operation and feature transformation are performed to extract the potential fault feature vector. This process makes full use of the combination of real-time data and three-dimensional models, making the extraction of fault features more accurate and comprehensive. After the fault feature vector is extracted, the fault analysis is further performed through the deep neural network model. The combination of convolutional neural network, long short-term memory network and attention mechanism fully explores the spatiotemporal correlation of fault features and ensures the accuracy and real-time performance of fault analysis. Finally, the fault feature vector undergoes multi-level feature transformation and classification to generate fault probability distribution and fault code data, and combined with the corresponding fault description information, the fault analysis data is obtained. This series of technical features and steps, through the organic combination of deep learning and 3D modeling, realizes the efficient analysis and prediction of GIS equipment operation faults, significantly improves the real-time, accuracy and comprehensiveness of fault detection, and solves the problem that traditional fault detection methods are difficult to accurately identify and predict faults in complex environments. Brief Description of the Drawings

[0105] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0106] Figure 1 It is a flowchart of a method for analyzing the operation faults of GIS devices based on 3D modeling in an embodiment of the present invention;

[0107] Figure 2 It is a schematic diagram of a system for analyzing the operation faults of GIS devices based on 3D modeling in an embodiment of the present invention. Detailed Embodiments

[0108] To make the above objects, features, and advantages of the present invention more understandable, the following will provide a detailed description of the specific embodiments of the present invention in conjunction with the drawings in the specification.

[0109] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0110] Embodiment 1

[0111] Refer to Figure 1 , Figure 1 which is a flowchart of a method for analyzing the operation faults of GIS devices based on 3D modeling in an embodiment of the present invention. As Figure 1 shown, it includes the following steps:

[0112] S101. Perform data cleaning on the collected GIS device operation status data, environmental data, and historical fault data to obtain a cleaned data set;

[0113] S102. Denoise the cleaned data set to obtain a denoised data set;

[0114] S103. Fill in the missing values in the denoised data set through a variational autoencoder to obtain a filled data set;

[0115] S104. Perform 3D modeling based on the filled data set to obtain a 3D model of the GIS device;

[0116] S105. Collect the real-time operation data of the GIS device, and extract the fault features from the 3D model of the GIS device through the real-time operation data to obtain a fault feature vector;

[0117] S106. Input the fault feature vector into the deep neural network model for fault analysis to obtain fault analysis data.

[0118] It should be noted that data cleaning and processing of the collected GIS device operation status data, environmental data, and historical fault data is the basis of the entire method. Specifically, data cleaning includes detecting and processing outliers in the operation status data. By identifying and removing values that deviate significantly from the normal range, the validity and accuracy of the data are ensured. For example, during the operation of a certain GIS device, if a sudden extremely high value appears in the reading of the temperature sensor, this may be a misreading caused by a sensor failure. Through the outlier detection algorithm, this anomaly can be identified and removed or replaced with a reasonable value. Next, the processed operation status data is fused with the environmental data, combining information from multiple data sources to form a more comprehensive dataset. After data fusion, feature engineering is carried out to extract key feature vectors, which can better reflect the operation status of the GIS device. For example, the features that can best reflect the device health status can be extracted from various sensor data such as temperature, humidity, current, and voltage. After completing feature engineering, a labeled dataset is constructed using these feature vectors and historical fault data, corresponding each feature vector to the corresponding fault type. Next, the labeled dataset is split into a training dataset and a validation dataset. By training a classification model based on the training dataset, such as a random forest or support vector machine model, a fault classification system can be initially established. The validation dataset is used to evaluate the classification model, and the model is optimized according to the evaluation results to ensure its high accuracy and robustness. The optimized classification model is applied to the feature vector dataset to classify the data, obtaining classified data, and cleaning it to finally form a cleaned dataset. After obtaining the cleaned dataset, data denoising processing is carried out next. By constructing a Gaussian mixture model, anomaly detection is performed on each data point, and its anomaly score is calculated. An anomaly threshold is set, and the data points with anomaly scores higher than the threshold are marked as anomaly points. The principal component loading matrix is extracted through principal component analysis, and projection transformation and kernel discriminant analysis are performed on the labeled dataset to finally remove the anomaly points and obtain a denoised dataset. For example, when analyzing the vibration data of a GIS device, it may be found that some data points are abnormal due to external interference. Through the above process, these abnormal points can be effectively identified and removed to ensure the purity of the data. Based on the denoised dataset, variational autoencoders are used to fill in the missing values. Variational autoencoders construct a neural network model containing an encoder and a decoder, map the data to the latent space, sample latent variables from it, and reconstruct the data through the decoder. By calculating the reconstruction loss and KL divergence loss, the variational autoencoder model is optimized to accurately fill in the missing values. For example, some sensor data is missing due to sensor failures. Through the processing of variational autoencoders, reasonable filling values can be generated to ensure the integrity and consistency of the data.Next, perform 3D modeling based on the filled dataset, construct 3D point cloud data through feature extraction, and perform point cloud filtering and segmentation to remove outliers and noise, extract key point cloud clusters, and finally complete 3D reconstruction and mesh optimization to obtain a 3D model of the GIS device with real texture and lighting effects. For example, through 3D reconstruction of the data from multiple sensors of the GIS device, the internal structure and operating status of the device can be intuitively displayed, providing important references for fault analysis. The acquisition and processing of real-time operation data are the key links of fault analysis. Through preprocessing, the real-time data is standardized and mapped onto the 3D model, and model segmentation and graph convolution operations are performed to extract graph feature data. Through graph attention network weighted and pooling operations, a pooled feature tensor is obtained. The pooled feature tensor is input into a fully connected layer for feature transformation to obtain a potential fault feature vector. In the practical application of GIS devices, for example, when the temperature of a certain part of the device suddenly rises, combined with the temperature distribution map of the 3D model, the fault location can be accurately located and relevant features can be extracted. Finally, the fault feature vector is input into a deep neural network model for fault analysis. Feature maps are extracted through a convolutional neural network, temporal features are analyzed through a long short-term memory network, features are weighted by an attention mechanism, and a fully connected layer performs a non-linear transformation to generate a fault probability distribution, and finally the fault code is classified. The fault code is combined with the description information to generate fault analysis data. For example, through in-depth learning analysis of the fault feature vector of a certain device, the possible fault types of the device, such as arc fault or ground fault, can be accurately identified, and detailed fault descriptions and recommended handling measures can be provided. This method not only improves the accuracy and real-time performance of fault detection, but also provides intuitive and detailed fault analysis results for maintenance personnel, greatly improving the maintenance and management efficiency of GIS devices.

[0119] By executing the above steps, the noise and outliers in the data are effectively removed by cleaning the collected GIS equipment operation status data, environmental data and historical fault data, ensuring the quality and reliability of the data. This step includes sub-steps such as outlier detection and processing, data fusion, feature engineering, label data set construction, data segmentation, classification model training and optimization. Each sub-step has been carefully designed to maximize the effect and efficiency of data cleaning. Next, the data is denoised by using the Gaussian mixture model to effectively remove outliers in the data, and the purity of the data is further improved by principal component analysis and kernel discriminant analysis to ensure that the data processed later is more accurate and reliable. The multi-level processing method of this step not only retains the main features of the data in the denoised data set, but also effectively reduces the interference of noise on the analysis results. Subsequently, the missing values ​​of the denoised data set are filled by the variational autoencoder, which further enhances the integrity and accuracy of the data. The variational autoencoder maps the standardized data set to the latent space through the dual processing of the encoder and decoder networks, and reconstructs the data by sampling latent variables, finally achieving the filling of missing values. The comprehensive calculation and optimization of reconstruction loss and KL divergence loss ensure the quality of the filled data set, so that the data still maintains a high degree of consistency and integrity in the filled state. On this basis, three-dimensional modeling is carried out according to the filled data set, and a high-precision three-dimensional model of GIS equipment is constructed through steps such as feature extraction, point cloud data construction, point cloud filtering, point cloud segmentation, three-dimensional reconstruction, mesh optimization, texture mapping and lighting setting. This model not only has real physical properties, but also can intuitively display the structure and status of the equipment through rendered images, providing a solid foundation for fault analysis. The collection and processing of real-time operation data is an important part of fault analysis. By preprocessing the real-time operation data, it is mapped to the three-dimensional model, and model segmentation, graph convolution operation, graph attention network weighting, pooling operation and feature transformation are performed to extract the potential fault feature vector. This process makes full use of the combination of real-time data and three-dimensional models, making the extraction of fault features more accurate and comprehensive. After the fault feature vector is extracted, the fault analysis is further performed through the deep neural network model. The combination of convolutional neural network, long short-term memory network and attention mechanism fully explores the spatiotemporal correlation of fault features and ensures the accuracy and real-time performance of fault analysis. Finally, the fault feature vector undergoes multi-level feature transformation and classification to generate fault probability distribution and fault code data, and combined with the corresponding fault description information, the fault analysis data is obtained. This series of technical features and steps, through the organic combination of deep learning and 3D modeling, realizes the efficient analysis and prediction of GIS equipment operation faults, significantly improves the real-time, accuracy and comprehensiveness of fault detection, and solves the problem that traditional fault detection methods are difficult to accurately identify and predict faults in complex environments.

[0120] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0121] (1) Detect and process outliers in the GIS device operation status data to obtain the processed operation status data;

[0122] (2) Perform data fusion based on the processed operation status data and environmental data to obtain the fusion data;

[0123] (3) Perform feature engineering on the fusion data, extract key feature vectors, and obtain the feature vector dataset;

[0124] (4) Construct a label dataset based on the feature vector dataset and historical fault data;

[0125] (5) Split the label dataset to obtain the training dataset and the validation dataset;

[0126] (6) Train a classification model based on the training dataset to obtain the trained classification model;

[0127] (7) Evaluate the trained classification model using the validation dataset to obtain the evaluation result;

[0128] (8) Optimize the trained classification model according to the evaluation result to obtain the optimized classification model;

[0129] (9) Apply the optimized classification model to the feature vector dataset to classify the data and obtain the classification data;

[0130] (10) Clean the classification data to obtain the cleaned dataset.

[0131] Specifically, it is first necessary to detect and process outliers in the GIS device operation status data. During the operation of the GIS device, sensors may generate outliers for various reasons, and these outliers will affect subsequent data processing and analysis. Therefore, it is necessary to identify these outliers through outlier detection algorithms. There are various methods for outlier detection, such as the 3σ principle based on statistics, the box plot method, or the isolation forest algorithm based on machine learning. For example, when the data of a certain current sensor fluctuates within a certain range during normal operation, but at a certain moment, the data of this sensor suddenly soars to an extremely high value, significantly exceeding the normal range, it can be identified as an outlier through the isolation forest algorithm. After identifying the outliers, methods such as deletion, interpolation, or replacement can be used to process them, so as to obtain the processed operation status data. Next, data fusion is performed based on the processed operation status data and environmental data to form a more comprehensive data set. The GIS device operation status data includes parameters such as voltage, current, and temperature, while the environmental data includes external temperature, humidity, etc. The fusion of these data can more comprehensively reflect the operation status of the device. Data fusion can adopt methods such as weighted average, principal component analysis (PCA), or feature fusion. For example, in a high-temperature environment in summer, the temperature data of the device may be affected by the environmental temperature. Through the data fusion method, the temperature change inside the device can be more accurately separated, so as to obtain a more real operation status of the device. After data fusion, feature engineering is performed on the fused data to extract key feature vectors. Feature engineering is to transform the original data into features that can better reflect the essence of the problem, thereby improving the performance of the model. Feature extraction can adopt statistical features, frequency domain features, or time series-based features. For example, statistical features such as mean, standard deviation, and kurtosis can be extracted from the current data, frequency components can be extracted from the frequency domain data, and lag features can be extracted from the time series data. Through feature engineering, the original data is transformed into feature vectors to obtain a feature vector data set. Then, a label data set is constructed based on the feature vector data set and historical fault data. The historical fault data records the fault conditions of the GIS device in different operation states and can be used as label data to be paired with the feature vectors. When constructing the label data set, it is necessary to ensure that each feature vector corresponds to a clear fault label. For example, a certain feature vector represents the operation state of the device under high temperature and high humidity conditions. If an arc fault has occurred in the device in this state, the label of this feature vector is arc fault. After the label data set is constructed, it is divided into a training data set and a validation data set. Data division usually adopts the method of random division, dividing the data set into a training set and a validation set. The training set is used for model training, and the validation set is used for model evaluation. When dividing the data, it is necessary to ensure that the distributions of the training set and the validation set are consistent to improve the credibility of the model evaluation results.A classification model can be trained based on a training dataset, and classification algorithms such as random forest, support vector machine (SVM), or deep neural network (DNN) can be selected. Through multiple iterative trainings of the training dataset, the model gradually learns the patterns and rules in the data. For example, using the random forest algorithm, multiple decision tree models can be constructed and their results can be voted on to improve the accuracy and stability of classification. After the model training is completed, the trained classification model is evaluated using a validation dataset to obtain an evaluation result. Evaluation metrics can include accuracy, precision, recall, and F1 value, etc. For example, a certain evaluation result shows that the accuracy of the classification model is 90%, indicating that the model can correctly classify the fault type in 90% of the cases. According to the evaluation result, the model can be optimized, such as adjusting the model parameters, adopting a more complex algorithm, or increasing the amount of training data, etc., to obtain an optimized classification model. The optimized classification model is applied to the feature vector dataset to classify the data and obtain classification data. For example, through the optimized random forest model, a certain feature vector can be classified as an arc fault, a ground fault, or other fault types. The classification data represents the fault type corresponding to each feature vector, providing a basis for subsequent data cleaning and fault analysis. Finally, the classification data is cleaned to obtain a cleaned dataset. The purpose of data cleaning is to remove noise and outliers in the classification data and improve the quality and reliability of the data. For example, some misclassification results may be generated during the classification process, and these misclassified data can be removed through data cleaning to obtain a more pure cleaned dataset.

[0132] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0133] (1) Construct a Gaussian mixture model based on the cleaned dataset to obtain Gaussian mixture model parameters;

[0134] (2) Perform anomaly detection on each data point in the cleaned dataset, calculate its anomaly score according to the Gaussian mixture model parameters, and obtain an anomaly score set;

[0135] (3) Set an anomaly threshold according to the anomaly score set, mark the data points with anomaly scores higher than the anomaly threshold as anomaly points, and obtain a marked dataset;

[0136] (4) Perform principal component analysis on the marked dataset, extract the principal component loading matrix, and obtain the principal component loading matrix;

[0137] (5) Perform projection transformation on the marked dataset according to the principal component loading matrix to obtain a projection dataset;

[0138] (6) Perform kernel discriminant analysis on the projection dataset, construct a discriminant model, and obtain the discriminant model;

[0139] (7) Use the discrimination model to discriminate the abnormal points in the projection data set to obtain the discrimination result of the abnormal points;

[0140] (8) According to the discrimination result of the abnormal points, remove the data discriminated as abnormal points from the projection data set to obtain a denoised projection data set;

[0141] (9) Perform back-projection transformation on the denoised projection data set to obtain a back-projection data set;

[0142] (10) Perform data format conversion on the back-projection data set to obtain a denoised data set.

[0143] It should be noted that a Gaussian mixture model is constructed for the cleaned dataset. The Gaussian mixture model (GMM) is a probabilistic model for clustering and density estimation. By assuming that the data consists of multiple Gaussian distributions, it can capture the latent structure of the data. When constructing the GMM, it is necessary to determine the parameters of the model, including the mean, covariance matrix, and mixing coefficients of each Gaussian distribution. The expectation-maximization (EM) algorithm can be used for parameter estimation, and the model parameters are repeatedly optimized until convergence. For example, for the current data of GIS equipment, the parameters of each Gaussian distribution can be obtained through GMM modeling, reflecting different patterns of the current data. After obtaining the Gaussian mixture model parameters, next, anomaly detection is performed on each data point in the cleaned dataset. According to the GMM parameters, the anomaly score of each data point can be calculated. The calculation method of the anomaly score is based on the probability density values of the data point under each Gaussian distribution. The lower the probability density value, the more likely the data point is an outlier. For example, the calculation result of the anomaly score of a certain current data point shows that its probability density values under all Gaussian distributions are very low, indicating that this data point may be abnormal. The anomaly scores of all data points are aggregated to form an anomaly score set. According to the anomaly score set, an anomaly threshold is set to distinguish normal points from abnormal points. The setting of the threshold can be based on experience or determined by statistical methods. For example, the 99th percentile of the anomaly score set is taken as the threshold. The data points with anomaly scores higher than the anomaly threshold are marked as abnormal points, and the marked dataset is obtained. The marked dataset contains the marking information of normal points and abnormal points, providing a basis for subsequent processing. Next, principal component analysis (PCA) is performed on the marked dataset to extract the principal component loading matrix. PCA is a dimensionality reduction method that projects the original data into a low-dimensional space through a linear transformation, so that the projected data can retain the variance information of the original data as much as possible. The principal component loading matrix is the core of the PCA transformation, indicating the directions of each principal component in the original feature space. For example, for multi-dimensional current data, PCA can extract the main change patterns, reduce the dimension of the data, and improve the calculation efficiency. According to the principal component loading matrix, a projection transformation is performed on the marked dataset to obtain a projected dataset. The projection transformation is to project the original data into the principal component space to obtain the dimension-reduced data. For example, for three-dimensional current data, through PCA, it can be projected into a two-dimensional or one-dimensional space, reducing the complexity of the data and facilitating subsequent analysis. Kernel discriminant analysis (KDA) is performed on the projected dataset to construct a discriminant model. KDA is a classification method that combines kernel methods and linear discriminant analysis. It maps the data into a high-dimensional space through a non-linear transformation, making the data of different classes more linearly separable in the high-dimensional space. The discriminant model is used to distinguish normal points from abnormal points. For example, through KDA, the normal current data and abnormal current data in the projected dataset can be separated, improving the accuracy of anomaly detection. The discriminant model is used to discriminate the abnormal points in the projected dataset to obtain the discriminant result of the abnormal points.The discrimination result indicates whether each data point is a normal point or an abnormal point. For example, through the KDA model, it can be determined whether a certain projected current data point is abnormal, and the discrimination result of the abnormal point is obtained. According to the discrimination result of the abnormal point, the data determined to be abnormal points is removed from the projected data set, and a denoised projected data set is obtained. After removing the abnormal points, only normal points remain in the projected data set, and the purity and reliability of the data are greatly improved. For example, by removing abnormal current data points, a more accurate and reliable current data set is obtained. An inverse projection transformation is performed on the denoised projected data set to restore it to the original feature space, and an inverse projection data set is obtained. The inverse projection transformation remaps the dimension-reduced data back to the original high-dimensional space. For example, two-dimensional current data is restored to three-dimensional space through the inverse projection transformation to ensure the integrity of the data. Finally, a data format conversion is performed on the inverse projection data set to obtain a denoised data set. The data format conversion converts the processed data into a standard format for subsequent analysis and processing. For example, the inverse projected current data is converted into a standard time series format for further fault analysis and modeling.

[0144] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0145] (1) Perform standardization processing on the denoised data set to obtain a standardized data set;

[0146] (2) Construct a variational autoencoder model based on the standardized data set, including an encoder and a decoder network, to obtain a variational autoencoder model;

[0147] (3) Apply the encoder network in the variational autoencoder model to the standardized data set to obtain a latent space distribution;

[0148] (4) Sample latent variables from the latent space distribution to obtain a set of latent variables;

[0149] (5) Input the set of latent variables into the decoder network of the variational autoencoder model to obtain a reconstructed data set;

[0150] (6) Calculate the reconstruction loss based on the standardized data set and the reconstructed data set;

[0151] (7) Calculate the KL divergence loss based on the latent space distribution;

[0152] (8) Add the reconstruction loss and the KL divergence loss to obtain the total loss;

[0153] (9) Update the parameters of the variational autoencoder model based on the total loss to obtain an updated variational autoencoder model;

[0154] (10) Use the updated variational autoencoder model to fill in the missing values in the denoised data set to obtain a filled data set.

[0155] Specifically, the denoised dataset is standardized. The purpose of standardization is to scale different features of the data to the same scale to eliminate the dimensional differences between features. Standardization usually adopts the mean-standard deviation standardization method, subtracting the mean value of each feature and then dividing by the standard deviation, so that the processed data has a distribution with a mean of 0 and a standard deviation of 1. For example, when processing data such as current and temperature of GIS equipment, standardization can eliminate the influence brought by different dimensions, making subsequent modeling more stable and effective. After obtaining the standardized dataset, a variational autoencoder (VAE) model is constructed based on these data. VAE is a generative model that can generate data and fill in missing values by learning the latent distribution of the data. The VAE model consists of an encoder and a decoder network. The encoder maps the input data to the latent space, and the decoder reconstructs the variables in the latent space into the original data. The encoder and decoder are usually implemented using multi-layer neural networks. For example, when processing the operation data of GIS equipment, the encoder network can compress the high-dimensional sensor data into a low-dimensional latent space, and the decoder reconstructs the high-dimensional data based on these low-dimensional representations. Applying the encoder network in the VAE model to the standardized dataset can obtain the latent space distribution. The latent space distribution is the probability distribution of the latent variables output by the encoder network, usually assumed to be a Gaussian distribution. In this process, the encoder network maps the data to the latent space by learning the features of the data. For example, after the current and temperature data of GIS equipment are processed by the encoder network, the resulting latent variable distribution reflects the latent structure and relationship of these data. Sampling latent variables from the latent space distribution can obtain a set of latent variables. The sampling process is usually achieved through the reparameterization trick, introducing random noise into the generation process of latent variables, so that the model can be trained by the gradient descent method. For example, sampling latent variables from the Gaussian distribution in the latent space can obtain multiple latent variables, which can be used for data reconstruction and generation. Inputting the set of latent variables into the decoder network of the VAE model can obtain the reconstructed dataset. The decoder network maps the latent variables back to the high-dimensional space and reconstructs an approximation of the original data. For example, the current and temperature data generated by the decoder network based on the latent variables should be very close to the original data, thus realizing data reconstruction. Based on the standardized dataset and the reconstructed dataset, the reconstruction loss can be calculated. The reconstruction loss usually adopts the mean squared error (MSE) or cross-entropy loss to measure the difference between the reconstructed data and the original data. The smaller the reconstruction loss, the closer the data generated by the decoder is to the original data. For example, the size of the reconstruction loss can reflect the accuracy of the decoder when reconstructing the operation data of GIS equipment. At the same time, the KL divergence loss is calculated according to the latent space distribution. The KL divergence loss measures the difference between the latent distribution and the standard normal distribution, encouraging the latent distribution to be close to the standard normal distribution, making the model have better generative ability and robustness. Adding the reconstruction loss and the KL divergence loss together gives the total loss.The total loss comprehensively considers the accuracy of data reconstruction and the rationality of the potential distribution, which is the optimization objective of model training. Based on the total loss, the parameters of the VAE model are updated, and the parameters of the encoder and decoder networks are adjusted through the backpropagation algorithm, enabling the model to better reconstruct data and learn the potential distribution. The optimized VAE model has higher accuracy and reliability in reconstructing data and handling missing values. Using the updated VAE model, missing values in the denoised dataset can be filled. The denoised dataset is input into the encoder network to obtain the potential space distribution, and potential variables are sampled from it, and then the filling values are generated through the decoder network. For example, in the operation data of GIS devices, if the temperature data at certain moments is missing due to sensor failures, the VAE model can generate reasonable filling values through the learned potential distribution, thus obtaining a filled dataset.

[0156] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0157] (1) Extract features from the filled dataset to obtain a feature dataset;

[0158] (2) Construct three-dimensional point cloud data based on the feature dataset to obtain the original point cloud data;

[0159] (3) Perform point cloud filtering on the original point cloud data to remove outliers and obtain the filtered point cloud data;

[0160] (4) Perform point cloud segmentation on the filtered point cloud data to extract key point cloud clusters and obtain the segmented point cloud data;

[0161] (5) Perform three-dimensional reconstruction on the segmented point cloud data to construct a three-dimensional mesh model and obtain a rough three-dimensional model;

[0162] (6) Optimize the mesh of the rough three-dimensional model, including vertex refinement and patch subdivision, to obtain the optimized three-dimensional model;

[0163] (7) Perform texture mapping on the optimized three-dimensional model and assign real textures to the model according to the feature dataset to obtain a three-dimensional model with textures;

[0164] (8) Set the lighting for the three-dimensional model with textures, including setting ambient light and point light sources, to obtain an illuminated three-dimensional model;

[0165] (9) Generate a rendered image based on the illuminated three-dimensional model to obtain a three-dimensional rendered image of the GIS device;

[0166] (10) Combine the three-dimensional rendered image of the GIS device with the feature dataset to obtain a three-dimensional model of the GIS device.

[0167] Specifically, feature extraction is performed on the filled dataset. The purpose of feature extraction is to transform the original data into features that can better reflect the essence of the problem, thereby providing high-quality data support for subsequent 3D modeling. Feature extraction can be achieved through methods such as statistical analysis, signal processing, and machine learning. For example, for the operation data of GIS devices, key features such as temperature, current, and voltage can be extracted, which can accurately reflect the operation status and health condition of the devices. After obtaining the feature dataset, the next step is to construct 3D point cloud data based on the feature dataset. Point cloud data is a commonly used 3D data representation method, consisting of a large number of points, and each point has 3D coordinates and other attributes (such as color, intensity, etc.). By performing spatial transformation and mapping on the feature data, the original point cloud data can be generated. For example, when processing the spatial layout data of GIS devices, according to the physical size and position of the devices, the position information of temperature sensors, current sensors, etc. can be mapped into 3D space to form a complete original point cloud data. The original point cloud data often contains noise and outliers, which will affect the accuracy of subsequent 3D modeling. Therefore, point cloud filtering is required. The purpose of point cloud filtering is to remove noise points and outliers in the point cloud data and improve the purity of the data. Commonly used point cloud filtering methods include statistical filtering, radius filtering, and conditional filtering, etc. For example, through the statistical filtering method, the neighborhood density of each point can be calculated, and the points with density lower than a certain threshold are regarded as outliers and removed, thereby obtaining the filtered point cloud data. The filtered point cloud data still contains a large number of points, and these points need to be further segmented to extract the key point cloud clusters. The purpose of point cloud segmentation is to divide the point cloud data into multiple independent parts, and each part corresponds to an object or a region. Commonly used point cloud segmentation methods include region-growing-based segmentation, clustering-based segmentation, and graph-based segmentation, etc. For example, when processing the point cloud data of GIS devices, the region-growing algorithm can be used to divide the point cloud data into different components according to the spatial position and attributes of the points, such as circuit breakers, busbars, and disconnectors, etc., thereby obtaining the segmented point cloud data. Next, 3D reconstruction is performed on the segmented point cloud data to construct a 3D mesh model. The purpose of 3D reconstruction is to convert the point cloud data into a 3D mesh representation to better display the shape and structure of the object. Commonly used 3D reconstruction methods include Delaunay triangulation, Poisson reconstruction, and surface reconstruction, etc. For example, through the Poisson reconstruction method, a continuous 3D surface can be generated according to the segmented point cloud data, thereby obtaining a rough 3D model. The rough 3D model often contains some irregular and rough parts, and mesh optimization is required. The purpose of mesh optimization is to refine the vertices and faces of the 3D model to make the model smoother and more detailed. Commonly used mesh optimization methods include vertex refinement, face subdivision, and mesh simplification, etc.For example, through the vertex refinement method, the number of vertices of a 3D model can be increased to make the model surface smoother; through the patch subdivision method, the patch structure of the 3D model can be refined to improve the detail expressiveness of the model, thus obtaining an optimized 3D model. The optimized 3D model needs to be texture mapped to endow the model with real textures according to the feature dataset. The purpose of texture mapping is to map a 2D image onto the surface of a 3D model to make the model look more realistic. Commonly used texture mapping methods include UV mapping, projection mapping, and procedural textures, etc. For example, through the UV mapping method, the color and intensity information in the feature data can be mapped onto the surface of the 3D model to make the model have real textures, thus obtaining a textured 3D model. To make the textured 3D model look more real, lighting settings are also required. The purpose of lighting settings is to add ambient light and point light sources to the 3D model to make the model have shadow and highlight effects. Commonly used lighting setting methods include ambient light setting, point light source setting, and lighting models, etc. For example, by setting ambient light and point light sources, the lighting effects in reality can be simulated to make the 3D model more realistic, thus obtaining an illuminated 3D model. Generating a rendered image based on the illuminated 3D model is a process of converting the 3D model into a 2D image. The generation of the rendered image is usually achieved through techniques such as ray tracing, rasterization, and global illumination. For example, through the ray tracing method, the propagation and reflection of light in 3D space can be simulated to generate a high-quality 2D rendered image, thus obtaining a 3D rendered image of the GIS device.

[0168] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0169] (1) Preprocess the real-time operation data, including data standardization and feature scaling, to obtain processed real-time data;

[0170] (2) Map the processed real-time data onto the 3D model of the GIS device to obtain a 3D model with data;

[0171] (3) Segment the 3D model with data to obtain a segmented model dataset;

[0172] (4) Construct a model graph based on the segmented model dataset to obtain model graph data;

[0173] (5) Perform graph convolution operations on the model graph data to extract graph features and obtain graph feature data;

[0174] (6) Input the graph feature data into the graph attention network to obtain attention-weighted graph features;

[0175] (7) Perform pooling operations on the attention-weighted graph features to obtain a pooled feature tensor;

[0176] (8) Input the pooled feature tensor into the fully connected layer for feature transformation to obtain the potential fault feature vector;

[0177] (9) Normalize the potential fault feature vector to obtain the normalized fault feature vector;

[0178] (10) Input the normalized fault feature vector into the classification layer to obtain the fault feature vector.

[0179] Specifically, preprocess the real-time operation data. The preprocessing steps include data standardization and feature scaling, aiming to transform data of different scales into the same scale to eliminate the dimensional difference and improve the consistency and comparability of the data. Data standardization usually adopts the mean-standard deviation standardization method, subtracting the mean value of each feature from its value and then dividing by its standard deviation, so that the processed data has a distribution with a mean of 0 and a standard deviation of 1. For example, when processing the current, temperature, and humidity data of GIS devices, through standardization, these data can be transformed into the same scale, making them equally important in subsequent processing and analysis. After completing data standardization and feature scaling, map the processed real-time data onto the 3D model of the GIS device to obtain a 3D model with data. This step combines real-time data with the 3D model, so that each data point has a clear physical position in 3D space. For example, when processing the temperature sensor data of a GIS device, the temperature value of each sensor can be mapped to the corresponding position in the 3D model of the device, so that the 3D model can not only display the geometric structure of the device but also intuitively show the temperature distribution. Next, perform model segmentation on the 3D model with data to obtain a segmented model dataset. The purpose of model segmentation is to divide the 3D model into multiple independent parts, and each part corresponds to a specific component or area. Common model segmentation methods include segmentation based on spatial position and segmentation based on attributes, etc. For example, in the 3D model of a GIS device, the model can be segmented into different components such as circuit breakers, busbars, and disconnectors according to the structure of the device, so as to obtain a segmented model dataset. Construct a model graph based on the segmented model dataset to obtain model graph data. A model graph is a method of representing a 3D model as a graph structure, where each node represents a segmented model part and each edge represents the connection relationship between parts. For example, a circuit breaker and a busbar can be represented as nodes in the model graph, and the connection between them is represented as an edge, thus constructing the model graph of the entire GIS device. Perform graph convolution operations on the model graph data to extract graph features and obtain graph feature data. A graph convolutional network (GCN) is a neural network specifically designed to process graph-structured data. By performing convolution operations on the graph, it can effectively extract the feature information of the graph. For example, through GCN, the connection features and spatial relationship features between various components of a GIS device can be extracted, and these features can reflect the operating state and potential faults of the device. Input the graph feature data into a graph attention network to obtain attention-weighted graph features. A graph attention network (GAT) is a network that adds an attention mechanism on the basis of GCN. By calculating the attention weights of each node, it can dynamically adjust the contribution of different nodes to the features, so as to more accurately extract graph features. For example, when processing the graph feature data of a GIS device, GAT can assign different weights according to the importance of each component, so that the features of key components receive more attention, thus obtaining more accurate fault features.Perform a pooling operation on the attention-weighted graph features to obtain a pooled feature tensor. The purpose of the pooling operation is to reduce the dimension and compress the feature data, retain the key features, and remove redundant information. Common pooling methods include max pooling and average pooling, etc. For example, through max pooling, the maximum value can be selected in each feature dimension, so as to obtain a representative pooled feature tensor, and these tensors can better reflect the operating state of the GIS device. Input the pooled feature tensor into the fully connected layer for feature transformation to obtain the potential fault feature vector. The fully connected layer is a neural network layer that maps the input features to the output feature space through linear transformation and non-linear activation functions. For example, through the fully connected layer, the pooled feature tensor can be converted into potential fault feature vectors, and these vectors can reflect the potential fault modes and features of the device. Perform normalization processing on the potential fault feature vector to obtain the normalized fault feature vector. The purpose of normalization processing is to scale the values of the feature vector to a specific range, usually between 0 and 1, to improve the stability and comparability of the features. For example, through normalization processing, the values of the potential fault feature vector can be restricted between 0 and 1, so that different features have the same dimension, which is convenient for subsequent classification and analysis. Input the normalized fault feature vector into the classification layer to obtain the fault feature vector. The classification layer is a neural network layer for fault classification, and through the trained classification model, the input feature vector can be mapped to the fault category. For example, in the fault detection of GIS devices, through the classification layer, the normalized fault feature vector can be classified into different fault types, such as arc fault, ground fault, etc., so as to obtain the fault feature vector. These fault feature vectors can provide the fault type and fault probability of the device, providing an important reference basis for the maintenance and management of the device.

[0180] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0181] (1) Construct an input tensor according to the fault feature vector to obtain the input tensor data;

[0182] (2) Input the input tensor data into the convolutional neural network model to obtain the convolutional feature map;

[0183] (3) Perform a max pooling operation on the convolutional feature map to obtain the pooled feature map;

[0184] (4) Flatten the pooled feature map into a one-dimensional vector to obtain the one-dimensional feature vector;

[0185] (5) Input the one-dimensional feature vector into the long short-term memory network to obtain the temporal feature data;

[0186] (6) Input the temporal feature data into the attention layer for feature weighting to obtain the weighted feature data;

[0187] (7) Input the weighted feature data into the fully connected layer for non-linear transformation to obtain the fault probability distribution;

[0188] (8) Classify according to the fault probability distribution to obtain the classification result;

[0189] (9) Map the classification result to a fault code to obtain the fault code data;

[0190] (10) Combine the fault code data with the corresponding fault description information to obtain the fault analysis data.

[0191] It should be noted that the input tensor is constructed based on the fault feature vector. The fault feature vector is extracted from the operation data of the GIS device, and these feature vectors need to be converted into a tensor form suitable for neural network processing. The input tensor is a multi-dimensional array that can represent the spatial structure and attributes of fault features. For example, if the fault feature vector includes multiple parameters such as temperature, current, voltage, etc., these parameters can be combined into a two-dimensional or three-dimensional tensor to represent the changes in fault features at different time points. The constructed input tensor data is input into a convolutional neural network (CNN) model. CNN is a deep learning model that is good at processing images and spatial data. Through convolutional operations, CNN can extract local features of the input data and gradually extract higher-level features through multiple layers of convolution. For example, when processing the fault features of the GIS device, CNN can extract the spatial relationships and change patterns in the feature vector through the convolutional layer, thereby obtaining a convolutional feature map. The maximum pooling operation is performed on the convolutional feature map. The purpose is to reduce the dimension and compress the feature map, retain key features, and remove redundant information. Maximum pooling selects the maximum value in the local area of the feature map, thereby reducing the size of the feature map and enhancing the robustness of the features. For example, through the maximum pooling operation, the most significant fault features in the convolutional feature map can be extracted to obtain a pooled feature map. The pooled feature map is flattened into a one-dimensional vector. The purpose is to convert the two-dimensional or multi-dimensional feature map into a one-dimensional feature vector for subsequent processing and analysis. The flattening operation unfolds the elements of the feature map row by row or column by column to form a long vector. For example, through the flattening operation, all the feature values in the pooled feature map can be combined into a one-dimensional feature vector for easy input into the subsequent neural network layer. The one-dimensional feature vector is input into a long short-term memory network (LSTM). LSTM is a recurrent neural network that is good at processing time series data and can capture long-term and short-term dependencies in the data. Through the LSTM network, temporal analysis can be performed on the input fault feature vector, and the time patterns and change trends in the feature data can be extracted. For example, when processing the time series fault data of the GIS device, LSTM can analyze the changes in the feature vector at different time points to obtain temporal feature data. The temporal feature data is input into the attention layer. The purpose of the attention layer is to weight the feature data, highlight key features, and suppress irrelevant features. The attention mechanism calculates the importance weights of each feature and weights the features according to the weights, thereby enhancing the model's attention to important features. For example, through the attention mechanism, the fault features of the GIS device can be weighted to highlight those features that are most important for fault identification, obtaining weighted feature data. The weighted feature data is input into the fully connected layer for non-linear transformation. The purpose is to map the weighted feature data to a fault probability distribution. The fully connected layer performs feature combination and transformation on the input data through linear transformation and non-linear activation functions, thereby obtaining a fault probability distribution.For example, through the fully connected layer, the weighted feature data can be converted into the probability distribution of various types of faults, reflecting the possibility of each fault. Classification is performed based on the fault probability distribution to obtain the classification result. The classification process is to take the fault type corresponding to the maximum probability in the fault probability distribution as the final classification result. For example, if in the fault probability distribution of a certain GIS device, the probability of arc fault is the largest, then it is classified as an arc fault to obtain the classification result. The classification result is mapped to a fault code to obtain the fault code data. The fault code is a concise fault identifier. By mapping the classification result to a predefined fault code, it is convenient to record and transmit fault information. For example, mapping the arc fault to the fault code "E01" and the ground fault to the fault code "G02" to form the fault code data. Finally, the fault code data is combined with the corresponding fault description information to obtain the fault analysis data. The fault description information includes a detailed explanation of the fault type, possible causes, and recommended handling measures. By combining the fault code with these description information, a complete fault analysis report can be generated. For example, for the fault code "E01", the corresponding description information can be provided: "Arc fault, possibly caused by insulation damage or too high voltage. It is recommended to check the insulation status and voltage level of the device." In this way, detailed fault analysis data can be provided for maintenance personnel to help them quickly locate and handle equipment faults.

[0192] An embodiment of the present invention also provides a GIS device operation fault analysis system based on 3D modeling. The GIS device operation fault analysis system based on 3D modeling specifically includes:

[0193] An acquisition module 201, configured to perform data cleaning processing on the acquired GIS device operation status data, environmental data, and historical fault data to obtain a cleaned data set;

[0194] A denoising module 202, configured to perform data denoising on the cleaned data set to obtain a denoised data set;

[0195] A filling module 203, configured to perform missing value filling on the denoised data set through a variational autoencoder to obtain a filled data set;

[0196] A modeling module 204, configured to perform 3D modeling based on the filled data set to obtain a 3D model of the GIS device;

[0197] An acquisition module 205, configured to acquire the real-time operation data of the GIS device, and extract fault features from the 3D model of the GIS device through the real-time operation data to obtain a fault feature vector;

[0198] An analysis module 206, configured to input the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data.

[0199] Through the collaborative work of the above-mentioned various modules, by performing data cleaning on the collected GIS device operation status data, environmental data, and historical fault data, noise and outliers in the data are effectively removed, ensuring the quality and reliability of the data. This step includes sub-steps such as outlier detection and handling, data fusion, feature engineering, labeled dataset construction, data splitting, classification model training and optimization, etc. Each sub-step is carefully designed to maximize the effect and efficiency of data cleaning. Next, data denoising is performed through a Gaussian mixture model to effectively eliminate outliers in the data, and the purity of the data is further improved through principal component analysis and kernel discriminant analysis, ensuring that the data for subsequent processing is more accurate and reliable. The multi-level processing method of this step enables the denoised dataset to not only retain the main features of the data but also effectively reduce the interference of noise on the analysis results. Subsequently, the variational autoencoder is used to fill in the missing values in the denoised dataset, further enhancing the integrity and accuracy of the data. The variational autoencoder maps the standardized dataset to the latent space through the dual processing of the encoder and decoder networks, and reconstructs the data by sampling latent variables, ultimately achieving the filling of missing values. The comprehensive calculation and optimization of the reconstruction loss and KL divergence loss ensure the quality of the filled dataset, enabling the data to still maintain a high degree of consistency and integrity after filling. On this basis, a three-dimensional model of the GIS device is constructed according to the filled dataset. Through steps such as feature extraction, point cloud data construction, point cloud filtering, point cloud segmentation, three-dimensional reconstruction, mesh optimization, texture mapping, and lighting settings, a high-precision three-dimensional model of the GIS device is constructed. This model not only has real physical characteristics but also can intuitively display the structure and status of the device through rendered images, providing a solid foundation for fault analysis. The acquisition and processing of real-time operation data are important links in fault analysis. Through the preprocessing of real-time operation data, it is mapped onto the three-dimensional model, and model segmentation, graph convolution operations, graph attention network weighting, pooling operations, and feature transformation are performed to extract potential fault feature vectors. This process makes full use of the combination of real-time data and the three-dimensional model, making the extraction of fault features more accurate and comprehensive. After the extraction of the fault feature vectors, further fault analysis is carried out through a deep neural network model. The combination of convolutional neural networks, long short-term memory networks, and attention mechanisms fully explores the spatio-temporal correlation of fault features, ensuring the accuracy and real-time nature of fault analysis. Finally, the fault feature vectors undergo multi-level feature transformation and classification to generate fault probability distribution and fault code data, and combined with the corresponding fault description information, fault analysis data is obtained. This series of technical features and steps, through the organic combination of deep learning and three-dimensional modeling, achieve the efficient analysis and prediction of the operation faults of GIS devices, significantly improving the real-time nature, accuracy, and comprehensiveness of fault detection, and solving the problem that traditional fault detection methods are difficult to accurately identify and predict faults in complex environments.

[0200] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A GIS equipment operation failure analysis method based on three-dimensional modeling, characterized by: include, Perform data cleaning on the collected GIS equipment operating status data, environmental data and historical fault data to obtain a cleaned data set; Performing data denoising on the cleaned data set to obtain a denoised data set; Filling missing values ​​in the denoised data set by a variational autoencoder to obtain a filled data set; Perform three-dimensional modeling according to the filled data set to obtain a three-dimensional model of the GIS equipment; Collecting real-time operation data of the GIS equipment, and extracting fault features of the three-dimensional model of the GIS equipment through the real-time operation data to obtain a fault feature vector; Inputting the fault feature vector into a deep neural network model to perform fault analysis to obtain fault analysis data; The step of filling missing values ​​of the denoised data set by using a variational autoencoder to obtain a filled data set comprises: Performing standardization processing on the denoised data set to obtain a standardized data set; Constructing a variational autoencoder model according to the standardized data set, including an encoder and a decoder network, to obtain a variational autoencoder model; Applying an encoder network in the variational autoencoder model to the standardized dataset to obtain a latent space distribution; Sampling latent variables from the latent space distribution to obtain a set of latent variables; Inputting the latent variable set into the decoder network of the variational autoencoder model to obtain a reconstructed data set; Calculate the reconstruction loss based on the standardized data set and the reconstructed data set; Calculate the KL divergence loss according to the latent space distribution; Add the reconstruction loss and the KL divergence loss to get the total loss; Based on the total loss, the variational autoencoder model is updated with parameters to obtain an updated variational autoencoder model; Filling missing values ​​in the denoised data set using the updated variational autoencoder model to obtain the filled data set; The step of collecting real-time operation data of the GIS device and extracting fault features of the three-dimensional model of the GIS device through the real-time operation data to obtain a fault feature vector includes: Preprocessing the real-time operation data, including data standardization and feature scaling, to obtain processed real-time data; Mapping the processed real-time data onto the three-dimensional model of the GIS device to obtain a three-dimensional model with data; Performing model segmentation on the three-dimensional model with data to obtain a segmented model data set; Construct a model graph according to the segmented model data set to obtain model graph data; Performing a graph convolution operation on the model graph data to extract graph features and obtain graph feature data; Inputting the graph feature data into a graph attention network to obtain attention-weighted graph features; Performing a pooling operation on the attention weighted graph features to obtain a pooled feature tensor; Inputting the pooled feature tensor into a fully connected layer to perform feature transformation to obtain a potential fault feature vector; Normalizing the potential fault feature vector to obtain a normalized fault feature vector; Inputting the normalized fault feature vector into a classification layer to obtain the fault feature vector; The step of inputting the fault feature vector into a deep neural network model for fault analysis to obtain fault analysis data comprises: Constructing an input tensor according to the fault feature vector to obtain input tensor data; Inputting the input tensor data into a convolutional neural network model to obtain a convolutional feature map; Performing a maximum pooling operation on the convolutional feature map to obtain a pooled feature map; Flattening the pooled feature map into a one-dimensional vector to obtain a one-dimensional feature vector; Inputting the one-dimensional feature vector into a long short-term memory network to obtain time series feature data; Input the time series feature data into the attention layer, perform feature weighting, and obtain weighted feature data; Inputting the weighted feature data into a fully connected layer, performing nonlinear transformation, and obtaining a fault probability distribution; performing classification according to the fault probability distribution, and obtaining a classification result; Mapping the classification result into a fault code to obtain fault code data; The fault code data is combined with corresponding fault description information to obtain the fault analysis data.

2. A GIS equipment operation failure analysis method based on three-dimensional modeling as claimed in claim 1, characterized in that: The step of performing data cleaning on the collected GIS equipment operating status data, environmental data and historical fault data to obtain a cleaned data set includes: Perform outlier detection and processing on the GIS equipment operation status data to obtain the processed operation status data; Performing data fusion according to the processed operating status data and environmental data to obtain fused data; Performing feature engineering on the fused data to extract key feature vectors to obtain a feature vector data set; Constructing a label data set according to the feature vector data set and the historical fault data; Performing data segmentation on the labeled data set to obtain a training data set and a verification data set; Training a classification model based on the training data set to obtain a trained classification model; Using the validation data set to evaluate the trained classification model to obtain an evaluation result; Optimizing the trained classification model according to the evaluation result to obtain an optimized classification model; Applying the optimized classification model to the feature vector data set to classify the data to obtain classified data; The classified data is cleaned to obtain a cleaned data set.

3. A GIS equipment operation failure analysis method based on three-dimensional modeling as claimed in claim 1, characterized in that: The step of performing data denoising on the cleaned data set to obtain a denoised data set includes: Constructing a Gaussian mixture model according to the cleaned data set to obtain Gaussian mixture model parameters; Performing anomaly detection on each data point in the cleaned data set, calculating its anomaly score according to the Gaussian mixture model parameters, and obtaining an anomaly score set; An anomaly threshold is set according to the anomaly score set, and data points with an anomaly score higher than the anomaly threshold are marked as anomalies, so as to obtain a marked data set; Performing principal component analysis on the labeled data set, extracting a principal component loading matrix, and obtaining a principal component loading matrix; Performing a projection transformation on the labeled data set according to the principal component loading matrix to obtain a projected data set; Performing a kernel discriminant analysis on the projected data set, constructing a discriminant model, and obtaining a discriminant model; Using the discriminant model to discriminate abnormal points in the projection data set to obtain abnormal point discrimination results; According to the outlier identification result, data identified as outliers are removed from the projection data set to obtain a denoised projection data set; Performing a back-projection transformation on the denoised projection data set to obtain a back-projection data set; The back-projection data set is converted into a data format to obtain the denoised data set.

4. A GIS equipment operation failure analysis method based on three-dimensional modeling as claimed in claim 1, characterized in that: The step of performing three-dimensional modeling according to the filled data set to obtain a three-dimensional model of the GIS device includes: Performing feature extraction on the filled data set to obtain a feature data set; Constructing three-dimensional point cloud data according to the feature data set to obtain original point cloud data; Performing point cloud filtering on the original point cloud data to remove outliers and obtain filtered point cloud data; Performing point cloud segmentation on the filtered point cloud data, extracting key point cloud clusters, and obtaining segmented point cloud data; Performing three-dimensional reconstruction on the segmented point cloud data, constructing a three-dimensional mesh model, and obtaining a rough three-dimensional model; Performing mesh optimization on the rough three-dimensional model, including vertex refinement and facet subdivision, to obtain an optimized three-dimensional model; Performing texture mapping on the optimized three-dimensional model, giving the model a real texture according to the feature data set, and obtaining a textured three-dimensional model; Performing lighting settings on the textured three-dimensional model, including setting ambient light and point light sources, to obtain an illuminated three-dimensional model; Generate a rendering image according to the lighting three-dimensional model to obtain a three-dimensional rendering image of the GIS device; The three-dimensional rendering image of the GIS device is combined with the feature data set to obtain the three-dimensional model of the GIS device.

5. A GIS equipment operation fault analysis system based on three-dimensional modeling, used to execute the GIS equipment operation fault analysis method based on three-dimensional modeling as claimed in any one of claims 1 to 4, characterized in that: include: The acquisition module is used to clean the collected GIS equipment operating status data, environmental data and historical fault data to obtain a cleaned data set; A denoising module, used for denoising the cleaned data set to obtain a denoised data set; A filling module, used for filling missing values ​​of the denoised data set by using a variational autoencoder to obtain a filled data set; A modeling module, used for performing three-dimensional modeling according to the filled data set to obtain a three-dimensional model of the GIS equipment; A collection module is used to collect real-time operation data of GIS equipment, and extract fault features of the three-dimensional model of the GIS equipment through the real-time operation data to obtain a fault feature vector; The analysis module is used to input the fault feature vector into a deep neural network model to perform fault analysis and obtain fault analysis data.

6. The GIS equipment operation fault analysis system based on three-dimensional modeling according to claim 5 is characterized in that: The acquisition module further comprises: It is used to clean the collected GIS equipment operation status data, environmental data and historical fault data to obtain a clean data set; Used to perform data denoising on the cleaned data set to obtain a denoised data set; Used to fill missing values ​​of the denoised data set through a variational autoencoder to obtain a filled data set; Used to perform three-dimensional modeling according to the filled data set to obtain a three-dimensional model of the GIS equipment; Used to collect real-time operation data of GIS equipment, and extract fault features of the three-dimensional model of the GIS equipment through the real-time operation data to obtain a fault feature vector; It is used to input the fault feature vector into a deep neural network model to perform fault analysis and obtain fault analysis data.

7. The GIS equipment operation fault analysis system based on three-dimensional modeling according to claim 6 is characterized in that: The analysis module further comprises: Used to construct an input tensor according to the fault feature vector to obtain input tensor data; Used to input the input tensor data into a convolutional neural network model to obtain a convolutional feature map; Used to perform a maximum pooling operation on the convolution feature map to obtain a pooled feature map; Used to flatten the pooled feature map into a one-dimensional vector to obtain a one-dimensional feature vector; Used to input the one-dimensional feature vector into a long short-term memory network to obtain time series feature data; Used to input the time series feature data into the attention layer, perform feature weighting, and obtain weighted feature data; Used to input the weighted feature data into a fully connected layer, perform nonlinear transformation, and obtain a fault probability distribution; Used to perform classification according to the fault probability distribution to obtain a classification result; Used to map the classification result into a fault code to obtain fault code data; Used to combine the fault code data with corresponding fault description information to obtain fault analysis data.

Images