GIS equipment fault diagnosis method and system based on Transform neural network

Abstract

The invention discloses a GIS equipment fault diagnosis method and system based on a Transform neural network, and relates to the technical field, and the method comprises the following steps: obtaining the key data of the operation state of GIS equipment in real time, and preprocessing the data, the key data comprising an original vibration signal, equipment temperature rise and the decomposition product concentration of insulating gas; carrying out mechanical fault identification and non-mechanical fault identification on the processed data, fusing identification results, and outputting fault classification information; constructing an equipment fault knowledge graph and generating a fault report in combination with the fault classification information; the fault report of the GIS equipment is visually displayed through a three-dimensional graphic technology; obtaining user feedback information, and dynamically updating the equipment fault knowledge graph; according to the method, advanced deep learning, knowledge graph construction, three-dimensional graph rendering and feedback learning mechanisms are introduced, so that the accuracy, robustness and interactivity of GIS equipment fault diagnosis are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of GIS equipment fault diagnosis, and more particularly to a GIS equipment fault diagnosis method and system based on a Transformer neural network. BACKGROUND

[0002] As a key component in modern power systems, gas insulated switchgear (GIS) is widely used in urban power grids, substations and other high-voltage power facilities due to its high reliability, compact size and ability to adapt to complex environments. With the development of intelligent and efficient power systems, GIS equipment plays an indispensable role in ensuring the stability and safety of power transmission. However, as the operating time of GIS equipment increases, its complex internal mechanical structure also faces various potential failure risks, which may not only cause power outages, but also lead to serious safety accidents and economic losses.

[0003] For example, the invention patent with the announcement number CN112329937B announces a GIS fault diagnosis method based on cases and fault reasoning, which includes the following steps: step S01: the user inputs the description information of the GIS fault data characteristic attributes to the user layer and proposes the query requirements of the GIS fault data characteristic attributes; step S02: the user layer pre-processes the description information of the GIS fault data characteristics to obtain the fault phenomenon characteristic values in the preset format; step S03: using a matching algorithm for case matching, calling the GIS equipment diagnosis cases corresponding to the description information of the GIS fault data characteristics in the fault case library to achieve the GIS equipment diagnosis cases with a preset similarity to the GIS equipment diagnosis cases for fault diagnosis; step S04: obtaining the diagnosis conclusion of the corresponding GIS equipment diagnosis cases. The present application can utilize existing data materials to avoid complex data modeling process, making the GIS equipment fault diagnosis conclusion more reliable and practical.

[0004] The above-mentioned technical solution has at least the following technical problems:

[0005] Traditional methods are difficult to effectively fuse multi-dimensional heterogeneous operation data (such as vibration, gas, temperature rise), often focusing on a single feature dimension, resulting in limited diagnostic accuracy and coverage. For complex nonlinear vibration characteristics in mechanical faults, traditional signal processing or shallow learning models are difficult to capture their deep temporal relationships, which may easily cause misdiagnosis or missed diagnosis. Existing systems lack dynamic absorption and model updating capabilities for user feedback, resulting in long-term reliance on static knowledge bases for fault diagnosis systems, which cannot adapt to new fault patterns caused by equipment aging or changes in operating environment.

[0006] In view of the above problems, the present application provides a solution. SUMMARY

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a GIS equipment fault diagnosis method and system based on Transformer neural networks. By introducing advanced deep learning, knowledge graph construction, 3D graphics rendering, and feedback learning mechanisms, the accuracy, robustness, and interactivity of GIS equipment fault diagnosis are significantly improved.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A fault diagnosis method for GIS equipment based on Transformer neural networks includes the following steps: real-time acquisition of key data on the operating status of GIS equipment and preprocessing the data, wherein the key data includes raw vibration signals, equipment temperature rise, and concentration of decomposition products of insulating gas; mechanical fault identification and non-mechanical fault identification of the processed data, and fusion of the identification results to output fault classification information; construction of an equipment fault knowledge graph and generation of fault reports based on the fault classification information; visualization of the GIS equipment fault reports using 3D graphics technology; acquisition of user feedback information and dynamic updating of the equipment fault knowledge graph.

[0009] In a preferred embodiment, preprocessing includes bandpass filtering of the original vibration signal, specifically: converting the acquired original vibration signal into an original digital signal; inputting the original digital signal into a bandpass filter for processing to obtain an intermediate signal with limited frequency characteristics; based on the intermediate signal, extracting signal segments of a preset window length in chronological order as an input vector sequence and defining a desired signal; inputting the input vector sequence into an LMS adaptive filter, performing weight iterative updates based on the difference between the input vector sequence and the desired signal until a preset number of iterations is reached, and outputting the processed vibration characteristics.

[0010] In a preferred embodiment, mechanical fault identification includes identifying faults based on Gramian angular domain and Transformer neural network on preprocessed vibration features. Specifically, the preprocessed vibration features are standardized to obtain a standardized time series; the time series is mapped to polar coordinate space, and the signal value at each time point is converted into a corresponding angle series; a two-dimensional Gramian angular domain matrix GASF is constructed based on the angle series; and the mechanical fault type is output from the GASF matrix through a Transformer neural network.

[0011] In a preferred embodiment, the GASF matrix is ​​used to output mechanical fault types via a Transformer neural network. Specifically, the GASF matrix is ​​divided into several patches, and each patch is flattened. A two-dimensional position code in the GASF matrix is ​​added to each flattened patch, and the position code is fused into the patch through vector addition to form a patch feature vector with spatial position information. The patch feature vector is input into a multi-layer Transformer encoder, and features are extracted from the patch feature vector using a learnable parameter matrix based on the multi-head self-attention mechanism in the encoder. The feature extraction result of the Transformer encoder is input into a feedforward neural network to extract high-dimensional features and obtain a classification label vector. The classification label vector is input into a Softmax classifier, and a fully connected layer is used to calculate the probability distribution of each fault type. Based on the probability distribution, the category with the highest probability is selected as the final fault type, and the fault location is obtained by combining the spatial position mapping of the GASF matrix.

[0012] In a preferred embodiment, non-mechanical fault identification specifically involves: normalizing the concentration of decomposition products of the collected insulating gas to construct a first feature vector; constructing an insulation state identification module to perform pattern analysis on the first feature vector to obtain a preliminary insulation state judgment result; based on the preliminary judgment result, extracting the corresponding electrical disturbance index to construct a second feature vector; inputting the second feature vector and the preliminary insulation state judgment result into a preset fault reasoning module to output the electrical fault type; selecting the corresponding heating area as the observation object according to the electrical fault type and extracting thermal anomaly indexes to form a third feature vector; according to a preset thermal anomaly judgment threshold, considering the third feature vector exceeding the threshold as an anomaly to obtain thermal anomaly features; and fusing the thermal anomaly features, the preliminary insulation state judgment result, and the electrical fault type to obtain the non-mechanical fault identification result.

[0013] In a preferred embodiment, constructing a device fault knowledge graph involves: constructing a multi-dimensional fault event vector based on fault classification information; performing named entity recognition on various entities in the multi-dimensional fault event vector using BERT and CRF methods; performing semantic standardization processing on the named entities based on a predefined power system device fault domain knowledge ontology; constructing relational triples between entities based on named entity recognition and semantic standardization processing; storing the relational triples in a structured manner using a Neo4j graph database; aligning the newly generated triples with existing graph entities, generating vector representations for each node in the graph using a graph embedding method based on GraphSAGE, and calculating cosine similarity; if the similarity between a new node and an existing node exceeds a preset threshold, they are determined to be the same entity or relation, and attribute merging and edge updates are performed; otherwise, new nodes or edges are added.

[0014] In a preferred embodiment, a fault report is generated by combining fault classification information, specifically as follows: Based on the constructed equipment fault knowledge graph, a multi-dimensional fault event vector is collected for each fault event node. An updated node embedding vector is obtained by aggregating entity nodes and their adjacent edges using a graph convolutional neural network. Based on the node embedding vector, a starting node and a target node are set, and the similarity between any two nodes in the graph and the edge weight of each relation edge are obtained. Based on the similarity and edge weight, the path connecting the starting node and the target node in the graph is scored. Using the A* path search algorithm, multiple causal paths with scores higher than a preset path threshold are extracted as inference chains. A structured fault report is generated based on the inference chains, and the structured fault reports are summarized into a standardized fault report, which includes a fault description section, a suggested handling section, and a prevention suggestion section.

[0015] In a preferred embodiment, 3D graphics technology is used to visualize fault reports from GIS equipment. Specifically, this involves: constructing a 3D model library to acquire and store standardized 3D models of GIS equipment; loading the 3D models into a graphics rendering engine; receiving fault reports based on the equipment's unique identifier and mapping the fault type and location to corresponding parts of the 3D model using a hierarchical structural resolution method; generating dynamic fault labels at the corresponding model parts; providing an interactive interface on the user terminal that supports rotation, scaling, and clicking on the 3D model, and displaying detailed information related to the fault after the user selects a label; and analyzing the received fault reports based on a generative agent to generate processing suggestions and preventative measures related to the fault, which are then displayed in the 3D model interface using semantic tags.

[0016] In a preferred embodiment, acquiring user feedback information and dynamically updating the equipment fault knowledge graph involves: collecting user feedback on fault diagnosis results through a user input interface; normalizing the terminology of text fields in the feedback information and generating a standardized phrase set; labeling and identifying the phrase set using BERT and CRF methods to extract feedback information entities and their corresponding confidence scores, whereby the feedback information entities include fault type entities, involved component entities, maintenance measure entities, and fault result entities, and the confidence scores are calculated using the average marginal probability of CRF; mapping the feedback information entities to a predefined multi-level labeling system and generating a structured feedback data table; performing quality verification on the structured feedback data table, including integrity verification and rationality judgment; converting the verified structured feedback data into high-dimensional semantic vectors; calculating the cosine similarity between the high-dimensional semantic vectors and the vectors of known historical fault records in the knowledge graph, and marking the current feedback as a potential new fault feedback if the similarity is lower than a preset feedback threshold; and embedding the feedback information marked as potential new fault feedback into the GIS equipment fault knowledge graph in the form of triples.

[0017] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By constructing a hierarchical fusion diagnostic architecture for multi-source heterogeneous data and combining the collaborative analysis of vibration, gas, electrical, and thermal features, this approach offers multiple advantages, including high precision, robustness, and real-time performance. On one hand, the combined processing of the original vibration signal using bandpass filtering and LMS adaptive filtering significantly improves signal quality and the accuracy of subsequent feature extraction. Vibration image modeling through Gramian angular domain analysis and Transformer networks enhances the ability to identify complex mechanical faults. On the other hand, the non-mechanical fault identification module achieves collaborative diagnosis of insulation, electrical, and thermal faults through joint inference of gas decomposition products, harmonic disturbances, and thermal anomaly indicators. In particular, principal component analysis, K-means clustering, and probabilistic inference models improve the resolution and logical consistency of anomaly pattern recognition. Finally, a unified fault classification result is output through multimodal feature fusion, making this solution not only suitable for complex and nonlinear equipment operating environments but also possessing strong engineering adaptability and practical applicability, providing effective support for intelligent operation and maintenance of power equipment.

[0018] 2. By integrating key technologies such as multimodal perception, knowledge graph construction, graph neural network reasoning, 3D visualization, and closed-loop updates based on user feedback, it possesses high levels of intelligence and engineering practicality. Its advantages are reflected in: improving the accuracy of fault event characterization through the fusion of multi-source data such as vibration, gas, and electrical data; achieving context-enhanced named entity recognition by combining BERT+CRF, effectively identifying complex fault terms; constructing a high-semantic-accuracy knowledge graph using structured ontology and graph embedding methods to support accurate causal reasoning and fault report generation; introducing graph convolution and A* path algorithms to achieve semantic path scoring, enhancing the interpretability of the diagnostic chain; using 3D models and semantic labels for visualization, improving the intuitiveness and interactivity of fault location; and finally, achieving dynamic graph updates through user feedback and semantic encoding models, constructing a knowledge self-evolution mechanism, and comprehensively improving the intelligent, visual, and closed-loop management capabilities of equipment operation and maintenance. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of a GIS equipment fault diagnosis method based on a Transformer neural network, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the structure of a GIS equipment fault diagnosis system based on a Transformer neural network, provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] refer to Figure 1 , Figure 1 A flowchart illustrating a GIS equipment fault diagnosis method based on a Transformer neural network, provided for embodiments of this application, includes the following steps: S1. Real-time acquisition of key data on the operating status of GIS equipment and preprocessing of the data. The key data includes the original vibration signal, equipment temperature rise, and concentration of decomposition products of insulating gas.

[0023] Data acquisition is achieved through sensors installed at key locations on the equipment. An accelerometer collects raw vibration signals, a temperature sensor monitors the equipment's temperature rise, and an SF6 gas concentration detector monitors the concentration of decomposition products of the insulating gas. The sampling frequency is set to 20480 Hz to ensure the capture of the dynamic characteristics of high-frequency signals.

[0024] Further preprocessing includes bandpass filtering of the original vibration signal, and normalization and interpolation based on temperature rise and decomposition product concentration.

[0025] The original vibration signal is bandpass filtered, specifically as follows: The acquired raw vibration signals are converted into raw digital signals; For example, the acquired raw vibration signal can be converted into a raw digital signal using an A / D converter.

[0026] The original digital signal is input into a bandpass filter for processing to obtain an intermediate signal with limited frequency characteristics; The bandpass filter is a finite impulse response filter, and its passband frequency range is between a preset lower limit frequency and an upper limit frequency. It is used to retain the effective signal components within the target frequency band and suppress interference outside the frequency band. The filter order is set according to the specific filtering accuracy requirements.

[0027] Based on the intermediate signal, signal segments of a preset window length are extracted in chronological order as the input vector sequence, and the desired signal is defined. The input vector sequence is input into the LMS adaptive filter, and the weight is iteratively updated based on the difference between the input vector sequence and the desired signal. The weight iterative update is performed iteratively based on the minimum mean square algorithm with a preset fixed step size until the preset number of iterations is reached, and the processed vibration characteristics are output.

[0028] The input vector sequence is a segment of signal with a fixed window length extracted chronologically from the intermediate signals of the current time and several previous time steps. The window length is consistent with the order of the LMS filter. The desired signal can originate from the actual reference signal, a template signal generated based on the system's prior model, or be synchronously acquired through an external channel, and is used to guide the filter weight update. The algorithm iterates using a fixed-step least mean square algorithm. This can be understood as follows: in the LMS adaptive filter, the error between the current input vector sequence and the desired signal is calculated at each step, and the weight vector is adjusted according to this error at a fixed step size. ,in For the updated weights, Given an input vector sequence, The difference between the input vector sequence and the desired signal. As the initial weights, The process involves setting a fixed step size and iterating continuously until a preset condition (such as the number of iterations) is met. This aims to minimize the output error and make the output signal approximate the desired signal, thus obtaining the processed vibration characteristics. .

[0029] It should be noted that the dual processing mechanism of bandpass filtering combined with LMS adaptive filtering in the preprocessing of the original vibration signal can significantly improve signal quality and the reliability of subsequent feature extraction. Specifically, the bandpass filter is used to initially remove low-frequency drift and high-frequency random noise, ensuring that only effective signal components within the target frequency band are retained, thus compressing interference in the frequency domain. Based on this, the LMS adaptive filter is introduced to further dynamically suppress residual interference and non-stationary background noise in the time domain. It iteratively optimizes weights using the minimum mean square error criterion, possessing real-time tracking and adaptive modeling capabilities, making it particularly suitable for processing nonlinear and non-stationary vibration signals. Furthermore, this method constructs the input vector through a sliding window and combines it with external or model-desired signals, allowing for flexible configuration according to the actual system characteristics. This enhances signal feature fidelity while improving the system's ability to detect weak faults and sudden anomalies. Therefore, this preprocessing scheme not only improves signal usability and analysis accuracy but also provides a more stable and high signal-to-noise ratio data foundation for subsequent state diagnosis and anomaly detection.

[0030] S2 identifies mechanical and non-mechanical faults in the processed data, merges the identification results, and outputs fault classification information.

[0031] Furthermore, mechanical fault identification includes fault identification based on Gramian angular domain and Transformer neural network on preprocessed vibration features, specifically: Preprocessed vibration characteristics The time series is standardized by mapping it to the interval [-1, 1]. ; The standardization process is specifically calculated using the following formula:

[0032] In the formula, For the first A vibration signal.

[0033] The time series is mapped to polar coordinate space, and the signal value at each time point is converted into a corresponding angle sequence. ; The specific calculation formula for converting this into the corresponding angle sequence is as follows:

[0034] The two-dimensional Gramian angular domain matrix GASF is constructed based on the angle sequence. The specific calculation formula is as follows:

[0035] formula, This represents the value of the GASF two-dimensional matrix at the i-th row and j-th column. , It is an angle sequence for signal value conversion.

[0036] The GASF matrix is ​​used to output the mechanical fault type through a Transformer neural network.

[0037] It should be noted that by mapping the standardized time series to polar coordinate space and then constructing a symmetric matrix using the cosine relationship of angle summation, not only is the global time-dependent structure and dynamic evolution trend of the vibration signal effectively preserved, but also one-dimensional time-series signals are encoded into a two-dimensional image representation with spatial characteristics. This representation has strong visualization capabilities, facilitating the subsequent extraction of local and global pattern features using image modeling methods (such as Transformer), thereby improving the accuracy and generalization ability of fault identification in complex systems.

[0038] Furthermore, the GASF matrix is ​​used to output the mechanical fault type through a Transformer neural network, specifically: The GASF matrix is ​​divided into several patches of size P×P, and the specific calculation formula is as follows:

[0039] In the formula, Let N be the size of the GASF matrix. This refers to dividing the GASF matrix into several fixed-size local sub-image blocks.

[0040] Each patch is represented as follows: .

[0041] Flatten each patch to form a corresponding one-dimensional vector. :

[0042] In the formula, k represents the k-th patch.

[0043] The flattened patch is then given a two-dimensional positional encoding in the GASF matrix. This positional encoding is then fused into the patch representation using vector addition, forming a patch feature vector with spatial location information. ; The specific formula for calculating the Patch feature vector is as follows:

[0044] In the formula, It is a two-dimensional positional encoding.

[0045] The Patch feature vector is input into a multi-layer Transformer encoder. Based on the multi-head self-attention mechanism in the encoder, features are extracted from the Patch feature vector using a learnable parameter matrix. The specific calculation formula is as follows:

[0046] In the formula, Q is the query vector, K is the key vector, and V is the value vector. To query the dot product of vector Q and key vector K, For the key / query vector dimension, To calculate the weighted sum, This is a self-attention mechanism.

[0047] in:

[0048]

[0049]

[0050] In the formula, , , This is the learnable parameter matrix.

[0051] The output of the Transformer encoder is sent to the feedforward neural network to extract high-dimensional features and obtain the classification label vector; The classification label vector is input into the Softmax classifier, and the probability distribution of each fault type is calculated through a fully connected layer. The probability distribution is calculated using the following formula:

[0052] In the formula, The probability of each fault type, For classification label vectors, The weights of the classifier, This is the bias term for the classifier.

[0053] Based on the probability distribution, the category with the highest probability is selected as the final fault type, and the fault location is obtained by combining the spatial location mapping of the GASF matrix.

[0054] Among them, mechanical failure types include poor contact, loose shielding, internal foreign objects, mechanical fatigue or wear.

[0055] It should be noted that by dividing the two-dimensional GASF matrix into small patches and flattening them into vectors, the matrix-to-sequence conversion is completed, providing an input format for the Transformer model. By introducing the positional information of the patches, the spatial structure of the original matrix is ​​preserved, effectively enhancing the model's perception of local semantics. A multi-head self-attention mechanism is used to model the long-range dependencies between patches, extracting global features and improving the understanding of complex fault modes. Global semantics are further aggregated through a feedforward network and classification label vectors, providing a discriminative representation for the final classification. A Softmax classifier outputs multi-class probability distributions, supporting fine-grained identification of various fault types. Finally, the prediction results are transformed into structured diagnostic information, enabling interpretable classification decisions and fault location, providing an operational basis for subsequent operation and maintenance. The entire process, while preserving temporal and spatial information, possesses high robustness, high recognition accuracy, and good interpretability, making it suitable for fault diagnosis tasks dominated by high-dimensional, complex, and nonlinear features.

[0056] Further, non-mechanical fault identification specifically includes: The concentration of the decomposition products of the collected insulating gas is normalized to construct a first feature vector. The concentration of the decomposition products includes sulfur dioxide, hydrogen sulfide, carbon tetrafluoride, carbon monoxide, carbon dioxide, and hydrogen fluoride. The first feature vector is used to characterize the decomposition gas characteristic pattern caused by processes such as aging of insulating materials, discharge, or pyrolysis inside the equipment. An insulation status identification module is constructed based on feature pattern recognition rules. The first feature vector is analyzed to determine whether the target device currently has an insulation abnormality and its fault type, and a preliminary insulation status judgment result is obtained. The preliminary insulation status judgment result includes normal, corona discharge, surface discharge, creepage discharge, and thermal aging. Based on the discharge type determined in the initial assessment, the corresponding electrical disturbance indexes are extracted to construct a second feature vector. The electrical disturbance indexes include the effective value change of current, the peak voltage shift, and the harmonic distortion rate of current and voltage. The second feature vector and the initial insulation state judgment result are input into the preset fault reasoning module, and the electrical fault type is output. The fault reasoning module constructs an electrical fault classification model based on the response characteristics of the electrical disturbance index under different electrical fault types and the conditional dependence of the insulation state on it. This model is used to determine the electrical fault type of the current disturbance. The electrical fault types include overcurrent, inrush current, ferroresonance, arc breakdown, and normal disturbance. The classification model can be implemented using a joint probabilistic reasoning method, which associates the electrical disturbance index with the insulation state to perform conditional classification inference. Based on the electrical fault type, the heating area corresponding to the physical mechanism of the fault type is selected as the observation object. Based on the equipment temperature rise in the selected area, thermal anomaly indicators are extracted to form a third feature vector. According to the electrical fault type, a corresponding thermal anomaly judgment threshold is dynamically set. Any third feature vector exceeding the thermal anomaly judgment threshold is considered an anomaly, thus obtaining the thermal anomaly characteristics. The thermal anomaly judgment thresholds include, but are not limited to, peak temperature thresholds and temperature rise rate thresholds, to improve the targeting and accuracy of thermal fault identification. The thermal anomaly indicators include local temperature peaks and temperature rise rates; the thermal anomaly characteristics include temperature anomalies and temperature rise rate anomalies. The thermal anomaly characteristics, the initial insulation condition assessment results, and the electrical fault type are fused to obtain non-mechanical fault identification results, which include normal, insulation material aging, internal partial discharge, conductor heating, etc.

[0057] The pattern analysis method in the insulation condition identification module first collects the normalized concentration of each decomposition product in sulfur hexafluoride gas and constructs a multidimensional ratio vector based on the ratio relationship between each component. Secondly, it uses principal component analysis to map the high-dimensional ratio vector to the principal component space, calculates the Euclidean distance between each sample in the space, and uses K-means clustering to determine the fault category of the current sample based on the density aggregation degree of sample points within the category, thereby realizing the classification and identification of states such as corona discharge, surface discharge, and pyrolysis.

[0058] It should be noted that by constructing a hierarchical and integrated multimodal fault identification process, the collaborative analysis and discrimination of insulating gas characteristics, electrical disturbance response, and thermal anomaly behavior are achieved, which has the following advantages and technical functions: On the one hand, by constructing the normalized concentration of sulfur hexafluoride decomposition products as the first feature vector, and combining principal component analysis and K-means clustering algorithms, high-resolution identification of insulation status is achieved, which can effectively distinguish different insulation anomaly types such as corona discharge, surface discharge, and thermal aging; on the other hand, the system selectively extracts electrical disturbance characteristics based on the initial insulation judgment results, constructs a second feature vector, and then... The joint probabilistic reasoning model establishes a conditional dependency between electrical indicators and insulation status, enhancing the identification accuracy of complex electrical faults (such as inrush current, ferroresonance, and arc breakdown). Furthermore, the scheme incorporates electrical fault types into the thermal anomaly analysis process, enabling adaptive configuration of the thermal observation area and judgment threshold, thus improving the ability to locate thermal faults such as equipment overheating and abnormal conductor heating. Finally, through multi-source feature fusion, it achieves comprehensive identification of non-mechanical composite faults, providing a logically dependent, responsive, and engineering-feasible technical approach for power equipment condition assessment and fault early warning.

[0059] Furthermore, the identification results are fused to output fault classification information, as shown in Table 1, which details the specific faults diagnosed by the system. Table 1

[0060] S3 constructs a knowledge graph of equipment faults and generates fault reports by combining fault classification information.

[0061] Furthermore, a knowledge graph of equipment failures is constructed, specifically as follows: A multidimensional fault event vector is constructed based on fault classification information. The multidimensional fault event vector includes, but is not limited to, fault type label, fault location, and multimodal feature fusion result. The multimodal feature fusion result includes standardized vibration signal, gas decomposition concentration, and electrical disturbance index. The fault type label is derived from the identification results of the Transformer neural network and the fault reasoning module in the fault diagnosis module; the fault location is derived from the spatial location mapped by the GASF matrix and the temperature rise anomaly observation point. Multimodal feature fusion employs a weighted principal component analysis method, normalizing the standardized vibration signal, gas decomposition concentration, and electrical disturbance index to the interval [0,1], and combining this with the perceptual weight matrix for feature projection to obtain a unified fusion vector. This multidimensional fault event vector serves as the basis for entity and relation generation in the knowledge graph construction.

[0062] Named entity recognition is performed on various entities in a multidimensional fault event vector using BERT and CRF methods. The named entities are semantically standardized based on a predefined knowledge ontology of power system equipment faults. The knowledge ontology predefines standard representations of terms such as typical equipment categories, fault types, sensor variables, and control strategies, along with their hierarchical relationships. A hybrid matching algorithm based on edit distance and context cosine similarity is used to match the recognition results with standard concepts in the ontology. Semantic standardization is then performed when the matching degree reaches a preset matching threshold.

[0063] Based on named entity recognition and semantic standardization, a triplet of relationships between entities is constructed through a rule-driven relation template. The triplet includes the following forms: (fault type, occurrence location, equipment location), (fault type, accompanying features, multimodal indicators), (fault type, triggering conditions, diagnostic context information), (fault type, suggested measures, operation and maintenance strategy). The relation triples are stored in a structured manner using a Neo4j graph database. The newly generated triples are aligned with the existing graph. A graph embedding method based on GraphSAGE is used to generate vector representations for each node in the graph and calculate the cosine similarity. If the similarity between the new node and the existing node exceeds a preset threshold, they are determined to be the same entity or relation, and attribute merging and edge updates are performed. Otherwise, a new node or edge is added.

[0064] Among them, BERT is a pre-trained language model based on the Transformer architecture, which can understand the bidirectional semantics of words in context. CRF is a statistical modeling method that is often used in sequence labeling tasks to optimize the dependencies between labels.

[0065] It should be noted that using BERT and CRF to perform named entity recognition on key terms, entity names and associated contexts in multidimensional fault event vectors can be understood as follows: using the BERT model to encode each word in a context-aware manner and extract its meaning in the sentence; then feeding these encoded vectors into the CRF layer, and combining the dependencies between the context and the labels to perform named entity recognition on key terms (such as equipment names, components, fault types, etc.) in the whole sentence.

[0066] By introducing a context-aware named entity recognition mechanism based on BERT+CRF and a multimodal fault information fusion mechanism, the accuracy and normalization ability for identifying complex, low-frequency, and semantically ambiguous fault terms are significantly improved. Simultaneously, GASF spatial mapping and Transformer neural networks are used to achieve automated identification of fault types and locations, enhancing the accuracy and semantic representation of entity relationships in the knowledge graph. Compared to traditional methods, this approach not only possesses stronger adaptability and self-learning capabilities but also dynamically integrates structured and unstructured information, making the constructed knowledge graph more closely resemble actual operation and maintenance scenarios and supporting high-precision intelligent diagnosis and reasoning services.

[0067] Furthermore, a fault report is generated by combining the fault classification information, specifically as follows: Based on the constructed equipment fault knowledge graph, a multi-dimensional fault event vector is collected for each fault event node; The updated node embedding vector is obtained by aggregating entity nodes and their adjacent edges using a graph convolutional neural network, wherein the graph convolutional neural network is selected from GCN, GAT, GraphSAGE or a combination thereof; Based on the updated node embedding vector, set the start node and the target node; The starting node is such as the fault type node to be analyzed, and the target node is such as the expected impact node or the known consequence node.

[0068] The similarity between any two nodes in the graph and the edge weight of each relation edge are obtained. The similarity is the cosine similarity between their embedding vectors, which is used to measure whether two nodes have semantic similarity in the graph. The edge weight is obtained by learning the scoring function of relation triples through the TransE embedding model and converting the score into the edge weight. Based on similarity and edge weight, the path connecting the starting node and the target node in the graph is scored. The score is the weighted product of the embedding similarity of all adjacent nodes on the path and the edge weight. Using the A* path search algorithm, multiple causal paths with scores higher than a preset path threshold are extracted as inference chains. These inference chains are used to explain the possible causes and consequences of the current fault. A structured fault report is generated based on the inference chain. The structured fault report includes fault description generation, which sorts the node sequence in the inference chain according to time order and spatial distribution, and maps node attributes to natural language description sentences through template matching, such as: "In a certain part of a certain device, a certain abnormal feature was detected, and based on historical events, it can be inferred that it is a certain fault type"; processing suggestion generation, which retrieves the most similar processing case node in the historical knowledge graph based on the embedding vector of the current fault node in the inference chain, and selects the corresponding processing action information as the suggested measures through a similarity weighting mechanism; and prevention measure generation, which traces the inference chain forward, extracts the preceding event patterns that led to the current fault, predicts potential future hidden danger nodes based on their evolution trend, and generates executable prevention measure statements according to the predefined risk control rules in the ontology. The structured fault reports are aggregated into a standardized fault report output, which includes a fault description section, a suggested handling section, and a prevention suggestion section.

[0069] It should be noted that graph convolutional networks can effectively capture the implicit patterns of fault event nodes and their upstream and downstream relationships, enhancing the interpretability of the diagnostic logic chain. Secondly, the path scoring mechanism based on node embedding similarity and edge weights makes the extraction of reasoning chains more targeted and semantically accurate. Thirdly, the introduction of structured templates and case similarity retrieval mechanisms makes fault descriptions, handling suggestions, and preventive measures highly readable and have practical guiding value.

[0070] S4 uses 3D graphics technology to visualize fault reports from GIS equipment, specifically as follows: A 3D model library is constructed to acquire and store standardized 3D models of GIS equipment. The 3D models are obtained by converting and standardizing the CAD data provided by the equipment manufacturer. The models include the internal structural information of the equipment and the geometric features of key components. The three-dimensional model is loaded into the graphics rendering engine, which receives fault reports based on the device's unique identifier and maps the fault type and location to the corresponding part in the three-dimensional model through a structural hierarchy resolution method. The graphics rendering engine supports real-time rendering and interactive operation of three-dimensional graphics. The graphics rendering engine is a graphics system built on WebGL or Unity3D, supporting GPU-accelerated rendering and multi-device compatible display.

[0071] Dynamic fault labels are generated at the corresponding model parts. The labels include color codes, graphic icons and text descriptions to indicate the fault location and fault type. An interactive interface is provided on the user terminal, which supports rotation, scaling and clicking operations on the 3D model, and displays detailed information related to the fault after the user selects a tag; The received fault reports are analyzed using a generative agent to generate processing suggestions and preventive measures related to the faults. These suggestions and measures are then displayed in the 3D model interface using semantic tags to assist users in fault diagnosis and decision-making.

[0072] The generative agent is based on a text generation model (such as fine-tunedGPT) based on the Transformer architecture. Its input is standardized fault report content (structured JSON format), and its output is processing suggestions and preventive measures. The content is generated through template sentences and labeled with semantic tags (such as "warning level: high" and "recommended operation: replace contact component"). The labeled content is bound to the additional information box of the fault location in the 3D model.

[0073] It should be noted that by constructing a standardized 3D model library for GIS equipment and combining it with a graphics rendering engine and unique equipment identification codes, accurate mapping and visualization of fault information in the 3D structure are achieved. This not only improves the intuitiveness and accuracy of equipment fault location but also supports multi-angle interactive operation, helping maintenance personnel to quickly identify and analyze fault locations. Simultaneously, a generative intelligent agent is introduced to intelligently analyze fault reports, generating handling suggestions and preventative measures, which are embedded into the 3D interface in the form of semantic tags. This enhances the system's decision support capabilities, significantly improves fault response efficiency and maintenance quality, and demonstrates a high degree of engineering practicality and intelligence.

[0074] S5 obtains user feedback information and dynamically updates the device fault knowledge graph, specifically: The system collects user feedback on the fault diagnosis results through a user input interface. The feedback includes an evaluation of the accuracy of the diagnosis results, actual repair measures and results, repair time and personnel information, and system improvement suggestions. The text fields in the feedback information are normalized using a domain-defined dictionary and a thesaurus expansion table. The normalized text field is processed by a syntactic analyzer to identify semantic boundaries and generate a standardized set of phrases. The phrase set is labeled and identified using BERT and CRF methods, and feedback information entities and corresponding confidence scores are extracted. The feedback information entities include fault type entities, involved component entities, maintenance measure entities, and fault result entities. The confidence scores are calculated using the average marginal probability of CRF. The feedback information entities are mapped to a predefined multi-level tag system, and a structured feedback data table is generated. Perform quality verification operations on the structured feedback data table, including integrity verification and rationality judgment. Among them, the integrity check ensures that all four types of entities have been extracted and that the confidence level is higher than the preset entity threshold. The reasonableness judgment is to call the built-in rule library to check the logical consistency between tags (for example, "insulation breakdown" should not co-occur with "handling method: fastening bolts") and mark abnormal fields. Convert the validated structured feedback data into high-dimensional semantic vectors; Semantic coding models such as SimCSE or Sentence-BERT can be used for conversion.

[0075] The cosine similarity between the high-dimensional semantic vector and the vector of the known historical fault record in the knowledge graph is calculated. If the similarity is lower than the preset feedback threshold, the current feedback is marked as a potential new fault feedback. Feedback information marked as potential new faults is embedded into the GIS equipment fault knowledge graph in the form of triples, such as ("fault type-affected component-maintenance suggestion"), to achieve dynamic updates of structured knowledge and continuous enrichment of diagnostic semantics.

[0076] The confidence level is calculated using the following formula:

[0077] In the formula, For confidence level, For the t-th position, Let t be the predicted label obtained from the CRF solution. This is a piece of user feedback information that was entered. To give input Under the condition that the label of the t-th position is The marginal probabilities are obtained by CRF dynamic programming. Starting position This is the termination position.

[0078] refer to Figure 2 , Figure 2 A schematic diagram of a GIS equipment fault diagnosis system based on a Transformer neural network, provided in this application embodiment, includes a data acquisition and preprocessing module, a fault diagnosis module, a knowledge graph construction and fault report generation module, a visualization module, and a feedback module. The modules are interconnected. The data acquisition and preprocessing module is used to acquire key data on the operating status of GIS equipment in real time and preprocess the data. The key data includes the original vibration signal, equipment temperature rise and concentration of decomposition products of insulating gas. The fault diagnosis module is used to identify mechanical and non-mechanical faults in the processed data, and to fuse the identification results to output fault classification information. The knowledge graph construction and fault report generation module is used to construct a knowledge graph of equipment faults and generate fault reports by combining fault classification information. The visualization module is used to visualize fault reports from GIS equipment using 3D graphics technology. The feedback module is used to obtain user feedback information and dynamically update the equipment fault knowledge graph.

[0079] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0080] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0081] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0084] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A GIS equipment fault diagnosis method based on a Transformer neural network, characterized in that, The method comprises the following steps: Real-time acquisition of key data of GIS device operation state and preprocessing of the data, wherein the key data comprises original vibration signals, device temperature rise and decomposition product concentration of insulating gas; Mechanical fault identification and non-mechanical fault identification are performed on the preprocessed data, and the identification results are fused to output fault classification information; A device fault knowledge graph is constructed, and a fault report is generated in combination with the fault classification information; The fault report of the GIS device is visually displayed through three-dimensional graphics technology; User feedback information is acquired, and the device fault knowledge graph is dynamically updated.

2. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 1, characterized in that, The preprocessing comprises band-pass filtering of the original vibration signals, specifically: The collected original vibration signals are converted into original digital signals, and a band-pass filter is used to process the original digital signals to obtain intermediate signals with limited frequency characteristics; Based on the intermediate signals, signal segments with a preset window length are sequentially intercepted as input vector sequences, and an expected signal is defined; The input vector sequences are input into an LMS adaptive filter, and weight iterative updating is performed according to the difference between the input vector sequences and the expected signal until a preset number of iterations is reached, and the processed vibration features are output.

3. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 2, characterized in that, The mechanical fault identification comprises fault identification of the preprocessed vibration features based on a Gramian angle domain and a Transformer neural network, specifically: The vibration features are converted into standardized time series; The time series are mapped to a polar coordinate space, and the signal values at each time point are converted into corresponding angle sequences; A two-dimensional Gramian angle domain matrix GASF is constructed according to the angle sequences; The GASF matrix is output through a Transformer neural network to output a mechanical fault type.

4. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 3, characterized in that, The GASF matrix is output through a Transformer neural network to output a mechanical fault type, specifically: The GASF matrix is divided into several patches, and each patch is flattened; A two-dimensional position encoding of each flattened patch in the GASF matrix is added, and the position encoding is fused into the patch through vector addition to form a patch feature vector with spatial position information; The patch feature vector is input into a multi-layer Transformer encoder, and the patch feature vector is feature-extracted through the multi-head self-attention mechanism and the learnable parameter matrix in the encoder; The feature extraction result of the Transformer encoder is input into a feedforward neural network to extract high-dimensional features and obtain a classification label vector; The classification label vector is input into a Softmax classifier, and the probability distribution of each fault type is calculated through a fully connected layer; According to the probability distribution, the class with the maximum probability is selected as the final fault type, and the fault position is obtained in combination with the spatial position mapping of the GASF matrix.

5. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 4, characterized in that, The non-mechanical fault identification, specifically: The collected decomposition product concentration of insulating gas is normalized to construct a first feature vector; An insulation state identification module is constructed to perform pattern analysis on the first feature vector to obtain an initial judgment result of the insulation state; Based on the initial judgment result, the corresponding electrical disturbance index is extracted to construct a second feature vector; The second feature vector and the initial judgment result of the insulation state are input into a preset fault reasoning module, and the electrical fault type is output; According to the electrical fault type, the corresponding heating area is selected as the observation object, and the thermal anomaly index is extracted to form a third feature vector; According to the preset thermal anomaly judgment threshold, the third feature vector exceeding the threshold is regarded as an anomaly, and a thermal anomaly feature is obtained; The thermal anomaly feature, the initial judgment result of the insulation state, and the electrical fault type are fused to obtain a non-mechanical fault recognition result, which includes normal, insulation material aging, internal partial discharge, and conductor heating.

6. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 5, characterized in that, The device fault knowledge graph is constructed, specifically: A multi-dimensional fault event vector is constructed according to the fault classification information; The entities in the multi-dimensional fault event vector are recognized by BERT and CRF, and semantic standardization processing is performed; Based on the named entity recognition and semantic standardization processing, the relationship triplets between entities are constructed; The relationship triplets are stored in the form of a Neo4j graph database; The newly generated triplets are aligned with the existing graph, the GraphSAGE-based graph embedding method is used to generate vector representations of nodes in the graph, and the cosine similarity is calculated; If the similarity between the new node and the existing node exceeds the preset threshold, it is determined that they are the same entity or relationship, and the attributes are merged and the edges are updated, otherwise a new node or edge is added.

7. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 6, characterized in that, The fault report is generated based on the fault classification information, specifically: Based on the constructed device fault knowledge graph, the multi-dimensional fault event vector of each fault event node is collected; The entity nodes and their adjacent edges are aggregated to obtain updated node embedding vectors; Based on the node embedding vectors, the starting node and the target node are set, and the similarity between any two nodes in the graph and the edge weight of each relationship edge are obtained; Based on the similarity and the edge weight, the path connecting the starting node and the target node in the graph is scored; Through the A* path search algorithm, multiple causal paths above the preset path threshold are extracted as reasoning chains according to the path score; A structured fault report is generated based on the reasoning chain, and the structured fault report is aggregated into a standardized fault report, which includes a fault description section, a suggestion processing section, and a prevention suggestion section.

8. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 1, characterized in that, The fault report of the GIS device is visualized through three-dimensional graphics technology, specifically: A three-dimensional model library is constructed to obtain and store standardized three-dimensional models of GIS devices; The three-dimensional model is loaded into a graphics rendering engine, receives a fault report based on the unique identification code of the device, and maps the fault type and location to the corresponding part in the three-dimensional model through a structural hierarchy analysis method; Dynamic fault labels are generated on the corresponding model parts; An interactive interface is provided on the user terminal to support rotation, scaling, and clicking operations on the three-dimensional model, and detailed information related to the fault is displayed after the user selects the label. Based on the generated agent, the received fault report is analyzed, the corresponding processing suggestion and preventive measure are generated, and the semantic label is displayed in the three-dimensional model interface.

9. The GIS equipment fault diagnosis method based on the Transformer neural network according to claim 1, characterized in that, The user feedback information is acquired, and the device fault knowledge graph is dynamically updated, specifically: The user feedback information on the fault diagnosis result is collected through a user input interface, the text field in the feedback information is normalized in terms, and a standardized phrase set is generated; The phrase set is labeled and recognized by BERT and CRF method, and the feedback information entity and the corresponding confidence are extracted; The feedback information entity is mapped to the pre-defined multi-level label system, and a structured feedback data table is generated; The quality checking operation is performed on the structured feedback data table, and the quality checking operation includes integrity checking and rationality judgment; The structured feedback data passing the checking is converted into a high-dimensional semantic vector; The high-dimensional semantic vector and the vector of the known historical fault record in the knowledge graph are calculated by cosine similarity, and if the similarity is lower than the preset feedback threshold, the current feedback is marked as potential new fault feedback; The feedback information marked as potential new fault feedback is embedded into the GIS device fault knowledge graph in the form of triple.

10. A system using the GIS equipment fault diagnosis method based on the Transformer neural network according to any one of claims 1-9, characterized in that, It includes data acquisition and preprocessing module, fault diagnosis module, knowledge graph construction and fault report generation module, visualization module and feedback module, and there is a connection between the modules: The data acquisition and preprocessing module is used for real-time acquisition of key data of GIS device running state and preprocessing of data, and the key data includes original vibration signal, device temperature rise and decomposition product concentration of insulating gas; The fault diagnosis module is used for mechanical fault identification and non-mechanical fault identification of the processed data, and the identification results are fused to output fault classification information; The knowledge graph construction and fault report generation module is used for constructing the device fault knowledge graph and generating the fault report combined with the fault classification information; The visualization module is used for visualizing the fault report of GIS device through three-dimensional graphics technology; The feedback module is used for acquiring user feedback information and dynamically updating device fault knowledge graph.

Images