A method for constructing a health status diagnosis framework for power main equipment based on adversarial learning

Abstract

The present invention discloses a method for constructing a health status diagnosis framework for power main equipment based on adversarial learning, which belongs to the technical field of power equipment diagnosis. The present invention performs entity and relationship modeling; collects multi-source data of power main equipment, and constructs a knowledge graph based on the multi-source data of power main equipment; designs an adversarial learning model architecture, and uses a lightweight model for a single scenario as a small model; converts the knowledge graph into graph structure data, performs graph convolution to generate high-order features, and inputs the high-order features into a large model; uses the knowledge graph as a training sample for the small model, and feeds the output of the small model back to the knowledge graph; sets a training strategy for the adversarial learning model and performs adversarial training; obtains online monitoring data, and when the large model output device fails, queries the historical association rules of the device in the knowledge graph, retrieves the historical diagnostic results of the small model of the device in the corresponding scenario, and generates a diagnostic path diagram; generates work order priorities, and arranges operation and maintenance personnel to perform repairs.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment diagnosis, and in particular to a method for constructing a health status diagnosis framework for power main equipment based on adversarial learning. Background Art

[0002] In modern power grids, power equipment is numerous and diverse, and its health is directly related to the safe and stable operation of the grid. A failure in an equipment device can cause a widespread power outage. Traditional power equipment diagnostic methods, which rely primarily on manual experience and simple monitoring methods, are no longer able to meet the needs of today's power grid development.

[0003] Traditional power equipment health status diagnosis relies heavily on manual analysis. Single-state diagnosis primarily relies on simple methods like threshold analysis and trend analysis based on procedures, making it difficult to accurately diagnose complex faults. Existing diagnostic models often target a single device or a single fault type, lacking versatility and comprehensiveness. Correlations between different devices and fault types are not effectively utilized, making it difficult to accurately assess and predict the overall health status of power equipment. Traditional diagnostic methods are unable to achieve real-time, dynamic monitoring and diagnosis, making it difficult to promptly identify potential equipment failures. In terms of operation and maintenance management, they also lack intelligent decision-making support, resulting in low operation and maintenance efficiency and failing to meet the modern power grid's requirements for efficient and timely power equipment operation and maintenance. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for constructing a health status diagnosis framework of a main power device based on adversarial learning, so as to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present application provides a method for constructing a health status diagnosis framework for a power main device based on adversarial learning, comprising the following steps:

[0007] Identify entities, diagnostic scenario edges, and fault type nodes, and perform entity and relationship modeling; collect multi-source data from power main equipment, perform data cleaning and standardization, extract knowledge from unstructured data, and build a knowledge graph based on the multi-source data from power main equipment;

[0008] Design an adversarial learning model architecture, using a Transformer-based graph neural network as the large model and a lightweight model for a single scenario as the small model. Convert the knowledge graph into graph-structured data, perform graph convolution to generate high-order features, and input the high-order features into the large model. Use the knowledge graph as training samples for the small model, feed the output of the small model back into the knowledge graph, and use the small model as the atomic module for knowledge graph reasoning. Set the training strategy for the adversarial learning model and perform adversarial training between the large and small models.

[0009] Obtain online monitoring data. When a large model output device fails, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnosis results of the small model of the device in the corresponding scenario, and generate a diagnostic path diagram through Neo4jBloom.

[0010] Based on the diagnostic path map, equipment importance, and fault urgency, work order priorities are generated and maintenance personnel are assigned to perform repairs.

[0011] In conjunction with the first aspect, in a first implementation of the first aspect of the present application, it is characterized in that the determining of entities, diagnostic scenario edges, and fault type nodes, and performing entity and relationship modeling, includes:

[0012] The entities include a power master device entity, a monitoring indicator entity, and an operation and maintenance personnel entity, and the attributes of the entities are defined; the diagnostic scenario edges include monitoring edges, triggering edges, and participating edges. The monitoring edges represent the relationship between the power master device and the monitoring indicators. The triggering edges indicate that the abnormal values of the monitoring indicators trigger a certain diagnostic scenario. The participating edges indicate that the operation and maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power master device. The diagnostic scenarios include scenarios based on monitoring indicators and scenarios based on fault history. Faults are classified according to their nature, fault type nodes are determined, and fault type attributes are defined.

[0013] Use ER diagrams to represent the conceptual structure between entities, attributes, and relationships, determine the cardinality of relationships between entities, convert the conceptual structure into a logical model, and select the graph database Neo4j for storage; based on the logical model, create corresponding nodes and edges in the graph database.

[0014] In combination with the first aspect, in a second implementation of the first aspect of the present application, collecting multi-source data of the power master equipment, performing data cleaning and standardization, extracting knowledge from the unstructured data, and constructing a knowledge graph based on the multi-source data of the power master equipment include:

[0015] Collect multi-source data from main power equipment, process missing values, outliers, duplicate values, and noise, perform Z-score standardization, uniformly encode text data into the UTF-8 standard encoding format, perform text cleaning, and use one-hot encoding to convert categorical text data into numerical data. Use named entity recognition algorithms to extract knowledge from unstructured data, and use the BERT model for training to identify entities in unstructured data and determine the relationships between entities.

[0016] Map the processed multi-source data of power main equipment into the knowledge graph.

[0017] In combination with the first aspect, in a third embodiment of the first aspect of the present application, the design of the adversarial learning model architecture uses a Transformer-based graph neural network as a large model and a lightweight model for a single scenario as a small model, including:

[0018] For large models, the knowledge graph is converted into graph-structured data, where nodes represent entities and edges represent relationships between entities. Node attributes are encoded and converted into vector representations. An adjacency matrix of the graph structure is constructed to represent the connection relationships between nodes. A graph neural network is used as the graph convolution layer to aggregate and update node features. Several graph convolution layers are stacked to capture the high-order structural information of the graph. The Transformer's multi-head self-attention mechanism is introduced into the graph neural network, and the node features output by the graph convolution layer are fused with the output of the Transformer layer to obtain a comprehensive feature representation. For fault type identification tasks, the output layer uses a softmax function for multi-classification. For fault severity assessment tasks, the output layer uses linear regression.

[0019] For the small model, the lightweight model uses a multi-layer perceptron and uses data of the corresponding scenario in the knowledge graph to train the small model.

[0020] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, converting the knowledge graph into graph structure data, performing graph convolution to generate high-order features, and inputting the high-order features into the large model includes:

[0021] Entities in the knowledge graph are mapped to graph nodes, assigned unique IDs, and the entity attributes are converted into initial feature vectors of the nodes through numerical conversion. Diagnostic scenario edges in the knowledge graph are converted into graph edges, and the types and weights of graph edges are defined. Graph edge types include directed edges and undirected edges. Node connection relationships are represented using sparse matrices, with non-zero elements indicating associations between nodes.

[0022] A graph convolutional neural network is designed. The input layer is the initial node features. The GAT layer is the first layer, which aggregates direct neighbor features. The GCN layer is the second layer, which captures second-order neighbor relationships. The output layer is used to output high-order features and perform parameter configuration. For the main power equipment nodes, real-time operating parameters and static attributes are integrated and normalized to [0,1]. For the monitoring indicator nodes, the monitoring values and their threshold ranges are encoded as two-dimensional vectors. Adjacency matrix preprocessing is performed. In the first layer, each main power equipment node aggregates the features of its directly associated monitoring indicator nodes. In the second layer, the main power equipment nodes further aggregate the features of the fault type nodes. The outputs of multiple GATs are spliced or averaged to generate high-order features containing multi-order neighbor information. The high-order features are input into the large model.

[0023] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, the knowledge graph is used as a training sample for the small model, the output of the small model is fed back to the knowledge graph, and the small model serves as an atomic module for knowledge graph reasoning, including:

[0024] By device type and diagnostic scenario, corresponding subgraphs are extracted from the knowledge graph. Monitoring edges in the knowledge graph are used as supervisory signals for feature engineering. A multi-layer perceptron is used to train a small model, normalize the output, and write real-time diagnostic results into the knowledge graph.

[0025] The small model is used as the atomic module for knowledge graph reasoning, and the method is as follows:

[0026] In the diagnostic rules of the fault type node of the knowledge graph, a model relationship is added to point to the corresponding small model node, and the calling interface and input and output definitions are stored; the knowledge graph reasoning engine detects the need to call single-scenario diagnosis, extracts real-time monitoring data through the model relationship, calls the small model API, and writes the small model output as reasoning evidence into the knowledge graph for fusion with the large model; the LSTM model is packaged as a Docker container, exposing the prediction interface, receiving the opening and closing current waveform data, and returning the probability of mechanical jamming.

[0027] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, setting a training strategy for the adversarial learning model and performing adversarial training of the large model and the small model includes:

[0028] The graph structure data is divided into training set, validation set and test set to initialize the large model and small model; the goal of the large model is to determine whether the equipment has a fault and the type of fault, using the cross-entropy loss function and adding a regularization term; the small model uses the cross-entropy loss function, introduces adversarial loss, and sets a comprehensive loss function; for the large model and small model, the Adam optimizer is selected; the batch size and number of training rounds for adversarial training are set, and adversarial training is performed; after each round of training, the performance of the large model and small model is evaluated using the validation set; based on the evaluation results of the validation set, the hyperparameters in the training strategy are adjusted; after training is completed, the large model and small model are tested using the test set, and the trained and tested large model and small model are deployed.

[0029] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, the online monitoring data is obtained, and when a large model output device fails, the historical association rules of the device are queried in the knowledge graph, the historical diagnosis results of the small model of the device in the corresponding scenario are retrieved, and a diagnostic path diagram is generated through Neo4jBloom, including:

[0030] Online monitoring data is acquired and converted into a lightweight format using ONNX, a Transformer-based graph neural network model. This model is deployed on edge nodes to support offline inference. A graph data input tensor is constructed, including node features and an adjacency matrix for the main power equipment. The large model outputs a fault type probability vector. When the confidence level of the highest-probability fault type exceeds a set threshold and the fault type is abnormal, a knowledge graph query process is triggered. A Cypher query is designed to sort query results based on rule update time and expert ratings, prioritizing high-priority rules.

[0031] Filter the diagnostic results of the small model, convert the relationships in the knowledge graph into a structured list; generate a diagnostic path diagram through Neo4jBloom.

[0032] In conjunction with the first aspect, in an eighth implementation of the first aspect of the present application, generating a work order priority based on the diagnostic path map, equipment importance, and fault urgency, and arranging operation and maintenance personnel to perform maintenance, includes:

[0033] Equipment importance is classified from high to low into special-grade equipment, first-grade equipment, and ordinary equipment. The voltage level, load rate, and historical fault count are obtained from the power main equipment entity in the knowledge graph, matched with the equipment importance level, and quantified as the equipment importance coefficient. The fault urgency is classified from high to low into red warning, yellow warning, and blue warning, quantified as the fault urgency coefficient. The confidence of the large model and the small model is calculated, and the work order priority is calculated by combining the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient.

[0034] Based on the diagnostic path diagram, basic equipment information, fault details and visual links are extracted as work order information; skill labels are obtained from the operation and maintenance personnel node in the knowledge graph and matched with the fault type; local operation and maintenance teams are matched according to the equipment location. When there are no idle personnel in the local team, it is upgraded to the superior operation and maintenance center; and the work order information is sent to the corresponding operation and maintenance personnel.

[0035] In conjunction with the first aspect, in a ninth implementation of the first aspect of the present application, calculating the confidence of the large model and the small model, combining the weight distribution of the large model and the small model, and the equipment importance coefficient and the fault urgency coefficient to calculate the work order priority includes:

[0036] Receive the large model inference results and obtain the confidence of the large model ; Perform time-weighted average of the historical confidence of the small model in the same scenario to obtain the confidence of the small model ; Preset the weight of the large model according to business needs and the weight of the small model ; Calculate the work order priority WOP, the formula is as follows:

[0037] ;

[0038] Where I is the equipment importance coefficient and E is the fault urgency coefficient.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. Based on large-scale model technology, this invention can achieve high-quality extraction, understanding, and learning of multimodal data in power production. By constructing an adversarial learning system between a large-parameter model and a small model, the comprehensive analysis capability of the large model and the single-scenario fine diagnosis capability of the small model are fully utilized, effectively improving the accuracy and reliability of diagnosis. At the same time, the use of knowledge graphs for deep mining and reasoning can more accurately identify equipment fault states and reduce misjudgments and missed judgments.

[0041] 2. The present invention constructs a diagnostic knowledge graph with transformers, high-voltage circuit breakers, GIS equipment and other main equipment as core nodes, covering a variety of equipment and diagnostic scenarios; large models and small models work together with the support of the knowledge graph to achieve multi-state comprehensive diagnosis, improve the versatility and comprehensiveness of the model, and can more comprehensively evaluate the health status of power equipment.

[0042] 3. The present invention obtains online monitoring data in real time through edge computing nodes to achieve real-time monitoring and rapid diagnosis of equipment; the adversarial learning of large and small models and the reasoning ability of knowledge graphs enable the system to automatically identify faults and generate diagnostic path diagrams; based on the diagnostic results and equipment information, the work order priority is automatically generated, and operation and maintenance personnel are arranged to carry out repairs, realizing the intelligence of diagnosis and operation and maintenance, and greatly improving the efficiency and timeliness of operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the steps of the method for constructing a health status diagnosis framework of a power main device based on adversarial learning according to the present invention;

[0044] Figure 2 It is a flowchart of the adversarial training program of the method for constructing the health status diagnosis framework of the main power equipment based on adversarial learning of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0047] like Figure 1 As shown in the schematic diagram of the steps of the method for constructing a power master device health status diagnosis framework based on adversarial learning, the present application provides a method for constructing a power master device health status diagnosis framework based on adversarial learning, comprising the following steps:

[0048] Step S100: Determine entities, diagnostic scenario edges, and fault type nodes, and perform entity and relationship modeling; collect multi-source data of power main equipment, perform data cleaning and standardization, extract knowledge from unstructured data, and build a knowledge graph based on the multi-source data of power main equipment;

[0049] Specifically, the entities include a power master device entity, a monitoring indicator entity, and an operation and maintenance personnel entity, and the attributes of the entities are defined; the diagnostic scenario edges include monitoring edges, triggering edges, and participating edges. The monitoring edges represent the relationship between the power master device and the monitoring indicators. The triggering edges indicate that the abnormal values of the monitoring indicators trigger a certain diagnostic scenario. The participating edges indicate that the operation and maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power master device. The diagnostic scenarios include scenarios based on monitoring indicators and scenarios based on fault history. Faults are classified according to their nature, fault type nodes are determined, and fault type attributes are defined.

[0050] Use ER diagrams to represent the conceptual structure between entities, attributes, and relationships, determine the cardinality of relationships between entities, convert the conceptual structure into a logical model, and select the graph database Neo4j for storage; based on the logical model, create corresponding nodes and edges in the graph database.

[0051] Furthermore, we collected multi-source data from main power equipment, processed missing values, outliers, duplicate values, and noise, performed Z-score standardization, encoded text data uniformly into the UTF-8 standard encoding format, and performed text cleaning. For categorized text data, we used one-hot encoding to convert it into numerical data. We also used named entity recognition algorithms to extract knowledge from unstructured data, and trained the BERT model to identify entities in unstructured data and determine the relationships between them.

[0052] Map the processed multi-source data of power main equipment into the knowledge graph.

[0053] In one specific embodiment, the main power equipment entities include: Transformer T1: Equipment number "T1001", model "SCB10-1000 / 10", rated capacity 1000 kVA, rated voltage 10 kV, manufacturer "XX Electric Co., Ltd.", commissioned on "2015-01-01", installed at "XX Substation No. 1 Main Transformer." Circuit breaker CB1: Equipment number "CB001", model "VS1-12 / 630-20", rated current 630 A, rated short-circuit breaking current 20 kA, manufacturer "YY Electric Co., Ltd.", commissioned on "2018-03-15", installed at "XX Substation 10 kV Outgoing Switchgear."

[0054] Monitoring indicator entities include: Transformer oil temperature: The indicator number is "I001," the indicator name is "Transformer oil temperature," the measurement unit is "°C," the normal range is "0-80," and the measurement time is real-time. For example, at a certain moment, the oil temperature of transformer T1 is 50°C. Circuit breaker opening and closing time: The indicator number is "I002," the indicator name is "Circuit breaker opening and closing time," the measurement unit is "ms," the normal range is "20-60," and the measurement time is collected during each opening and closing operation. The opening time of circuit breaker CB1 on a certain occasion was 30ms.

[0055] The O&M personnel entities include: O&M Personnel A: Personnel ID "P001", name "Zhang San", position "Electric Power Equipment O&M Engineer", skill level "Senior", and work experience "8 years". O&M Personnel B: Personnel ID "P002", name "Li Si", position "Electric Power Equipment O&M Technician", skill level "Intermediate", and work experience "5 years".

[0056] A monitoring edge exists between transformer T1 and its oil temperature, expressed as "Transformer T1 - Monitoring - Transformer Oil Temperature," with a monitoring frequency of "15 minutes." A monitoring edge exists between circuit breaker CB1 and its opening and closing times, expressed as "Circuit Breaker CB1 - Monitoring - Breaker Opening and Closing Time," with a monitoring frequency of "Every Operation."

[0057] When the transformer oil temperature exceeds 80°C, the "Transformer Overheat Diagnosis Scenario" is triggered, with the relationship represented as "Transformer Oil Temperature Exceeds 80°C - Trigger - Transformer Overheat Diagnosis Scenario." When the circuit breaker opening and closing time exceeds 60ms or is less than 20ms, the "Circuit Breaker Operation Abnormal Diagnosis Scenario" is triggered, with the relationship represented as "Circuit Breaker Opening and Closing Time Abnormal - Trigger - Circuit Breaker Operation Abnormal Diagnosis Scenario."

[0058] Maintenance personnel A participates in the routine inspection of transformer T1, represented by the relationship "Maintenance personnel A - Participation - Transformer T1 maintenance work," with participation scheduled on Mondays, Wednesdays, and Fridays. Maintenance personnel B participates in the repair of circuit breaker CB1 on May 10, 2023, represented by the relationship "Maintenance personnel B - Participation - Circuit breaker CB1 fault repair (May 10, 2023)."

[0059] Scenarios based on monitoring indicators include: Transformer overheating diagnosis scenario: This scenario is triggered when the transformer oil temperature exceeds the normal range. Circuit breaker operation abnormality diagnosis scenario: This scenario is triggered when the circuit breaker opening and closing time is abnormal.

[0060] Scenarios based on fault history include: if transformer T1 has experienced a winding short-circuit fault before, in subsequent operation, the possibility of the winding short-circuit fault occurring again is diagnosed, forming a "transformer T1 winding short-circuit fault recurrence diagnosis scenario."

[0061] The fault type nodes and attributes are as follows: Electrical Fault: The fault number is "F001," the fault name is "Electrical Fault," and the fault description is "An abnormality has occurred in the electrical part of the power equipment, such as a short circuit or open circuit." The fault probability is "Medium," the fault severity is "High," and the handling recommendation is "Immediately shut down the power for maintenance and check the electrical connections and insulation." Mechanical Fault: The fault number is "F002," the fault name is "Mechanical Fault," and the fault description is "Damage or jamming of mechanical components of the power equipment." The fault probability is "Medium," the fault severity is "Medium," and the handling recommendation is "Inspect, repair, or replace the mechanical components."

[0062] Multi-source data collection and processing is performed, using a named entity recognition algorithm (trained on the BERT model) to extract entities from operation and maintenance logs. For example, from the sentence "On June 1, 2023, operator A inspected transformer T1, and the equipment was operating normally," entities such as "operator A" and "transformer T1" are identified, and the relationship between them is determined as "operator A participated in the operation and maintenance of transformer T1." Based on these entities, relationships, and attributes, corresponding nodes and edges are created in the Neo4j graph database.

[0063] like Figure 2 As shown in the flowchart of the adversarial training program of the method for constructing the health status diagnosis framework of the main power equipment based on adversarial learning, step S200: designing the adversarial learning model architecture, using the Transformer-based graph neural network as the large model and the lightweight model for a single scenario as the small model; converting the knowledge graph into graph structure data, performing graph convolution to generate high-order features, and inputting the high-order features into the large model; using the knowledge graph as the training sample of the small model, feeding the output of the small model back to the knowledge graph, and the small model as the atomic module of the knowledge graph reasoning; setting the training strategy of the adversarial learning model, and performing adversarial training of the large model and the small model;

[0064] Specifically, for large models, the knowledge graph is converted into graph-structured data, where nodes represent entities and edges represent relationships between entities. Node attributes are encoded and converted into vector representations. An adjacency matrix of the graph structure is constructed to represent the connection relationships between nodes. A graph neural network is used as the graph convolution layer to aggregate and update node features. Several graph convolution layers are stacked to capture the high-order structural information of the graph. The Transformer's multi-head self-attention mechanism is introduced into the graph neural network, and the node features output by the graph convolution layer are fused with the output of the Transformer layer to obtain a comprehensive feature representation. For fault type recognition tasks, the output layer uses a softmax function for multi-classification. For fault severity assessment tasks, the output layer uses linear regression.

[0065] For the small model, the lightweight model uses a multi-layer perceptron and uses data of the corresponding scenario in the knowledge graph to train the small model.

[0066] Furthermore, the entities in the knowledge graph are mapped to graph nodes and assigned unique IDs. The attributes of the entities are converted into initial feature vectors of the nodes through numerical conversion. The diagnostic scenario edges in the knowledge graph are converted into graph edges, and the types and weights of graph edges are defined. The types of graph edges include directed edges and undirected edges. A sparse matrix is used to represent the node connection relationship, and non-zero elements indicate the existence of associations between nodes.

[0067] A graph convolutional neural network is designed. The input layer is the initial node features. The GAT layer is the first layer, which aggregates direct neighbor features. The GCN layer is the second layer, which captures second-order neighbor relationships. The output layer is used to output high-order features and perform parameter configuration. For the main power equipment nodes, real-time operating parameters and static attributes are integrated and normalized to [0,1]. For the monitoring indicator nodes, the monitoring values and their threshold ranges are encoded as two-dimensional vectors. Adjacency matrix preprocessing is performed. In the first layer, each main power equipment node aggregates the features of its directly associated monitoring indicator nodes. In the second layer, the main power equipment nodes further aggregate the features of the fault type nodes. The outputs of multiple GATs are spliced or averaged to generate high-order features containing multi-order neighbor information. The high-order features are input into the large model.

[0068] Furthermore, we extract corresponding subgraphs from the knowledge graph based on device type and diagnostic scenario, and use the monitoring edges in the knowledge graph as supervisory signals for feature engineering. We use a multi-layer perceptron to train the small model, normalize the output, and write the real-time diagnostic results into the knowledge graph.

[0069] The small model is used as the atomic module for knowledge graph reasoning, and the method is as follows:

[0070] In the diagnostic rules of the fault type node of the knowledge graph, a model relationship is added to point to the corresponding small model node, and the calling interface and input and output definitions are stored; the knowledge graph reasoning engine detects the need to call single-scenario diagnosis, extracts real-time monitoring data through the model relationship, calls the small model API, and writes the small model output as reasoning evidence into the knowledge graph for fusion with the large model; the LSTM model is packaged as a Docker container, exposing the prediction interface, receiving the opening and closing current waveform data, and returning the probability of mechanical jamming.

[0071] Furthermore, the graph structure data is divided into training set, validation set and test set to initialize the large model and small model; the goal of the large model is to determine whether the equipment has a fault and the type of fault, using the cross-entropy loss function and adding a regularization term; the small model uses the cross-entropy loss function, introduces adversarial loss, and sets a comprehensive loss function; for the large model and small model, the Adam optimizer is selected; the batch size and number of training rounds of adversarial training are set to perform adversarial training; after each round of training, the performance of the large model and small model is evaluated using the validation set; based on the evaluation results of the validation set, the hyperparameters in the training strategy are adjusted; when the training is completed, the large model and small model are tested using the test set, and the trained and tested large model and small model are deployed.

[0072] In one embodiment, the nodes and attributes are encoded as follows:

[0073] Transformer node (T1): Unique ID: T1; Attributes: Rated capacity of 1000 kVA is coded as 0.8 (maximum rated capacity is 1250 kVA); Rated voltage of 10 kV is coded as 0.5 (maximum rated voltage is 20 kV); Real-time oil temperature of 60°C is coded as 0.75 (normal oil temperature range is 0-80°C). The initial eigenvector is [0.8, 0.5, 0.75].

[0074] Circuit breaker node (CB1): Unique ID: CB1; Attributes: Rated current 630A is encoded as 0.63 (maximum rated current 1000A); Open / Close time 35ms is encoded as 0.6 (normal range 20-60ms). Initial eigenvector is [0.63, 0.6].

[0075] Monitoring indicator node (oil temperature): Unique ID: I1; Attributes: Monitoring value 60°C, threshold range [0,80]°C, coded as [0.75,0.8].

[0076] Monitoring indicator node (opening and closing time): Unique ID: I2; Attributes: Monitoring value 35ms, threshold range [20,60]ms, coded as [0.6,0.8].

[0077] Fault type node (overheating fault): Unique ID: F1; Attributes: Fault probability 0.3 is encoded as 0.3; Fault severity rating 2 (severity range 1-5) is encoded as 0.4. Initial feature vector is [0.3, 0.4].

[0078] Fault type node (abnormal opening and closing fault): Unique ID: F2; Attributes: Fault probability 0.2 is coded as 0.2; Fault severity assessment 1.5 is coded as 0.3. Initial feature vector is [0.2, 0.3].

[0079] Transformer T1 has a monitoring edge with oil temperature monitoring indicator I1. The edge type is directed (from T1 to I1) and the weight is 0.8.

[0080] There is a monitoring edge between circuit breaker CB1 and the opening and closing time monitoring indicator I2. The edge type is directed edge (from CB1 to I2) and the weight is 0.9.

[0081] The abnormal oil temperature monitoring indicator I1 (exceeding 80°C) triggers the overheating fault F1: the edge type is a directed edge (from I1 to F1) with a weight of 0.7.

[0082] The abnormality of the opening and closing time monitoring indicator I2 (less than 20ms or greater than 60ms) triggers the opening and closing abnormal fault F2: the edge type is a directed edge (from I2 to F2) with a weight of 0.8.

[0083] Generate an adjacency matrix.

[0084] The graph convolutional neural network parameters are configured as follows: Input layer: initial node feature vectors. GAT layer (first layer): number of heads is 3, and the number of hidden units in each head is 16. GCN layer (second layer): number of hidden units is 32. Output layer: output dimension is 64.

[0085] After the first GAT layer, transformer T1 aggregates the features of oil temperature monitoring indicator I1, and circuit breaker CB1 aggregates the features of opening and closing time monitoring indicator I2. After the second GCN layer, transformer T1 further aggregates the features of overheating fault F1, and circuit breaker CB1 further aggregates the features of abnormal opening and closing fault F2. Examples of high-order features obtained by concatenating multiple GAT outputs are as follows: transformer T1's high-order features: [0.2, 0.3, 0.1, ...] (64-dimensional vector), circuit breaker CB1's high-order features: [0.1, 0.4, 0.2, ...] (64-dimensional vector).

[0086] A small model was trained to extract a subgraph from the knowledge graph containing transformer T1, oil temperature monitoring indicator I1, and overheating fault F1. The oil temperature monitoring edge was used as the supervisory signal, and the input features obtained through feature engineering were [0.75, 0.8] (encoding the oil temperature monitoring value and threshold range). A subgraph containing circuit breaker CB1, opening and closing time monitoring indicator I2, and abnormal opening and closing fault F2 was extracted. The input features were [0.6, 0.8] (encoding the opening and closing time monitoring value and threshold range).

[0087] The multilayer perceptron architecture consists of two neurons in the input layer, 16 neurons in the hidden layer, and two neurons in the output layer (corresponding to the probability of a fault occurring or not occurring). Training data: 100 transformer overheating scenario samples and 100 circuit breaker opening and closing abnormality scenario samples were extracted from the knowledge graph and divided into training, validation, and test sets with a ratio of 8:1:1. Training was performed using the Adam optimizer with a learning rate of 0.001 and 50 training rounds.

[0088] The graph data (including all node and edge information) was split into training, validation, and test sets with a ratio of 7:2:1. Adversarial training was performed on the large and small models. The batch size was 16, and the number of training epochs was 100. The large model used the cross-entropy loss function with a regularization coefficient of 0.01, while the small model used the cross-entropy loss function with an adversarial loss weight of 0.2. The learning rate of the Adam optimizer was 0.001.

[0089] In each round of training, the small model is trained first, and then the output of the small model is used as the input to train the large model. The performance of the large and small models is evaluated every 10 rounds using the validation set, and hyperparameters such as the learning rate are adjusted based on the evaluation results. A final test of the large and small models is performed using the test set, and metrics such as accuracy and recall are recorded. The trained and tested large and small models are then deployed in a real-world system for power equipment fault diagnosis.

[0090] A relationship pointing to the small model for the transformer overheating scenario is added to the diagnostic rule for overheating fault F1, storing the call interface and input and output definitions. When the knowledge graph inference engine detects that transformer T1 needs to be diagnosed for overheating, it extracts the oil temperature monitoring data [0.75, 0.8] and calls the small model API. The small model returns a probability of 0.2 for the overheating fault. This result is written into the knowledge graph as inference evidence for integration with the large model.

[0091] Step S300: Obtain online monitoring data. When a large model output device fails, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnosis results of the small model of the device in the corresponding scenario, and generate a diagnostic path diagram through Neo4jBloom.

[0092] Specifically, online monitoring data is acquired, and the Transformer-based graph neural network model is converted into a lightweight format using ONNX. This is then deployed on edge nodes to support offline inference. A graph data input tensor is constructed, including the node features and adjacency matrix of the main power equipment. The large model outputs a fault type probability vector. When the confidence level of the highest-probability fault type exceeds a set threshold and the fault type is abnormal, the knowledge graph query process is triggered. A Cypher query is designed to sort the query results based on the rule update time and expert ratings, prioritizing high-priority rules.

[0093] Filter the diagnostic results of the small model, convert the relationships in the knowledge graph into a structured list; generate a diagnostic path diagram through Neo4jBloom.

[0094] In one specific embodiment, real-time online monitoring data for transformer T001 is obtained through edge nodes, including oil temperature (current value 85°C), winding temperature (current value 90°C), and load factor (current value 88%). A Transformer-based graph neural network model is converted to the ONNX lightweight format and deployed on the edge node. A graph data input tensor is constructed, with the features of the main power equipment node being [normalized oil temperature value (0.85, oil temperature range 0-100°C), normalized winding temperature value (0.9, winding temperature range 0-100°C), normalized load factor value (0.88, load factor range 0-100%)]. The adjacency matrix contains the connection weights between transformer T001 and other relevant monitoring indicator nodes (such as oil temperature sensor nodes and winding temperature sensor nodes). The edge weight between the oil temperature sensor and the transformer is 0.9, and the edge weight between the winding temperature sensor and the transformer is 0.8). The large model performs offline inference and outputs a fault type probability vector: [overheating fault: 0.8, insulation aging fault: 0.1, normal: 0.1]. Because the confidence level of 0.8 for the overheating fault is greater than the set threshold (0.7) and this fault type is not normal, the knowledge graph query process is triggered.

[0095] Perform a knowledge graph query, and the query results are:

[0096] Association rules: Rule 1: When the oil temperature exceeds 80°C and the load factor exceeds 80%, an overheating warning is triggered (Updated: 2024-10-01, Expert Rating: 8 points). Rule 2: When the winding temperature exceeds 95°C, an overheating fault may occur (Updated: 2023-05-15, Expert Rating: 6 points).

[0097] Small model diagnosis results: Small model diagnosis result 1: Diagnosis performed on November 10, 2024, with an overheating fault confidence level of 0.75 (derived from the small model). Small model diagnosis result 2: Diagnosis performed on August 20, 2024, with an overheating fault confidence level of 0.6 (derived from the small model).

[0098] Filter the diagnosis results of the small model, retaining only those with a confidence level greater than 0.6. After filtering, retain the diagnosis performed on November 10, 2024 (overheating fault confidence level 0.75). Convert the relationships in the knowledge graph into a structured list.

[0099] Neo4jBloom generates a diagnostic path graph based on the filtered and processed information. The graph includes a transformer T001 node, an overheating fault node, related monitoring indicator nodes (such as oil temperature, winding temperature, and load factor), a diagnostic rule node (when the oil temperature exceeds 80°C and the load factor exceeds 80%, an overheating warning is triggered), and a small model diagnosis result node (diagnosis performed on November 10, 2024, overheating fault confidence level 0.75). Edges between nodes represent relationships, such as the monitoring relationship between transformer T001 and the oil temperature node, the fault triggering relationship between transformer T001 and the overheating fault node, the application rule relationship between transformer T001 and the diagnostic rule node, and the small model diagnosis relationship between transformer T001 and the small model diagnosis result node.

[0100] Step S400: Based on the diagnostic path map, equipment importance and fault urgency, a work order priority is generated and operation and maintenance personnel are arranged to perform maintenance.

[0101] Specifically, the importance of equipment is classified from high to low into special-grade equipment, first-grade equipment, and ordinary equipment. The voltage level, load rate, and historical fault count are obtained from the power main equipment entity in the knowledge graph, matched with the equipment importance level, and quantified as the equipment importance coefficient. The urgency of the fault is classified from high to low into red warning, yellow warning, and blue warning, and quantified as the fault urgency coefficient. The confidence of the large model and the small model is calculated, and the work order priority is calculated by combining the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient.

[0102] Based on the diagnostic path diagram, basic equipment information, fault details and visual links are extracted as work order information; skill labels are obtained from the operation and maintenance personnel node in the knowledge graph and matched with the fault type; local operation and maintenance teams are matched according to the equipment location. When there are no idle personnel in the local team, it is upgraded to the superior operation and maintenance center; and the work order information is sent to the corresponding operation and maintenance personnel.

[0103] Furthermore, the inference results of the large model are received to obtain the confidence of the large model. ; Perform time-weighted average of the historical confidence of the small model in the same scenario to obtain the confidence of the small model ; Preset the weight of the large model according to business needs and the weight of the small model ; Calculate the work order priority WOP, the formula is as follows:

[0104] ;

[0105] Where I is the equipment importance coefficient and E is the fault urgency coefficient.

[0106] In a specific embodiment, relevant information about transformer T001 is obtained from the knowledge graph:

[0107] Voltage level: 110kV (110kV corresponds to level 1 equipment in the equipment importance classification rules); load rate: 88% (relatively high); number of historical failures: 2 times (in the past year).

[0108] Based on the equipment importance classification rules (special-grade equipment is 500kV and above main equipment; first-grade equipment is 110kV-220kV main equipment with a high load rate or a high number of historical failures; ordinary equipment is auxiliary equipment below 110kV), transformer T001 is classified as first-grade equipment. Quantifying the equipment importance coefficient I: First-grade equipment corresponds to an I=1.2.

[0109] The fault type probability vector for the large-scale model output transformer T001 is [overheating fault: 0.8, insulation aging fault: 0.1, normal: 0.1], and the overheating fault confidence level is 0.8. Based on the fault urgency grading rules (red alert: large-scale model confidence level ≥ 0.9 or three consecutive small-scale model alerts; yellow alert: large-scale model confidence level 0.8-0.9 or high small-scale model confidence in a single scenario; blue alert: large-scale model confidence level 0.7-0.8 or the first occurrence of an abnormal signal), this fault is classified as a yellow alert. The fault urgency coefficient E is quantified: the yellow alert corresponds to an E=1.5.

[0110] After receiving the inference results of the large model, it is found that the large model's confidence in the overheating fault is 0.8. The historical confidences of the small model in the overheating fault scenario are: the diagnosis performed on 2024-11-10, the overheating fault confidence is 0.75, and the diagnosis performed on 2024-08-20, the overheating fault confidence is 0.6.

[0111] The time-weighted average of the historical confidence of the small model (time decay factor: within the last month = 1.0, within 2-3 months = 0.8, more than 3 months = 0.5) yields a result of 0.615.

[0112] According to business requirements, the preset weight of the large model is 0.6, the weight of the small model is 0.4, and the calculated work order priority WOP is 1.3068.

[0113] Based on the diagnostic path diagram, basic equipment information (such as equipment model and installation location), fault details (large model diagnosis of overheating fault with a confidence level of 0.8; small model historical diagnostic results, etc.) and visualization links (diagnostic path diagram link generated by Neo4jBloom) of transformer T001 are extracted as work order information.

[0114] The skill tag is obtained from the operator node in the knowledge graph and matched with the overheating fault type to find an operator with the skills to handle transformer overheating faults. For example, operator A (skill tags: Transformer Fault Diagnosis, Overheating Fault Handling) is matched based on the equipment location (Transformer T001 is located at XX substation). If operator A is available in the local operation and maintenance team, the work order information is sent to operator A.

[0115] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for constructing a health status diagnosis framework for power main equipment based on adversarial learning, characterized in that: The following steps are involved: Determine entities, diagnostic scenario edges, and fault type nodes, and perform entity and relationship modeling; Collect multi-source data from main power equipment, perform data cleaning and standardization, extract knowledge from unstructured data, and build a knowledge graph based on the multi-source data from main power equipment; Design an adversarial learning model architecture, using a Transformer-based graph neural network as the large model and a lightweight model for a single scenario as the small model. Convert the knowledge graph into graph-structured data, perform graph convolution to generate high-order features, and input the high-order features into the large model. Use the knowledge graph as training samples for the small model, feed the output of the small model back into the knowledge graph, and use the small model as the atomic module for knowledge graph reasoning. Set the training strategy for the adversarial learning model and perform adversarial training between the large and small models. For large models, the knowledge graph is converted into graph-structured data, where nodes represent entities and edges represent relationships between entities. Node attributes are encoded and converted into vector representations. An adjacency matrix of the graph structure is constructed to represent the connection relationships between nodes. A graph neural network is used as the graph convolution layer to aggregate and update node features. Several graph convolution layers are stacked to capture the high-order structural information of the graph. The Transformer's multi-head self-attention mechanism is introduced into the graph neural network, and the node features output by the graph convolution layer are fused with the output of the Transformer layer to obtain a comprehensive feature representation. For fault type identification tasks, the output layer uses a softmax function for multi-classification. For fault severity assessment tasks, the output layer uses linear regression. For the small model, the lightweight model uses a multi-layer perceptron and uses the data of the corresponding scene in the knowledge graph to train the small model; Obtain online monitoring data. When a large model output device fails, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnosis results of the small model of the device in the corresponding scenario, and generate a diagnostic path diagram through Neo4jBloom. Based on the diagnostic path map, equipment importance, and fault urgency, work order priorities are generated and maintenance personnel are assigned to perform repairs.

2. The method for constructing a health status diagnosis framework for power main equipment based on adversarial learning according to claim 1 is characterized in that: The determining of entities, diagnosis scenario edges, and fault type nodes, and entity and relationship modeling, includes: The entities include a power master device entity, a monitoring indicator entity, and an operation and maintenance personnel entity, and the attributes of the entities are defined; the diagnostic scenario edges include monitoring edges, triggering edges, and participating edges. The monitoring edges represent the relationship between the power master device and the monitoring indicators. The triggering edges indicate that the abnormal values of the monitoring indicators trigger a certain diagnostic scenario. The participating edges indicate that the operation and maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power master device. The diagnostic scenarios include scenarios based on monitoring indicators and scenarios based on fault history. Faults are classified according to their nature, fault type nodes are determined, and fault type attributes are defined. Use ER diagrams to represent the conceptual structure between entities, attributes, and relationships, determine the cardinality of relationships between entities, convert the conceptual structure into a logical model, and select the graph database Neo4j for storage; based on the logical model, create corresponding nodes and edges in the graph database.

3. The method for constructing a power main equipment health status diagnosis framework based on adversarial learning according to claim 1 is characterized in that: The method includes collecting multi-source data of power main equipment, performing data cleaning and standardization, extracting knowledge from unstructured data, and building a knowledge graph based on the multi-source data of power main equipment, including: Collect multi-source data from main power equipment, process missing values, outliers, duplicate values, and noise, perform Z-score standardization, uniformly encode text data into the UTF-8 standard encoding format, perform text cleaning, and use one-hot encoding to convert categorical text data into numerical data. Use named entity recognition algorithms to extract knowledge from unstructured data, and use the BERT model for training to identify entities in unstructured data and determine the relationships between entities. Map the processed multi-source data of power main equipment into the knowledge graph.

4. The method for constructing a health status diagnosis framework for power main equipment based on adversarial learning according to claim 1, characterized in that: The process of converting the knowledge graph into graph structure data, performing graph convolution to generate high-order features, and inputting the high-order features into the large model includes: Entities in the knowledge graph are mapped to graph nodes, assigned unique IDs, and the entity attributes are converted into initial feature vectors of the nodes through numerical conversion. Diagnostic scenario edges in the knowledge graph are converted into graph edges, and the types and weights of graph edges are defined. Graph edge types include directed edges and undirected edges. Node connection relationships are represented using sparse matrices, with non-zero elements indicating associations between nodes. A graph convolutional neural network is designed. The input layer is the initial node features. The GAT layer is the first layer, which aggregates direct neighbor features. The GCN layer is the second layer, which captures second-order neighbor relationships. The output layer is used to output high-order features and perform parameter configuration. For the main power equipment nodes, real-time operating parameters and static attributes are integrated and normalized to [0,1]. For the monitoring indicator nodes, the monitoring values and their threshold ranges are encoded as two-dimensional vectors. Adjacency matrix preprocessing is performed. In the first layer, each main power equipment node aggregates the features of its directly associated monitoring indicator nodes. In the second layer, the main power equipment nodes further aggregate the features of the fault type nodes. The outputs of multiple GATs are spliced or averaged to generate high-order features containing multi-order neighbor information. The high-order features are input into the large model.

5. The method for constructing a power main equipment health status diagnosis framework based on adversarial learning according to claim 1, characterized in that: The knowledge graph is used as a training sample for the small model, the output of the small model is fed back to the knowledge graph, and the small model serves as an atomic module for knowledge graph reasoning, including: By device type and diagnostic scenario, corresponding subgraphs are extracted from the knowledge graph. Monitoring edges in the knowledge graph are used as supervisory signals for feature engineering. A multi-layer perceptron is used to train a small model, normalize the output, and write real-time diagnostic results into the knowledge graph. The small model is used as the atomic module for knowledge graph reasoning, and the method is as follows: In the diagnostic rules of the fault type node of the knowledge graph, a model relationship is added to point to the corresponding small model node, and the calling interface and input and output definitions are stored; the knowledge graph reasoning engine detects the need to call single-scenario diagnosis, extracts real-time monitoring data through the model relationship, calls the small model API, and writes the small model output as reasoning evidence into the knowledge graph for fusion with the large model; the LSTM model is packaged as a Docker container, exposing the prediction interface, receiving the opening and closing current waveform data, and returning the probability of mechanical jamming.

6. The method for constructing a power main equipment health status diagnosis framework based on adversarial learning according to claim 1, characterized in that: The training strategy of setting the adversarial learning model and performing adversarial training of the large model and the small model includes: The graph structure data is divided into training set, validation set and test set to initialize the large model and small model; the goal of the large model is to determine whether the equipment has a fault and the type of fault, using the cross-entropy loss function and adding a regularization term; the small model uses the cross-entropy loss function, introduces adversarial loss, and sets a comprehensive loss function; for the large model and small model, the Adam optimizer is selected; the batch size and number of training rounds for adversarial training are set, and adversarial training is performed; after each round of training, the performance of the large model and small model is evaluated using the validation set; based on the evaluation results of the validation set, the hyperparameters in the training strategy are adjusted; after training is completed, the large model and small model are tested using the test set, and the trained and tested large model and small model are deployed.

7. The method for constructing a power main equipment health status diagnosis framework based on adversarial learning according to claim 1, characterized in that: The online monitoring data is obtained. When a large model output device fails, the historical association rules of the device in the knowledge graph are queried, the historical diagnosis results of the small model of the device in the corresponding scenario are retrieved, and a diagnostic path diagram is generated through Neo4jBloom, including: Online monitoring data is acquired and converted into a lightweight format using ONNX, a Transformer-based graph neural network model. This model is deployed on edge nodes to support offline inference. A graph data input tensor is constructed, including node features and an adjacency matrix for the main power equipment. The large model outputs a fault type probability vector. When the confidence level of the highest-probability fault type exceeds a set threshold and the fault type is abnormal, a knowledge graph query process is triggered. A Cypher query is designed to sort query results based on rule update time and expert ratings, prioritizing high-priority rules. Filter the diagnostic results of the small model, convert the relationships in the knowledge graph into a structured list; generate a diagnostic path diagram through Neo4jBloom.

8. The method for constructing a power main equipment health status diagnosis framework based on adversarial learning according to claim 1, characterized in that: Based on the diagnostic path map, equipment importance, and fault urgency, the work order priority is generated and maintenance personnel are arranged for repairs, including: Equipment importance is classified from high to low into special-grade equipment, first-grade equipment, and ordinary equipment. The voltage level, load rate, and historical fault count are obtained from the power main equipment entity in the knowledge graph, matched with the equipment importance level, and quantified as the equipment importance coefficient. The fault urgency is classified from high to low into red warning, yellow warning, and blue warning, quantified as the fault urgency coefficient. The confidence of the large model and the small model is calculated, and the work order priority is calculated by combining the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient. Based on the diagnostic path diagram, basic equipment information, fault details and visual links are extracted as work order information; skill labels are obtained from the operation and maintenance personnel node in the knowledge graph and matched with the fault type; local operation and maintenance teams are matched according to the equipment location. When there are no idle personnel in the local team, it is upgraded to the superior operation and maintenance center; and the work order information is sent to the corresponding operation and maintenance personnel.

9. The method for constructing a health status diagnosis framework for power main equipment based on adversarial learning according to claim 8, characterized in that: The confidence of the large model and the small model is calculated, and the work order priority is calculated by combining the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient, including: Receive the large model inference results and obtain the confidence of the large model ; Perform time-weighted average of the historical confidence of the small model in the same scenario to obtain the confidence of the small model ; Preset the weight of the large model according to business needs and the weight of the small model ; Calculate the work order priority WOP, the formula is as follows: ； Where I is the equipment importance coefficient and E is the fault urgency coefficient.

Images