Electric power primary equipment health state diagnosis framework construction method based on adversarial learning

Abstract

The invention discloses an adversarial learning-based electric power primary equipment health state diagnosis framework construction method, and belongs to the technical field of electric power equipment diagnosis. According to the method, entity and relation modeling is carried out; collecting multi-source data of the power master device, and constructing a knowledge graph based on the multi-source data of the power master device; designing an adversarial learning model architecture, and taking a lightweight model for a single scene as a 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; taking the knowledge graph as a training sample of the small model, and feeding back the output of the small model to the knowledge graph; setting a training strategy of the adversarial learning model, and carrying out adversarial training; obtaining online monitoring data, when a large model output device fails, querying a historical association rule of the device in the knowledge graph, calling a small model historical diagnosis result of the device in a corresponding scene, and generating a diagnosis path graph; and generating a work order priority, and arranging operation and maintenance personnel to carry out maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment diagnosis, and specifically 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, there are numerous and diverse power equipment, and their health status is directly related to the safe and stable operation of the power grid. Once a device fails, it may trigger large-scale power outages. Traditional power equipment diagnosis methods mainly rely on manual experience and simple monitoring means, and it is difficult to meet the current power grid development needs.

[0003] Traditional power equipment health status diagnosis relies to a large extent on manual analysis. Single-state diagnosis mainly uses simple methods such as threshold analysis and trend analysis according to regulations, and it is difficult to accurately judge complex faults. Existing diagnosis models often target single equipment or single fault types, lacking generality and comprehensiveness. The correlation information between different equipment and fault types cannot be effectively utilized, and it is difficult to accurately evaluate and predict the overall health status of power equipment. Traditional diagnosis methods cannot achieve real-time and dynamic monitoring and diagnosis, and it is difficult to detect potential fault hazards of equipment in a timely manner. In terms of operation and maintenance management, there is also a lack of intelligent decision support, resulting in low operation and maintenance efficiency and unable to meet the high efficiency and timeliness requirements of modern power grids for 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 for power main equipment based on adversarial learning to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: In the first aspect, the present application provides a method for constructing a health status diagnosis framework for power main equipment based on adversarial learning, including the following steps: 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 construct a knowledge graph based on the multi-source data of power main equipment; Design an adversarial learning model architecture, use a graph neural network based on Transformer as the large model, and use 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 the training sample of the small model, feedback the output of the small model to the knowledge graph, and use the small model as the atomic module for knowledge graph reasoning; set the training strategy of the adversarial learning model and perform adversarial training on the large model and the small model; Obtain online monitoring data. When a fault occurs in the device output by the large model, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnostic results of the small model of the device in the corresponding scenario, and generate a diagnostic path graph through Neo4jBloom; Based on the diagnostic path graph, device importance, and fault urgency, generate a work order priority and arrange maintenance personnel for repair.

[0006] Combined with the first aspect, in the first implementation manner of the first aspect of this application, it is characterized in that, the determining entities, diagnostic scenario edges, and fault type nodes, and performing entity and relationship modeling includes: The entities include power main equipment entities, monitoring index entities, and maintenance personnel entities, and define the attributes of the entities; the diagnostic scenario edges include monitoring edges, trigger edges, and participation edges. The monitoring edge represents the relationship between the power main equipment and the monitoring index, the trigger edge represents that the abnormal value of the monitoring index triggers a certain diagnostic scenario, and the participation edge represents that the maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power main equipment. The diagnostic scenarios include scenarios based on monitoring indexes and scenarios based on fault history; classify faults according to fault nature, determine fault type nodes, and define fault type attributes; Use an ER diagram to represent the conceptual structure among entities, attributes, and relationships, determine the cardinality of the relationships between entities, convert the conceptual structure into a logical model, and select the graph database Neo4j for storage; according to the logical model, create corresponding nodes and edges in the graph database.

[0007] Combined with the first aspect, in the second implementation manner of the first aspect of this application, the collecting multi-source data of power main equipment, performing data cleaning and standardization, extracting knowledge from unstructured data, and constructing a knowledge graph based on the multi-source data of power main equipment includes: Collect multi-source data of power main equipment, perform missing value processing, outlier processing, duplicate value processing, and noise processing, perform Z-score standardization, uniformly encode text data into UTF-8 standard encoding format, perform text cleaning, and for categorical text data, use one-hot encoding to convert it into numerical data; use a named entity recognition algorithm to extract knowledge from unstructured data, use a BERT model for training, identify entities in unstructured data, and determine the relationships between entities; Map the processed multi-source data of power main equipment to the knowledge graph.

[0008] Combined with the first aspect, in the third implementation manner of the first aspect of this application, the designing an adversarial learning model architecture, using a graph neural network based on Transformer as the large model, and using a lightweight model for a single scenario as the small model includes: For large models, the knowledge graph is transformed into graph-structured data. In the graph-structured data, nodes represent entities and edges represent the relationships between entities. Encode the attributes of the nodes and convert them into vector representations. Construct the adjacency matrix of the graph structure to represent the connection relationships between nodes. Use a graph neural network as the graph convolutional layer to aggregate and update the node features. Stack several graph convolutional layers to capture the high-order structure information of the graph. Introduce the multi-head self-attention mechanism of the Transformer into the graph neural network, and fuse the node features output by the graph convolutional layer with the output of the Transformer layer to obtain a comprehensive feature representation. For the fault type recognition task, the output layer uses the softmax function for multi-classification. For the fault severity assessment task, the output layer uses linear regression. For small models, the lightweight model uses a multi-layer perceptron and is trained using the data corresponding to the scenario in the knowledge graph.

[0009] Combined with the first aspect, in the fourth implementation manner of the first aspect of this application, the transformation of the knowledge graph into graph-structured data, performing graph convolution to generate high-order features, and inputting the high-order features into the large model includes: Map the entities in the knowledge graph to graph nodes, assign unique IDs, and numerically convert the attributes of the entities into the initial feature vectors of the nodes. Convert the diagnostic scenario edges in the knowledge graph into graph edges, define the types and weights of the graph edges, and the types of graph edges include directed edges and undirected edges. Use a sparse matrix to represent the node connection relationships, and non-zero elements indicate the existence of associations between nodes. Design a graph convolutional neural network. The input layer is the initial node features, the GAT layer is the first layer to aggregate the direct neighbor features, the GCN layer is the second layer to capture the second-order neighbor relationships, and the output layer is used to output the high-order features and perform parameter configuration. For the power main equipment nodes, fuse the real-time operation parameters and static attributes and normalize them to [0, 1]. For the monitoring index nodes, encode the monitoring values and their threshold ranges into two-dimensional vectors. Perform preprocessing on the adjacency matrix. In the first layer, each power main equipment node aggregates the features of the monitoring index nodes directly associated with it. In the second layer, the power main equipment nodes further aggregate the features of the fault type nodes. Concatenate or average the outputs of the multi-head GAT to generate high-order features containing multi-order neighbor information. Input the high-order features into the large model.

[0010] Combined with the first aspect, in the fifth implementation manner of the first aspect of this application, using the knowledge graph as the training sample of the small model, and feeding back the output of the small model to the knowledge graph, and the small model as the atomic module for knowledge graph reasoning includes: Extract corresponding subgraphs from the knowledge graph according to the device type and diagnostic scenario, use the monitoring edges in the knowledge graph as supervision signals for feature engineering; train small models using a multi-layer perceptron, perform output normalization, and write the real-time diagnostic results into the knowledge graph; The small model is used as an atomic module for knowledge graph reasoning, and the method is as follows: In the diagnostic rules of the fault type nodes in the knowledge graph, add model relationships that point to the corresponding small model nodes, and store the call interfaces and input / output definitions; when the knowledge graph inference engine detects that single-scenario diagnosis needs to be called, extract real-time monitoring data through the model relationship, call the small model API, and write the output of the small model into the knowledge graph as inference evidence for the large model to fuse; package the LSTM model into a Docker container, expose the prediction interface, receive the closing and opening current waveform data, and return the probability of mechanical jamming.

[0011] Combined with the first aspect, in the sixth implementation manner of the first aspect of the present application, the training strategy of the set adversarial learning model is used to perform adversarial training on the large model and the small model, including: Divide the graph structure data into training sets, validation sets, and test sets, and initialize the large model and the small model; the goal of the large model is to judge whether the device has a fault and the type of the fault, use the cross-entropy loss function, and add a regularization term; the small model uses the cross-entropy loss function, introduces an adversarial loss, and sets a comprehensive loss function; for the large model and the small model, select the Adam optimizer; set the batch size and the number of training rounds for adversarial training, and perform adversarial training; after each round of training, evaluate the performance of the large model and the small model using the validation set; according to the evaluation results of the validation set, adjust the hyperparameters in the training strategy; when the training is completed, test the large model and the small model using the test set, and deploy the trained and tested large model and small model.

[0012] Combined with the first aspect, in the seventh implementation manner of the first aspect of the present application, the online monitoring data is obtained. When the large model outputs that the device has a fault, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnostic results of the small model of the device in the corresponding scenario, and generate a diagnostic path graph through Neo4jBloom, including: Obtain online monitoring data, convert the graph neural network model based on Transformer into a lightweight format through ONNX, and deploy it on the edge node to support offline reasoning; construct a graph data input tensor, including: power main equipment node features and adjacency matrix; the output of the large model is a fault type probability vector. When the confidence level of the fault type with the highest probability is greater than the set threshold and the fault type is abnormal, trigger the knowledge graph query process; design a Cypher query, sort the query results according to the update time and expert score of the rules, and preferentially display 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 graph through Neo4jBloom.

[0013] Combined with the first aspect, in the eighth implementation manner of the first aspect of this application, generating a work order priority based on the diagnostic path graph, equipment importance, and fault urgency, and arranging maintenance personnel for repair includes: The grading of equipment importance from high to low includes special-level equipment, first-level equipment, and ordinary equipment. Obtain the voltage level, load rate, and historical fault times from the power main equipment entity in the knowledge graph, match the level of equipment importance, and quantify it as an equipment importance coefficient; the fault urgency from high to low includes red warning, yellow warning, and blue warning, and is quantified as a fault urgency coefficient; calculate the confidence levels of the large model and the small model, combine the weight distributions of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient, and calculate the work order priority. Based on the diagnostic path graph, extract equipment basic information, fault details, and visualization links as work order information; obtain skill tags from the maintenance personnel node in the knowledge graph and match them with the fault types; match the local maintenance team according to the equipment location, and when there are no idle personnel in the local team, escalate to the superior maintenance center; send the work order information to the corresponding maintenance personnel.

[0014] Combined with the first aspect, in the ninth implementation manner of the first aspect of this application, calculating the confidence levels of the large model and the small model, combining the weight distributions of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient, and calculating the work order priority includes: Receive the inference result of the large model to obtain the confidence level of the large model ; perform time-weighted averaging on the historical confidence levels of the small model in the same scenario to obtain the confidence level of the small model ; preset the weight of the large model according to business requirements and the weight of the small model ; calculate the work order priority WOP, and the formula is as follows: ; where I is the equipment importance coefficient and E is the fault urgency coefficient.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. The present invention is based on large model technology and can achieve high-quality extraction, understanding, and learning of multimodal data in power production. By constructing an adversarial learning system of large parameter models and small models, the comprehensive analysis ability of large models and the single-scene fine diagnosis ability of small models are fully utilized, effectively improving the accuracy and reliability of diagnosis. At the same time, through in-depth mining and reasoning using knowledge graphs, the fault states of equipment can be more accurately identified, reducing misjudgment and missed judgment.

[0016] 2. The present invention constructs a diagnostic knowledge graph with main equipment such as transformers, high-voltage circuit breakers, and GIS equipment as core nodes, covering a variety of equipment and diagnostic scenarios. The large model and small model work together with the support of the knowledge graph to achieve multi-state comprehensive diagnosis, improving the generality and comprehensiveness of the model, and being able to more comprehensively evaluate the health status of power equipment.

[0017] 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 the large model and small model and the reasoning ability of the knowledge graph enable the system to automatically identify faults and generate a diagnostic path map. Based on the diagnostic results and equipment information, the work order priority is automatically generated, and maintenance personnel are arranged for repair, realizing the intelligence of diagnosis and operation and maintenance, and greatly improving the operation and maintenance efficiency and timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of the steps of the method for constructing a health status diagnosis framework for power main equipment based on adversarial learning of the present invention; Figure 2 is a flowchart of the adversarial training program of the method for constructing a health status diagnosis framework for power main equipment based on adversarial learning of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Embodiment: As Figure 1 - Figure 2 shown, the present invention provides a technical solution, As Figure 1 shown in the schematic diagram of the steps of the method for constructing a health status diagnosis framework for power main equipment based on adversarial learning, the present application provides a method for constructing a health status diagnosis framework for power main equipment based on adversarial learning, including the following steps: 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 construct a knowledge graph based on the multi-source data of power main equipment; Specifically, the entities include power main equipment entities, monitoring index entities, and operation and maintenance personnel entities, and define the attributes of the entities; the diagnostic scenario edges include monitoring edges, trigger edges, and participation edges. The monitoring edge represents the relationship between the power main equipment and the monitoring index, the trigger edge represents that the abnormal value of the monitoring index triggers a certain diagnostic scenario, and the participation edge represents that the operation and maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power main equipment. The diagnostic scenarios include scenarios based on monitoring indexes and scenarios based on fault histories; classify faults according to fault natures, determine fault type nodes, and define fault type attributes; Use an ER diagram to represent the conceptual structure among entities, attributes, and relationships, determine the cardinality of the relationships between entities, transform the conceptual structure into a logical model, and select the graph database Neo4j for storage; create corresponding nodes and edges in the graph database according to the logical model.

[0021] Furthermore, collect multi-source data of power main equipment, perform missing value processing, outlier processing, duplicate value processing, and noise processing, perform Z-score standardization, uniformly encode the text data into the UTF-8 standard encoding format, perform text cleaning, and for categorical text data, use one-hot encoding to convert it into numerical data; use a named entity recognition algorithm to extract knowledge from unstructured data, use the BERT model for training, identify entities in the unstructured data, and determine the relationships between entities; Map the processed multi-source data of power main equipment to the knowledge graph.

[0022] In a specific embodiment, the power main equipment entities include: Transformer T1: The equipment number is "T1001", the model is "SCB10-1000 / 10", the rated capacity is 1000 kVA, the rated voltage is 10 kV, the manufacturer is "XX Electric Co., Ltd.", the commissioning time is "2015-01-01", and the installation location is "No. 1 Main Transformer of XX Substation". Circuit Breaker CB1: The equipment number is "CB001", the model is "VS1-12 / 630-20", the rated current is 630 A, the rated short-circuit breaking current is 20 kA, the manufacturer is "YY Electric Co., Ltd.", the commissioning time is "2018-03-15", and the installation location is "10 kV outgoing switchgear of XX Substation".

[0023] Monitoring indicator entities include: Transformer oil temperature: indicator number is "I001", indicator name is "Transformer oil temperature", measurement unit is "℃", normal range is "0-80", and measurement time is real-time collection. For example, at a certain moment, the oil temperature of transformer T1 is 50℃. Circuit breaker opening and closing time: indicator number is "I002", indicator name is "circuit breaker opening and closing time", measurement unit is "ms", normal range is "20-60", and measurement time is collected during each opening and closing operation. The opening time of circuit breaker CB1 is 30ms.

[0024] The operation and maintenance personnel entities include: Operation and maintenance personnel A: personnel number "P001", name "Zhang San", position "power equipment operation and maintenance engineer", skill level "senior", work experience "8 years". Operation and maintenance personnel B: personnel number "P002", name "Li Si", position "power equipment operation and maintenance technician", skill level "intermediate", work experience "5 years".

[0025] There is a monitoring edge between transformer T1 and transformer oil temperature, the relationship is expressed as "transformer T1-monitoring-transformer oil temperature", and the monitoring frequency is "15 minutes". There is a monitoring edge between circuit breaker CB1 and circuit breaker opening and closing time, the relationship is expressed as "circuit breaker CB1-monitoring-circuit breaker opening and closing time", and the monitoring frequency is "every operation".

[0026] When the transformer oil temperature exceeds 80°C, the "transformer overheat diagnosis scenario" is triggered, and the relationship is expressed 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 abnormality diagnosis scenario" is triggered, and the relationship is expressed as "circuit breaker opening and closing time abnormality-trigger-circuit breaker operation abnormality diagnosis scenario".

[0027] Maintenance personnel A participated in the daily inspection of transformer T1, and the relationship is expressed as "maintenance personnel A-participated-transformer T1 maintenance work", and the participation time is "every Monday, Wednesday, and Friday". Maintenance personnel B participated in the fault repair work of circuit breaker CB1 on "2023-05-10", and the relationship is expressed as "maintenance personnel B-participated-circuit breaker CB1 fault repair (2023-05-10)".

[0028] 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.

[0029] Scenarios based on fault history include: if transformer T1 has experienced a winding short-circuit fault before, in subsequent operation, a diagnosis is performed on the possibility of the winding short-circuit fault occurring again, forming a "transformer T1 winding short-circuit fault recurrence diagnosis scenario".

[0030] The fault type nodes and attributes are as follows: Electrical fault: The fault number is "F001", the fault name is "Electrical fault", the fault description is "Abnormalities occur in the electrical part of the power equipment, such as short circuits and open circuits", the probability of fault occurrence is "medium", the severity of the fault is "high", and the treatment suggestion is "Immediately cut off the power for maintenance, and check the electrical connection part and insulation condition". Mechanical fault: The fault number is "F002", the fault name is "Mechanical fault", the fault description is "Problems such as damage and jamming occur in the mechanical components of the power equipment", the probability of fault occurrence is "medium", the severity of the fault is "medium", and the treatment suggestion is "Check, repair or replace the mechanical components".

[0031] Perform multi-source data collection and processing, and use the named entity recognition algorithm (trained based on the BERT model) to extract entities from the operation and maintenance logs. For example, from "On June 1, 2023, Operator A performed an inspection on Transformer T1, and the equipment was operating normally", identify entities such as "Operator A" and "Transformer T1", and determine the relationship between them as "Operator A - Participate in - Maintenance work of Transformer T1". Create corresponding nodes and edges in the Neo4j graph database according to the above entities, relationships and attributes.

[0032] As Figure 2 As shown in the adversarial training program flowchart of the method for constructing a health status diagnosis framework of power main equipment based on adversarial learning, step S200: Design the adversarial learning model architecture, use the graph neural network based on Transformer as the large model, and use the lightweight model for a single scenario as the small model; Convert the knowledge graph into graph structure data, perform graph convolution to generate high-order features, and input the high-order features into the large model; Use the knowledge graph as the training sample of the small model, and feedback the output of the small model to the knowledge graph. The small model serves as the atomic module for knowledge graph reasoning; Set the training strategy of the adversarial learning model, and perform adversarial training on the large model and the small model. Specifically, for the large model, convert the knowledge graph into graph structure data. In the graph structure data, nodes represent entities, and edges represent the relationships between entities; Encode the attributes of the nodes and convert them into vector representations; Construct the adjacency matrix of the graph structure to represent the connection relationships between nodes; Use the graph neural network as the graph convolution layer to aggregate and update the node features; Stack several graph convolution layers to capture the high-order structure information of the graph; Introduce the multi-head self-attention mechanism of Transformer into the graph neural network, and fuse the node features output by the graph convolution layer with the output of the Transformer layer to obtain a comprehensive feature representation; For the fault type recognition task, the output layer uses the softmax function for multi-classification; For the fault severity assessment task, the output layer uses linear regression. For small models, the lightweight model uses a multi-layer perceptron and is trained using data corresponding to the scenarios in the knowledge graph.

[0033] Furthermore, map the entities in the knowledge graph to graph nodes, assign unique IDs, and numerically convert the attributes of the entities into initial node feature vectors; convert the diagnostic scenario edges in the knowledge graph into graph edges, define the types and weights of the graph edges, and the types of graph edges include directed edges and undirected edges; use a sparse matrix to represent the node connection relationship, and non-zero elements indicate the existence of an association between nodes. Design a graph convolutional neural network. The input layer is the initial node features, the GAT layer is the first layer, aggregating the features of direct neighbors, the GCN layer is the second layer, capturing second-order neighbor relationships, and the output layer is used to output high-order features and configure parameters; for power main equipment nodes, fuse real-time operating parameters and static attributes and normalize them to [0, 1]; for monitoring index nodes, encode the monitored values and their threshold ranges into two-dimensional vectors; perform preprocessing on the adjacency matrix; in the first layer, each power main equipment node aggregates the features of the monitoring index nodes directly associated with it, and in the second layer, the power main equipment nodes further aggregate the features of the fault type nodes; splice or average the outputs of the multi-head GAT to generate high-order features containing multi-order neighbor information; input the high-order features into the large model.

[0034] Furthermore, extract the corresponding subgraphs from the knowledge graph according to the equipment type and diagnostic scenarios, use the monitoring edges in the knowledge graph as supervision signals for feature engineering; use a multi-layer perceptron to train the small model, perform output standardization, and write the real-time diagnosis results into the knowledge graph. The small model is used as an atomic module for knowledge graph reasoning, and the method is as follows: In the diagnostic rules of the fault type nodes in the knowledge graph, add model relationships pointing to the corresponding small model nodes, and store the call interfaces and input / output definitions; when the knowledge graph inference engine detects the need to call single-scenario diagnosis, extract real-time monitoring data through the model relationship, call the small model API, and write the output of the small model into the knowledge graph as inference evidence for the large model to fuse; package the LSTM model into a Docker container, expose the prediction interface, receive the opening and closing current waveform data, and return the probability of mechanical jamming.

[0035] Furthermore, divide the graph structure data into a training set, a validation set, and a test set, and initialize the large model and the small model. The goal of the large model is to determine whether a device has a fault and the type of the fault. The cross-entropy loss function is adopted, and a regularization term is added. The small model uses the cross-entropy loss function, introduces the adversarial loss, and sets the comprehensive loss function. For both the large model and the small model, select the Adam optimizer. Set the batch size and the number of training epochs for adversarial training, and conduct adversarial training. After each round of training, use the validation set to evaluate the performance of the large model and the small model. According to the evaluation results of the validation set, adjust the hyperparameters in the training strategy. When the training is completed, use the test set to test the large model and the small model, and deploy the trained and tested large model and small model.

[0036] In a specific embodiment, the node and attribute encoding are as follows: Transformer node (T1): Unique ID: T1; Attributes: Rated capacity of 1000 kVA is encoded as 0.8 (the maximum rated capacity is 1250 kVA); Rated voltage of 10 kV is encoded as 0.5 (the maximum rated voltage is 20 kV); Real-time oil temperature of 60 °C is encoded as 0.75 (the normal oil temperature range is 0 - 80 °C). The initial feature vector is [0.8, 0.5, 0.75].

[0037] Circuit breaker node (CB1): Unique ID: CB1; Attributes: Rated current of 630 A is encoded as 0.63 (the maximum rated current is 1000 A); Switching time of 35 ms is encoded as 0.6 (the normal switching time range is 20 - 60 ms). The initial feature vector is [0.63, 0.6].

[0038] Monitoring index node (oil temperature): Unique ID: I1; Attributes: Monitoring value of 60 °C, threshold range [0, 80] °C, encoded as [0.75, 0.8].

[0039] Monitoring index node (switching time): Unique ID: I2; Attributes: Monitoring value of 35 ms, threshold range [20, 60] ms, encoded as [0.6, 0.8].

[0040] Fault type node (overheating fault): Unique ID: F1; Attributes: Fault occurrence probability of 0.3 is encoded as 0.3; Fault severity assessment is 2 (severity range 1 - 5) encoded as 0.4. The initial feature vector is [0.3, 0.4].

[0041] Fault type node (abnormal switching fault): Unique ID: F2; Attributes: Fault occurrence probability of 0.2 is encoded as 0.2; Fault severity assessment is 1.5 encoded as 0.3. The initial feature vector is [0.2, 0.3].

[0042] Transformer T1 and the oil temperature monitoring index I1 have a monitoring edge: the edge type is a directed edge (from T1 to I1), and the weight is 0.8.

[0043] Circuit breaker CB1 and the opening and closing time monitoring index I2 have a monitoring edge: the edge type is a directed edge (from CB1 to I2), and the weight is 0.9.

[0044] The abnormal oil temperature monitoring index I1 (exceeding 80 °C) triggers an overheating fault F1: the edge type is a directed edge (from I1 to F1), and the weight is 0.7.

[0045] The abnormal opening and closing time monitoring index I2 (less than 20 ms or greater than 60 ms) triggers an abnormal opening and closing fault F2: the edge type is a directed edge (from I2 to F2), and the weight is 0.8.

[0046] Generate an adjacency matrix.

[0047] The parameter configuration of the graph convolutional neural network is as follows: Input layer: The initial feature vector of the node. GAT layer (the first layer): The number of heads is 3, and the number of hidden units for each head is 16. GCN layer (the second layer): The number of hidden units is 32. Output layer: The output dimension is 64.

[0048] After the first GAT layer, Transformer T1 aggregates the features of the oil temperature monitoring index I1, and circuit breaker CB1 aggregates the features of the opening and closing time monitoring index I2. After the second GCN layer, Transformer T1 further aggregates the features of the overheating fault F1, and circuit breaker CB1 further aggregates the features of the abnormal opening and closing fault F2. Examples of the high-order features obtained after splicing the outputs of the multi-head GAT are as follows: High-order features of Transformer T1: [0.2, 0.3, 0.1,...] (64-dimensional vector), high-order features of circuit breaker CB1: [0.1, 0.4, 0.2,...] (64-dimensional vector).

[0049] Perform small model training, and extract the subgraph containing Transformer T1, the oil temperature monitoring index I1, and the overheating fault F1 from the knowledge graph. Using the oil temperature monitoring edge as the supervision signal, the input features obtained by feature engineering are [0.75, 0.8] (encoding of the oil temperature monitoring value and the threshold range). Extract the subgraph containing circuit breaker CB1, the opening and closing time monitoring index I2, and the abnormal opening and closing fault F2. The input features are [0.6, 0.8] (encoding of the opening and closing time monitoring value and the threshold range).

[0050] The structure of the multi-layer perceptron is as follows: 2 neurons in the input layer, 16 neurons in the hidden layer, and 2 neurons in the output layer (corresponding to the probabilities of fault occurrence and non-occurrence). Training data: 100 samples of transformer overheating scenarios and 100 samples of abnormal circuit breaker opening and closing scenarios are extracted from the knowledge graph and divided into a training set, a validation set, and a test set according to a ratio of 8:1:1. The training process is as follows: Use the Adam optimizer with a learning rate of 0.001 and 50 training epochs.

[0051] Divide the graph-structured data (including all node and edge information) into a training set, a validation set, and a test set according to a ratio of 7:2:1. Conduct adversarial training between the large model and the small model. The batch size is 16, and the number of training epochs is 100. The large model uses the cross-entropy loss function with a regularization coefficient of 0.01, the small model uses the cross-entropy loss function, the adversarial loss weight is 0.2, and the learning rate of the Adam optimizer is 0.001.

[0052] In each round of training, first train the small model, and then use the output of the small model as the input of the large model to train the large model. Evaluate the performance of the large model and the small model using the validation set every 10 rounds, and adjust hyperparameters such as the learning rate according to the evaluation results. Use the test set to conduct the final test on the large model and the small model, and record metrics such as accuracy and recall. Deploy the trained and tested large model and small model to the actual system for power equipment fault diagnosis.

[0053] Add a relationship pointing to the small model of the transformer overheating scenario to the diagnostic rule of the overheating fault F1, and store the call interface and input / output definitions. When the knowledge graph inference engine detects that an overheating diagnosis of transformer T1 is required, extract the oil temperature monitoring data [0.75, 0.8], call the small model API, the small model returns the probability of overheating fault occurrence as 0.2, and write this result as inference evidence into the knowledge graph for the large model to fuse.

[0054] Step S300: Obtain the online monitoring data. When the large model outputs that a device has a fault, query the historical association rules of this device in the knowledge graph, retrieve the historical diagnostic results of the small model of this device in the corresponding scenario, and generate a diagnostic path graph through Neo4jBloom; Specifically, obtain the online monitoring data, convert the graph neural network model based on Transformer into a lightweight format through ONNX, and deploy it on the edge node to support offline inference; construct the graph data input tensor, including: power main equipment node features and adjacency matrix; the output of the large model is a fault type probability vector. When the confidence of the fault type with the highest probability is greater than the set threshold and this fault type is abnormal, trigger the knowledge graph query process; design a Cypher query, sort the query results according to the update time and expert score of the rules, and give priority to displaying high-priority rules. Screen the diagnostic results of the small model, convert the relationships in the knowledge graph into a structured list; generate a diagnostic path graph through Neo4jBloom.

[0055] In a specific embodiment, real-time online monitoring data of transformer T001 is obtained through edge nodes, including oil temperature (current value 85°C), winding temperature (current value 90°C), load rate (current value 88%), etc. The graph neural network model based on Transformer is converted into the ONNX lightweight format and then deployed on the edge nodes. A graph data input tensor is constructed. The feature of the power main equipment node is [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 rate value (0.88, load rate range 0 - 100%)]. The adjacency matrix contains the connection relationship weights between transformer T001 and other related monitoring index nodes (such as oil temperature sensor node, winding temperature sensor node, etc.) (the connection edge weight between the oil temperature sensor and the transformer is 0.9, and the connection edge weight between the winding temperature sensor and the transformer is 0.8). The large model performs offline inference, and the output fault type probability vector is [overheating fault: 0.8, insulation aging fault: 0.1, normal: 0.1]. Since the confidence level of the overheating fault, 0.8, is greater than the set threshold (0.7) and this fault type is not normal, the knowledge graph query process is triggered.

[0056] Perform a knowledge graph query, and the query results are as follows: Association rules: Rule 1: When the oil temperature exceeds 80°C and the load rate exceeds 80%, an overheating warning is triggered (update time: 2024 - 10 - 01, expert score: 8 points). Rule 2: When the winding temperature exceeds 95°C, there may be an overheating fault (update time: 2023 - 05 - 15, expert score: 6 points).

[0057] Diagnostic results of the small model: Small model diagnostic result 1: The diagnosis was carried out on November 10, 2024, with an overheating fault confidence level of 0.75 (obtained from the small model). Small model diagnostic result 2: The diagnosis was carried out on August 20, 2024, with an overheating fault confidence level of 0.6 (obtained from the small model).

[0058] Screen the diagnostic results of the small model, and only retain the results with a confidence level greater than 0.6. After screening, retain the diagnosis carried out on November 10, 2024 (overheating fault confidence level 0.75). Convert the relationships in the knowledge graph into a structured list.

[0059] Neo4jBloom generates a diagnostic path diagram based on the above-screened and processed information. The diagram includes the transformer T001 node, the overheating fault node, relevant monitoring index nodes (such as oil temperature, winding temperature, load rate, etc.), the diagnostic rule node (when the oil temperature exceeds 80°C and the load rate exceeds 80%, an overheating warning is triggered), and the small model diagnostic result node (diagnosis conducted on November 10, 2024, with an overheating fault confidence level of 0.75). The edges between the nodes represent the relationships between them, 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, the small model diagnostic relationship between transformer T001 and the small model diagnostic result node, etc.

[0060] Step S400: Generate a work order priority based on the diagnostic path diagram, equipment importance, and fault urgency, and arrange maintenance personnel for repair.

[0061] Specifically, the classification of equipment importance from high to low includes special-level equipment, first-level equipment, and ordinary equipment. Obtain the voltage level, load rate, and historical fault times from the power main equipment entity in the knowledge graph, match the level of equipment importance, and quantify it as the equipment importance coefficient; the fault urgency from high to low includes red warning, yellow warning, and blue warning, and quantify it as the fault urgency coefficient; calculate the confidence levels of the large model and the small model, and combine the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the fault urgency coefficient, to calculate the work order priority; Based on the diagnostic path diagram, extract the equipment basic information, fault details, and visualization link as work order information; obtain the skill tags from the maintenance personnel node in the knowledge graph and match them with the fault type; match the territorial maintenance team according to the equipment location, and when there are no idle personnel in the territorial team, escalate to the higher-level operation and maintenance center; send the work order information to the corresponding maintenance personnel.

[0062] Furthermore, receive the large model inference result to obtain the confidence level of the large model ; perform time-weighted averaging on the historical confidence levels of the small model in the same scenario to obtain the confidence level of the small model ; preset the weight of the large model according to business requirements and the weight of the small model ; calculate the work order priority WOP, and the formula is as follows: ; where I is the equipment importance coefficient and E is the fault urgency coefficient.

[0063] In a specific embodiment, obtain the relevant information of transformer T001 from the knowledge graph: Voltage level: 110 kV (corresponding to Class I equipment in the classification rules for equipment importance); Load rate: 88% (relatively high); Number of historical faults: 2 times (in the past year).

[0064] According to the classification rules for equipment importance (extra-class equipment is the main equipment of 500 kV and above; Class I equipment is the main equipment of 110 kV - 220 kV with a relatively high load rate or a relatively large number of historical faults; ordinary equipment is auxiliary equipment below 110 kV), it is determined that transformer T001 is Class I equipment. Quantify the equipment importance coefficient I: The corresponding coefficient I for Class I equipment is I = 1.2.

[0065] It is known that the probability vector of the fault type of transformer T001 output by the large model is [Overheating fault: 0.8, Insulation aging fault: 0.1, Normal: 0.1], and the confidence level of the overheating fault is 0.8. According to the classification rules for fault emergency levels (red warning is for the large model confidence level ≥ 0.9 or the small model gives continuous 3 warnings; yellow warning is for the large model confidence level of 0.8 - 0.9 or high confidence level of the single-scenario small model; blue warning is for the large model confidence level of 0.7 - 0.8 or the first occurrence of an abnormal signal), it is determined that this fault is a yellow warning. Quantify the fault emergency level coefficient E: The corresponding coefficient E for the yellow warning is E = 1.5.

[0066] Receive the inference result of the large model, and obtain that the confidence level of the overheating fault by the large model is 0.8. The historical confidence levels of the small model in the overheating fault scenario are: For the diagnosis conducted on November 10, 2024, the confidence level of the overheating fault is 0.75; for the diagnosis conducted on August 20, 2024, the confidence level of the overheating fault is 0.6.

[0067] Perform time-weighted averaging on the historical confidence levels of the small model (time decay factor: within the most recent 1 month = 1.0, within 2 - 3 months = 0.8, more than 3 months = 0.5), and the result is 0.615.

[0068] According to the business requirements, the preset weight of the large model is 0.6, and the weight of the small model is 0.4. Calculate the work order priority WOP as 1.3068.

[0069] Based on the diagnostic path diagram, extract the equipment basic information of transformer T001 (such as equipment model, installation location, etc.), fault details (the large model diagnoses an overheating fault with a confidence level of 0.8; historical diagnostic results of the small model, etc.), and the visualization link (the link to the diagnostic path diagram generated by Neo4jBloom) as the work order information.

[0070] Obtain skill tags from the operation and maintenance personnel node of the knowledge graph, match them with the overheating fault types, and find the operation and maintenance personnel with the skills to handle transformer overheating faults. For example, operation and maintenance personnel A (skill tags: transformer fault diagnosis, overheating fault handling). Match the affiliated operation and maintenance team according to the equipment location (transformer T001 is located at XX substation), and find that operation and maintenance personnel A in the affiliated team has free time, and send the work order information to operation and maintenance personnel A.

[0071] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for constructing a health status diagnosis framework of power main equipment based on adversarial learning, characterized in that, The steps include: 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 construct a knowledge graph based on the multi-source data of power main equipment; Design an adversarial learning model architecture, use a graph neural network based on Transformer as the large model, and use 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 the training sample of the small model, feedback the output of the small model to the knowledge graph, and use the small model as the atomic module for knowledge graph reasoning; Set the training strategy of the adversarial learning model and perform adversarial training of the large model and the small model; Obtain online monitoring data. When the large model outputs that a device has a fault, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnostic results of the small model of the device in the corresponding scenario, and generate a diagnostic path graph through Neo4jBloom; Generate a work order priority based on the diagnostic path graph, device importance, and fault urgency, and arrange maintenance personnel for repair.

2. The method for constructing a health status diagnosis framework of power main equipment based on adversarial learning according to claim 1, wherein The determination of entities, diagnostic scenario edges, and fault type nodes, and the performance of entity and relationship modeling include: The entities include power main equipment entities, monitoring index entities, and maintenance personnel entities, and define the attributes of the entities; the diagnostic scenario edges include monitoring edges, trigger edges, and participation edges. The monitoring edge represents the relationship between the power main equipment and the monitoring index, the trigger edge represents that the abnormal value of the monitoring index triggers a certain diagnostic scenario, and the participation edge represents that the maintenance personnel participate in a certain diagnostic scenario or the operation and maintenance work of the power main equipment. The diagnostic scenarios include scenarios based on monitoring indexes and scenarios based on fault history; classify faults according to fault nature, determine fault type nodes, and define fault type attributes; Use an ER diagram to represent the conceptual structure among entities, attributes, and relationships, determine the cardinality of the relationships between entities, convert the conceptual structure into a logical model, and select the graph database Neo4j for storage; create corresponding nodes and edges in the graph database according to the logical model.

3. The method for constructing a health status diagnosis framework of power main equipment based on adversarial learning according to claim 1, wherein The collection of multi-source data of power main equipment, the performance of data cleaning and standardization, the extraction of knowledge from unstructured data, and the construction of a knowledge graph based on the multi-source data of power main equipment include: Collect multi-source data of power main equipment, perform missing value processing, outlier processing, duplicate value processing, and noise processing, perform Z-score standardization, uniformly encode text data into the UTF-8 standard encoding format, perform text cleaning, and for categorical text data, use one-hot encoding to convert it into numerical data; use a named entity recognition algorithm to extract knowledge from unstructured data, use a BERT model for training, identify entities in unstructured data, and determine the relationships between entities; Map the processed multi-source data of power main equipment to the knowledge graph.

4. The method for constructing a health status diagnosis framework of a power main equipment based on adversarial learning according to claim 1, wherein The described design of the adversarial learning model architecture uses a Transformer-based graph neural network as the large model and a lightweight model for a single scenario as the small model, including: For the large model, convert the knowledge graph into graph-structured data. In the graph-structured data, nodes represent entities and edges represent the relationships between entities; encode the attributes of the nodes and convert them into vector representations; construct the adjacency matrix of the graph structure to represent the connection relationships between nodes; use a graph neural network as the graph convolutional layer to aggregate and update the node features; stack several graph convolutional layers to capture the high-order structure information of the graph; introduce the multi-head self-attention mechanism of Transformer into the graph neural network, and fuse the node features output by the graph convolutional layer with the output of the Transformer layer to obtain a comprehensive feature representation; for the fault type recognition task, the output layer uses the softmax function for multi-classification; for the fault severity assessment task, the output layer uses linear regression; For the small model, the lightweight model uses a multi-layer perceptron and is trained using the data corresponding to the scenario in the knowledge graph.

5. The method for constructing a health status diagnosis framework of a power main equipment based on adversarial learning according to claim 1, wherein The conversion of the knowledge graph into graph-structured data, performing graph convolution to generate high-order features, and inputting the high-order features into the large model includes: Map the entities in the knowledge graph to graph nodes, assign unique IDs, and numerically convert the attributes of the entities into the initial feature vectors of the nodes; convert the diagnostic scenario edges in the knowledge graph into graph edges, define the types and weights of the graph edges, and the types of graph edges include directed edges and undirected edges; use a sparse matrix to represent the node connection relationships, and non-zero elements indicate the existence of associations between nodes; Design a graph convolutional neural network. The input layer is the initial node features, the GAT layer is the first layer to aggregate the features of direct neighbors, the GCN layer is the second layer to capture the second-order neighbor relationships, and the output layer is used to output high-order features and perform parameter configuration; for the power main equipment nodes, fuse the real-time operating parameters and static attributes and normalize them to [0, 1]; for the monitoring index nodes, encode the monitored values and their threshold ranges into two-dimensional vectors; perform preprocessing on the adjacency matrix; in the first layer, each power main equipment node aggregates the features of the monitoring index nodes directly associated with it, and in the second layer, the power main equipment nodes further aggregate the features of the fault type nodes; splice or average the outputs of the multi-head GAT to generate high-order features containing multi-order neighbor information; input the high-order features into the large model.

6. The method for constructing a health status diagnosis framework of power main equipment based on adversarial learning according to claim 1, wherein Using the knowledge graph as the training sample of the small model, the output of the small model is fed back to the knowledge graph, and the small model is used as the atomic module for knowledge graph reasoning, including: Extract the corresponding subgraphs from the knowledge graph according to the device type and diagnostic scenario, use the monitoring edges in the knowledge graph as the supervision signals, and perform feature engineering; use a multi-layer perceptron to train the small model, perform output standardization, and write the real-time diagnostic results into the knowledge graph; The method for using the small model as the atomic module for knowledge graph reasoning is as follows: In the diagnostic rules of the fault type nodes in the knowledge graph, add model relationships that point to the corresponding small model nodes, and store the call interfaces and input / output definitions. When the knowledge graph inference engine detects the need to call single-scenario diagnosis, it extracts real-time monitoring data through the model relationships, calls the small model API, and writes the output of the small model as inference evidence into the knowledge graph for the large model to fuse. Package the LSTM model into a Docker container, expose the prediction interface, receive the closing and opening current waveform data, and return the probability of mechanical jamming.

7. The method for constructing a health status diagnosis framework of power main equipment based on adversarial learning according to claim 1, wherein The training strategy of the set adversarial learning model for adversarial training of the large model and the small model includes: Divide the graph structure data into training set, validation set, and test set, and initialize the large model and the small model. The goal of the large model is to determine whether the device has a fault and the fault type. It uses the cross-entropy loss function and adds a regularization term. The small model uses the cross-entropy loss function, introduces adversarial loss, and sets a comprehensive loss function. For both the large model and the small model, select the Adam optimizer. Set the batch size and number of training rounds for adversarial training, and conduct adversarial training. After each round of training, evaluate the performance of the large model and the small model using the validation set. According to the evaluation results of the validation set, adjust the hyperparameters in the training strategy. When the training is completed, test the large model and the small model using the test set, and deploy the trained and tested large model and small model.

8. The method for constructing a health status diagnosis framework of a power main equipment based on adversarial learning according to claim 1, wherein When the large model outputs that the device has a fault, query the historical association rules of the device in the knowledge graph, retrieve the historical diagnostic results of the small model of the device in the corresponding scenario, and generate a diagnostic path graph through Neo4jBloom, including: Obtain online monitoring data, convert the graph neural network model based on Transformer into a lightweight format through ONNX, and deploy it on the edge node to support offline inference. Construct the input tensor of the graph data, including: the node features of the power main device and the adjacency matrix. The output of the large model is a fault type probability vector. When the confidence of the fault type with the highest probability is greater than the set threshold and the fault type is abnormal, trigger the knowledge graph query process. Design a Cypher query, sort the query results according to the update time and expert score of the rules, and give priority to displaying high-priority rules. Filter the diagnostic results of the small model, convert the relationships in the knowledge graph into a structured list, and generate a diagnostic path graph through Neo4jBloom.

9. The method for constructing a health status diagnosis framework of power main equipment based on adversarial learning according to claim 1, wherein Generate the work order priority based on the diagnostic path graph, device importance, and fault urgency, and arrange maintenance personnel for repair, including: The classification of equipment importance from high to low includes special-grade equipment, first-grade equipment, and ordinary equipment. The voltage level, load rate, and historical failure times are obtained from the power main equipment entity in the knowledge graph, and the level of equipment importance is matched and quantified as the equipment importance coefficient. The urgency of the failure from high to low includes red warning, yellow warning, and blue warning, which are quantified as the failure urgency coefficient. Calculate the confidence levels of the large model and the small model, and combine the weight distribution of the large model and the small model, as well as the equipment importance coefficient and the failure urgency coefficient, to calculate the work order priority. Based on the diagnostic path diagram, extract the equipment basic information, failure details, and visualization link as the work order information. Obtain the skill tags from the operation and maintenance personnel node in the knowledge graph and match them with the failure type. Match the territorial operation and maintenance team according to the equipment location. When there are no idle personnel in the territorial team, escalate to the higher-level operation and maintenance center. Send the work order information to the corresponding operation and maintenance personnel.

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

Images