Automatic operation and inspection method and system for plant station equipment based on mapping knowledge domain

Abstract

The invention relates to the technical field of automatic operation and maintenance of a power system, in particular to an automatic operation and maintenance method and system for plant-station equipment based on a knowledge graph, and the method comprises the steps: collecting static ledger data, real-time operation data and historical operation and maintenance data of the plant-station equipment, constructing a multi-dimensional association knowledge graph, creating a digital twinborn body for target equipment in a plant station, and carrying out the operation and maintenance of the plant-station equipment. Establishing a bidirectional data channel between the digital twin and a corresponding equipment entity node in the knowledge graph; and based on a state evaluation rule and a fault propagation path model preset in the knowledge graph, performing real-time health state evaluation on the target equipment by using the state evaluation rule and the fault propagation path model to generate a health index, and when an abnormality is monitored, tracing along the topological relation between the electrical connection and the equipment in the knowledge graph, the fault source is positioned, the confidence probability is calculated, risk level evaluation and processing priority ranking are carried out according to the confidence probability, an operation and maintenance diagnosis work order containing an execution time sequence is obtained, and the fault reason can be quickly traced and positioned.

Description

Technical Field

[0001] This invention relates to the field of automated operation and maintenance technology for power systems, and in particular to an automated operation and maintenance method and system for power plant equipment based on knowledge graphs. Background Technology

[0002] As key nodes in the power grid's energy conversion and distribution, power plants (such as substations, converter stations, and switching stations) rely on the reliable operation of their internal equipment (such as transformers, circuit breakers, and disconnectors) to directly impact the safety and stability of the entire power system. Traditional power plant equipment operation and maintenance models primarily rely on periodic inspections, preventative testing, and post-accident repairs. However, with the continuous expansion of the power grid and the increasing complexity of equipment, this model has revealed many limitations that urgently need to be addressed.

[0003] The plant involves multi-source heterogeneous data, including equipment static parameters, real-time operating data, historical maintenance data, and environmental data. This data is scattered across different information systems such as the Production Management System (PMS), Asset Management System, and online monitoring master station. These systems are independent of each other, with inconsistent data formats and standards, forming serious data silos. Maintenance personnel find it difficult to effectively correlate, integrate, and deeply analyze data across systems and time periods, making it impossible to form a comprehensive and three-dimensional understanding of equipment status. Currently, equipment anomaly identification is mostly based on simple threshold alarms, while complex fault diagnosis heavily relies on the personal experience and knowledge of maintenance experts. The entire diagnostic process lacks an intelligent reasoning platform that can integrate domain knowledge, historical cases, and real-time data. During operation and maintenance, it cannot accurately reflect the actual health status of the equipment. Relying solely on traditional threshold alarms is a passive and delayed alarm system, only triggered when equipment parameters significantly exceed limits. It cannot identify early fault symptoms characterized by subtle changes and trend deterioration of multiple parameters, thus lacking the ability to predict and warn of potential faults.

[0004] Therefore, there is an urgent need in this field for a new operation and maintenance method and system that can deeply integrate dynamic and static data, possess intelligent reasoning capabilities, and closely link with the real-time status of equipment, so as to achieve automation, intelligence, and precision in the operation and maintenance of plant equipment. Summary of the Invention

[0005] In view of this, in order to overcome the shortcomings of the prior art, this invention provides a knowledge graph-based automated operation and maintenance method and system for plant equipment. By constructing a panoramic knowledge graph that is synchronously mapped and dynamically evolved with the physical entities of the equipment, and based on this, closed-loop automation of perception, diagnosis, and decision-making is achieved. This breaks down data barriers between systems and constructs a knowledge graph that integrates static attributes, dynamic real-time data, and historical experience. Through the coupling of digital twins and knowledge graphs, real-time simulation and trend prediction of equipment status are achieved. It can identify complex degradation patterns that cannot be reflected by a single parameter, and achieve early and accurate warning of faults. By using knowledge graphs to make judgments on fault diagnosis from point to network, the cause of the fault can be quickly traced and located. The generated operation and maintenance strategy is based on the actual health status of the equipment and risk assessment.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides an automated operation and maintenance method for plant equipment based on knowledge graphs, comprising the following steps: Collect static ledger data, real-time operation data and historical operation and maintenance data of plant equipment, and construct a multi-dimensional relational knowledge graph with equipment entities as the core and integrating equipment topology relationships, real-time monitoring data streams and historical operation and maintenance knowledge; A digital twin is created for the target equipment in the plant. The digital twin establishes a two-way data channel with the corresponding equipment entity node in the knowledge graph. The state update of the digital twin is driven by real-time data stream. The simulation prediction state of the digital twin is synchronously mapped and updated with the attributes of the corresponding entity node in the knowledge graph. Based on the pre-set state assessment rules and fault propagation path model in the knowledge graph, the knowledge graph is traversed in real time. The target device node and associated sensor data nodes and historical defect nodes are used as input. The state assessment rules and fault propagation path model are used to perform real-time health status assessment of the target device to generate a health index. When an anomaly is detected, the source is traced along the electrical connection and equipment topology relationship in the knowledge graph to locate the root cause of the fault and calculate the confidence probability. Based on the confidence probability, risk level assessment and processing priority ranking are performed to obtain operation and maintenance diagnosis work orders containing execution time sequence. After the work order is processed, the maintenance results are fed back to the knowledge graph to optimize the status assessment rules and fault propagation path model.

[0007] As a further aspect of the present invention, the construction of the multidimensional association knowledge graph includes attaching timestamps and device spatial coordinate information to real-time running data, and performing entity alignment based on spatiotemporal constraints and entity attribute similarity, including the following steps: The real-time running data is appended with timestamps and device spatial coordinate information, and a quality score is calculated for each data point, wherein the quality score is dynamically calculated based on signal strength, signal-to-noise ratio and sensor health status; Entity alignment is performed based on spatiotemporal constraints and entity attribute similarity, and dynamic credibility weighted fusion is performed on the collected multi-source monitoring data. The fusion process adopts a weighted algorithm based on data quality and historical credibility. Extract entities and relationships related to equipment names, defects, and maintenance actions from unstructured text, and use a natural language processing model to achieve entity extraction.

[0008] As a further aspect of the present invention, the formula for calculating the quality score of each data point is as follows: ,in, For the first Quality score for each data point For the first Signal strength of each data point For the first Signal-to-noise ratio of each data point For the first The health status of the sensor corresponding to each data point This is a credibility decay factor based on the frequency of historical anomalies. , The attenuation coefficient is... This represents the number of recent anomalies in the sensor. When performing dynamic reliability-weighted fusion of collected multi-source monitoring data, a dynamic weighted fusion based on data quality, historical reliability, and real-time environmental factors is adopted, and the formula is: ,in, In the formula, For fusion value, For the first Sensor readings at each data point For the first Quality score for each data point For dynamic credibility, For the first Historical accuracy for each data point As environmental factor weights, , and To control the weighting balance coefficients of real-time quality, historical accuracy, and environmental factors, , This is a correction item based on real-time environmental data.

[0009] As a further aspect of the present invention, the creation of a digital twin for the target equipment within the plant includes the following steps: Based on the criticality analysis of equipment in the knowledge graph, target equipment is selected and the modeling target of the digital twin is determined. The criticality analysis calculates the criticality score through equipment failure history, power grid topology centrality, and maintenance cost. Equipment with a score higher than the threshold is given priority for modeling. Combining physical mechanisms and data-driven approaches, a multimodal simulation model of the target device is constructed. The physical mechanism model in the multimodal simulation model includes a transformer thermal model, in the following form: In the formula, The top oil temperature is used as the state variable of the model; The ambient temperature is used as an input variable for the model. The rated oil temperature difference is obtained from the equipment ledger in the knowledge graph; The load rate from real-time data is used as an input variable for the model; The thermal time constant, The heat dissipation index, and As model parameters, where It is 1.6-1.8; Using the static parameters and historical operating data of the target device in the knowledge graph, the parameters of the multimodal simulation model are initialized and calibrated. During calibration, the optimal parameter combination is searched to minimize the model prediction error. The calibrated model is encapsulated as a service and assigned a unique identifier, which is associated with the corresponding device entity node in the knowledge graph.

[0010] As a further aspect of the present invention, the bidirectional data channel is used to input the updated device parameters, adjacent device status and topology change information from the knowledge graph into the digital twin to initialize or dynamically adjust the parameters of its physical mechanism model; and to write the device health index, predicted values ​​of key state quantities and remaining life assessment results output by the digital twin based on real-time data simulation into the dynamic attribute fields of the corresponding device nodes in the knowledge graph.

[0011] As a further aspect of the present invention, the digital twin establishes a bidirectional data channel with the corresponding device entity node in the knowledge graph, including the following steps: The communication infrastructure is built based on message middleware, and a unique publish / subscribe topic is defined for each device's data stream; The knowledge graph publishes real-time running data to a designated topic, and the digital twin subscribes to and receives the data to drive model simulation; The digital twin publishes the simulation results to the feedback topic, and the knowledge graph subscribes to and receives the data to update the attributes of the corresponding device nodes.

[0012] As a further aspect of the present invention, when the bidirectional data channel performs data interaction between the digital twin and the knowledge graph, a message queue telemetry transmission is used as the message middleware, and a unique topic is defined for the data stream of each device. The knowledge graph publishes the received real-time data of device load current and ambient temperature to the designated topic through the message middleware. The digital twin subscribes to the topic, receives the data, and performs simulation calculations. The digital twin publishes the calculation results, including predicted oil temperature, health index, and remaining lifespan, to the feedback topic through the message middleware. After the knowledge graph subscribes to the feedback topic, it updates the dynamic attributes of the corresponding device entity node.

[0013] As a further aspect of the present invention, the bidirectional data channel employs a message queue telemetry transmission protocol for data interaction. Within the bidirectional data channel, a hybrid approach combining strong tracking Kalman filtering and deep learning is used to perform online correction of the model parameters of the digital twin. The parameter recursion formula is as follows: in, for The model parameter vector at time step, for The actual observed value at time , Based on The model prediction values ​​of the time parameters, Here is the Kalman gain matrix. For deep learning weights, This is the feature extraction output of a convolutional neural network for recent observation sequences.

[0014] As a further aspect of the present invention, a real-time health status assessment of the target device is performed using state assessment rules and a fault propagation path model to generate a health index. The generation of the health index includes processing a local subgraph centered on the target device using a graph neural network model, comprising the following steps: Using the target device node as the root node, the knowledge graph is traversed to obtain associated real-time data nodes, historical defect nodes, similar device nodes, and environment nodes. Time alignment and feature extraction are then performed to form a comprehensive state vector of the target device. The health index of the target device is calculated by multi-factor weighted fusion. The health index is obtained by processing local subgraphs through a pre-trained graph neural network model. The graph neural network model integrates graph attention mechanism and temporal convolutional network. It highlights the influence of key neighbor nodes through attention weights, outputs a health status embedding vector, and then outputs the health index through a regression layer. The health index is compared with a preset threshold to generate a device health status level, which is then visualized and alerted. When calculating the health index, a local subgraph centered on the target device is extracted from the knowledge graph and input into a pre-trained graph neural network model. Through message passing and node feature aggregation, the health status embedding vector of the target device is learned, and then the health index is output through a regression layer.

[0015] As a further aspect of the present invention, tracing the source of a fault along the electrical connections and device topology relationships in the knowledge graph to locate the root cause and calculate the confidence probability includes the following steps: When an anomaly is detected, an anomaly evidence set is extracted from the knowledge graph, and the fault propagation subgraph is obtained by traversing around the anomaly node. The fault propagation subgraph is mapped to a Bayesian network, where nodes represent device status or fault phenomena, and edges represent fault propagation relationships. The confidence probability of each potential fault root cause is calculated based on Bayes' theorem. The potential fault root causes are sorted according to their confidence probability, and the root cause with the highest confidence probability and its tracing path are highlighted and visualized.

[0016] As a further aspect of the present invention, the fault propagation subgraph is a temporal knowledge graph. The source tracing process includes analyzing the order of occurrence and propagation delay of abnormal evidence in the temporal graph, and matching and verifying it with the expected propagation characteristics of the fault physical model to correct the confidence probability; wherein, the confidence probability of each potential fault root cause is calculated based on Bayes' theorem as follows: In the formula, Root cause of the fault The prior probability, Let be the likelihood probability. Let be the confidence probability, representing the probability of finding the evidence set in the given evidence. Under these conditions, the root cause of the failure The posterior probability of occurrence; For the evidence set The marginal probability.

[0017] Secondly, the present invention also provides a knowledge graph-based automated operation and maintenance system for plant equipment, comprising: The data perception and access layer is configured to collect static ledger data, real-time operation data, and historical operation and maintenance data of plant equipment from the plant's monitoring and data acquisition system, online monitoring devices, inspection robots, production management system, and asset management system. The knowledge center layer, connected to the data perception and access layer, includes a knowledge graph management module and a digital twin engine module. The knowledge graph management module is used to store and manage a multi-dimensional associated knowledge graph centered on equipment entities, integrating equipment topology relationships, real-time monitoring data streams, and historical operation and maintenance knowledge. The digital twin engine module is used to create and run digital twins for target equipment within the plant. The intelligent reasoning and analysis layer interacts with the knowledge center layer and includes a status assessment module and a fault diagnosis module. The status assessment module is configured to perform real-time health status assessment of the target device based on the status assessment rules preset in the knowledge graph to generate a health index. The fault diagnosis module is configured to perform source tracing reasoning using the fault propagation path model in the knowledge graph when an anomaly is detected, locate the root cause of the fault, and calculate the confidence probability. The operation and maintenance decision and service platform, connected to the intelligent reasoning and analysis layer, includes a decision generation module and a service interface module. The decision generation module is configured to assess risk levels and prioritize processing based on the confidence probability, and generate an operation and maintenance diagnostic work order containing the execution sequence. The service interface module is used to push the work order to the operation and maintenance management system and feed back the inspection results to the knowledge graph. The human-computer interaction layer connects to the operation and maintenance decision-making and service platform and the intelligent reasoning and analysis layer, providing operation and maintenance personnel with a graphical operation interface for system configuration, status visualization, early warning information display, diagnostic result query and work order processing.

[0018] As a further aspect of the present invention, in the knowledge center layer, the knowledge graph management module is built on a time-series graph database and is used to store and query the sequence of entities, relationships and their attributes changing over time; the digital twin engine module and the knowledge graph management module interact bidirectionally through a message bus and an application programming interface to achieve state synchronization and model parameter correction between the digital twin and the corresponding device entity nodes in the knowledge graph.

[0019] As a further aspect of the present invention, the data perception and access layer includes a data quality assessment unit, configured to calculate a quality score for the collected real-time data and perform dynamic credibility-weighted fusion of multi-source monitoring data.

[0020] As a further aspect of the present invention, the state evaluation module in the intelligent reasoning and analysis layer integrates a graph neural network model, which is configured to take a local knowledge subgraph centered on the target device as input, learn the device's health status embedding vector through message passing and node feature aggregation, and output a health index.

[0021] Compared with existing technologies, the knowledge graph-based automated operation and maintenance method and system for plant equipment provided by this invention has the following beneficial effects: This invention constructs a multi-dimensional knowledge graph, unifying and associating the static attributes, dynamic states, spatial topology, and historical knowledge of equipment using a graph structure. This breaks down system barriers. Furthermore, by introducing dynamic credibility weighted fusion based on data quality scoring and historical accuracy, it effectively eliminates the influence of abnormal and low-quality data, creating a digital twin of critical equipment that is linked to the knowledge graph in real time. High-fidelity simulations, such as transformer thermal models, are performed using mechanistic models to achieve perception of the internal state of the equipment. Learning from the equipment-centric knowledge subgraph can comprehensively associate multi-dimensional information of the equipment, generating a more comprehensive and accurate health index, avoiding the limitations of a single criterion. Once an anomaly is detected, the system can automatically reason along the electrical connections and topological relationships in the knowledge graph, mapping the fault propagation subgraph to a Bayesian network, highlighting the fault root cause with the highest confidence probability and its complete propagation path, helping maintenance personnel quickly understand the fault mechanism, greatly shortening fault troubleshooting time, and reducing over-reliance on expert experience. This invention effectively solves the core pain points in the operation and maintenance of traditional power plant equipment, such as reliance on manual labor, information fragmentation, delayed early warning, and difficulty in decision-making. It significantly improves the equipment status control capability, fault early warning capability, and intelligent operation and maintenance management level, and has great value for ensuring the safe, stable, and economical operation of the power system.

[0022] These or other aspects of the invention will become more apparent from the following description of embodiments. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based automated operation and maintenance method for plant equipment according to the present invention.

[0024] Figure 2 This is a flowchart illustrating the bidirectional data channel interaction process in a knowledge graph-based automated operation and maintenance method for plant equipment according to the present invention.

[0025] Figure 3 This is a flowchart illustrating the operation and maintenance process of an automated operation and maintenance method for plant equipment based on knowledge graphs, according to the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0027] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0028] See Figures 1 to 3 As shown, embodiments of this application provide a knowledge graph-based automated operation and maintenance method for plant equipment, including the following steps: Step S10: Collect static ledger data, real-time operation data and historical operation and maintenance data of plant equipment, and construct a multi-dimensional relational knowledge graph with equipment entities as the core and integrating equipment topology relationships, real-time monitoring data flow and historical operation and maintenance knowledge.

[0029] In this step, the construction of the multidimensional association knowledge graph includes attaching timestamps and device spatial coordinate information to the real-time running data, and performing entity alignment based on spatiotemporal constraints and entity attribute similarity, including the following steps: The real-time running data is appended with timestamps and device spatial coordinate information, and a quality score is calculated for each data point, wherein the quality score is dynamically calculated based on signal strength, signal-to-noise ratio and sensor health status; Entity alignment is performed based on spatiotemporal constraints and entity attribute similarity, and dynamic credibility weighted fusion is performed on the collected multi-source monitoring data. The fusion process adopts a weighted algorithm based on data quality and historical credibility. Extract entities and relationships related to equipment names, defects, and maintenance actions from unstructured text, and use a natural language processing model to achieve entity extraction.

[0030] In this embodiment, the formula for calculating the quality score for each data point is as follows: ,in, For the first Quality score for each data point For the first Signal strength of each data point For the first Signal-to-noise ratio of each data point For the first The health status of the sensor corresponding to each data point This is a credibility decay factor based on the frequency of historical anomalies. , The attenuation coefficient is... This represents the number of recent anomalies in the sensor.

[0031] For example, in a 110kV smart substation, a data quality assessment was performed on the oil temperature sensor (number T001) of the No. 1 main transformer. This sensor recently exhibited three abnormal readings due to a humid environment. =3. The system automatically calculates the quality score and integrates multi-source data such as load rate and ambient temperature to improve the accuracy of oil temperature monitoring. The data used to calculate the quality score for each data point is as follows: Sensor data: signal strength =0.95, ranging from 0 to 1; Signal-to-noise ratio =20dB; Sensor health status =0.9, based on self-diagnostic report; number of abnormalities =3; Attenuation coefficient =0.1.

[0032] Therefore, the quality score is calculated as follows: ,in, The quality score calculation formula is obtained .

[0033] When performing dynamic reliability-weighted fusion of collected multi-source monitoring data, a dynamic weighted fusion based on data quality, historical reliability, and real-time environmental factors is adopted, and the formula is: ,in, In the formula, For fusion value, For the first Sensor readings at each data point For the first Quality score for each data point For dynamic credibility, For the first Historical accuracy for each data point As environmental factor weights, , and The weighting balance coefficients for controlling real-time quality, historical accuracy, and environmental factors are adjusted. , This is a correction item based on real-time environmental data.

[0034] For example, based on sensor data and quality score calculation results, if the ambient temperature is weighted during the fusion calculation... =0.8, historical accuracy =0.9, weighting coefficient =0.5, =0.3, =0.2, then the dynamic reliability is =0.5×0.85+0.3×0.9+0.2×0.8=0.855. The combined oil temperature value after fusing the above sensor readings is... ,in, =0.5℃, the final oil temperature value was corrected from the original 75℃ to 75.3℃.

[0035] This embodiment evaluates the reliability of each data point by calculating a quality score, performs dynamic reliability-weighted fusion of multi-source monitoring data, and utilizes the quality score... As a reliability attribute storage for data points, the fusion value As a real-time monitoring value attribute of the device status node, the fused value As a dynamic attribute directly of the device entity node, entity alignment and fusion ensure seamless integration of real-time data flow in the knowledge graph with static ledgers and historical knowledge, forming a multi-dimensional relational structure. Spatiotemporal constraints and entity alignment ensure that multi-source data is correctly mapped to entity nodes in the knowledge graph, so that the purified data can be used as input for digital twins and state assessment modules, ensuring the accuracy of simulation and reasoning.

[0036] Step S20: Create a digital twin for the target equipment in the plant. The digital twin establishes a two-way data channel with the corresponding equipment entity node in the knowledge graph. The state update of the digital twin is driven by real-time data stream. The simulation prediction state of the digital twin synchronously maps and updates the attributes of the corresponding entity node in the knowledge graph.

[0037] In this step, creating a digital twin of the target equipment within the plant includes the following steps: Based on the criticality analysis of equipment in the knowledge graph, target equipment is selected and the modeling target of the digital twin is determined. The criticality analysis calculates the criticality score through equipment failure history, power grid topology centrality, and maintenance cost. Equipment with a score higher than the threshold is given priority for modeling. Combining physical mechanisms and data-driven approaches, a multimodal simulation model of the target device is constructed. The physical mechanism model in the multimodal simulation model includes a transformer thermal model, in the following form: In the formula, The top oil temperature is used as the state variable of the model; The ambient temperature is used as an input variable for the model. The rated oil temperature difference is obtained from the equipment ledger in the knowledge graph; The load rate from real-time data is used as an input variable for the model; The thermal time constant, The heat dissipation index, and As model parameters, where It is 1.6-1.8; Using the static parameters and historical operating data of the target device in the knowledge graph, the parameters of the multimodal simulation model are initialized and calibrated. During calibration, the optimal parameter combination is searched to minimize the model prediction error. The calibrated model is encapsulated as a service and assigned a unique identifier, which is associated with the corresponding device entity node in the knowledge graph.

[0038] For example, in a substation, if main transformer No. 1 is selected as the target equipment due to its high historical fault frequency and location at a critical node in the power grid, the system prioritizes creating a digital twin for it based on its criticality analysis score (85 points, threshold 80) to simulate thermal behavior and predict lifespan. The criticality analysis indicators include fault history (weight 0.4, score 70), topological centrality (weight 0.4, score 90), and maintenance cost (weight 0.2, score 80). The overall score is calculated as: 0.4 × 70 + 0.4 × 90 + 0.2 × 80 = 84. Therefore, in the model parameters, the transformer thermal model... Initialization parameters , Rated oil temperature difference ℃.

[0039] In this embodiment, the bidirectional data channel is used to input the updated device parameters, adjacent device status and topology change information from the knowledge graph into the digital twin to initialize or dynamically adjust the parameters of its physical mechanism model; and to write the device health index, predicted values ​​of key state quantities and remaining life assessment results output by the digital twin based on real-time data simulation into the dynamic attribute fields of the corresponding device nodes in the knowledge graph.

[0040] The establishment of a bidirectional data channel between the digital twin and the corresponding device entity node in the knowledge graph includes the following steps: The communication infrastructure is built based on message middleware, and a unique publish / subscribe topic is defined for each device's data stream; The knowledge graph publishes real-time running data to a designated topic, and the digital twin subscribes to and receives the data to drive model simulation; The digital twin publishes the simulation results to the feedback topic, and the knowledge graph subscribes to and receives the data to update the attributes of the corresponding device nodes.

[0041] Taking main transformer No. 1 as an example, the digital twin of main transformer No. 1 interacts with the knowledge graph in real time via a 5G network: the twin subscribes to oil temperature data topics, the knowledge graph publishes load change information, and data within the channel is transmitted after compression and privacy protection. See also Figure 2As shown, when establishing a bidirectional data channel, the MQTT protocol is used. Publish / subscribe rules are defined for the topic ("device / transformer1 / temperature"). The 5G edge gateway handles data routing, reducing reliance on the cloud. Sensitive data is annotated using differential privacy, with Laplace noise added to the differential privacy layer at a scale parameter of 0.1 to ensure that oil temperature data is obfuscated during transmission; for example, an oil temperature of 75℃ is obfuscated to 75℃±0.05℃. Network bandwidth is then monitored, using 5G parameters with latency <10ms and bandwidth of 100Mbps. The edge node is deployed locally at the substation. During data synchronization, the knowledge graph publishes real-time parameters, which the twin receives and drives model simulation. The results are then written back into the dynamic attribute fields of the knowledge graph.

[0042] In this embodiment, when the bidirectional data channel performs data interaction between the digital twin and the knowledge graph, a message queue telemetry transmission is used as the message middleware, and a unique topic is defined for the data stream of each device. The knowledge graph publishes the received real-time data of device load current and ambient temperature to the designated topic through the message middleware. The digital twin subscribes to the topic, receives the data, and performs simulation calculations. The digital twin publishes the calculation results, including predicted oil temperature, health index, and remaining lifespan, to the feedback topic through the message middleware. After the knowledge graph subscribes to the feedback topic, it updates the dynamic attributes of the corresponding device entity node.

[0043] The bidirectional data channel uses a message queue telemetry transmission protocol to achieve data interaction. Within this bidirectional data channel, a hybrid approach combining strong tracking Kalman filtering and deep learning is employed to perform online correction of the digital twin's model parameters. The parameter recursion formula is as follows: in, for The model parameter vector at time step, for The actual observed value at time , Based on The model prediction values ​​of the time parameters, Here is the Kalman gain matrix. For deep learning weights, This is the feature extraction output of a convolutional neural network for recent observation sequences.

[0044] For example, if the thermal time constant of the digital twin of main transformer No. 1 When the thermal model parameters need online calibration, Kalman filtering can be used to process steady-state data, CNN can be used to capture temporal anomalies, and the backup model can be switched when the prediction error continuously exceeds the limit. Let the input data include the actual oil temperature observation sequence. =[75.0,75.2,75.5]℃, model predicted value =[74.8,75.1,75.3]℃; When performing hybrid correction, let the Kalman gain matrix be... CNN extracts features from the three most recent time points, and the weights are... =0.2, corrected parameter During anomaly detection, if the error exceeds 2°C for 5 consecutive times, the system will automatically switch to the backup LSTM model.

[0045] Step S30: Based on the pre-set state assessment rules and fault propagation path model in the knowledge graph, traverse the knowledge graph in real time, take the target device node and associated sensor data nodes and historical defect nodes as input, use the state assessment rules and fault propagation path model to perform real-time health status assessment of the target device to generate a health index, and when an anomaly is detected, trace the source along the electrical connection and device topology relationship in the knowledge graph to locate the root cause of the fault and calculate the confidence probability.

[0046] In this step, a real-time health status assessment of the target device is performed using state assessment rules and a fault propagation path model to generate a health index. The generation of the health index includes processing a local subgraph centered on the target device using a graph neural network model, comprising the following steps: Using the target device node as the root node, the knowledge graph is traversed to obtain associated real-time data nodes, historical defect nodes, similar device nodes, and environment nodes. Time alignment and feature extraction are then performed to form a comprehensive state vector of the target device. The health index of the target device is calculated by multi-factor weighted fusion. The health index is obtained by processing local subgraphs through a pre-trained graph neural network model. The graph neural network model integrates graph attention mechanism and temporal convolutional network. It highlights the influence of key neighbor nodes through attention weights, outputs a health status embedding vector, and then outputs the health index through a regression layer. The health index is compared with a preset threshold to generate a device health status level, which is then visualized and alerted. When calculating the health index, a local subgraph centered on the target device is extracted from the knowledge graph and input into a pre-trained graph neural network model. Through message passing and node feature aggregation, the health status embedding vector of the target device is learned, and then the health index is output through a regression layer.

[0047] In this embodiment, the graph attention mechanism is integrated into the graph neural network model, located in the graph attention layer. As the core module for processing local subgraphs of the knowledge graph, its working mechanism is as follows: for each device node in the knowledge graph, the graph attention layer takes the node feature vector as input, such as real-time device data and historical defect features. During processing, it calculates the attention coefficient between the node and each neighboring node and outputs the weighted aggregated node features. The temporal convolutional network is integrated into the graph neural network model, located after the graph attention layer, and is used to process the temporal dimension of node features. It takes the node feature sequence output by the graph attention layer as input, and through the causal convolutional layer in the temporal convolutional network, ensures that the output of each time step depends only on the current and historical time steps. The residual block in the temporal convolutional network contains multiple residual units, each composed of dilated convolution, weight normalization, and an activation function, used to capture long-term dependencies. The node embedding vector after extracting the temporal features is used as the output to capture the trend degradation pattern of the device state. When operating the graph neural network model integrating graph attention mechanism and temporal convolutional network, taking the No. 1 main transformer in the document as an example, the input is: a local subgraph including the No. 1 main transformer node, the associated oil temperature sensor node (T001), and the historical defect node (buffer oil leakage in 2024). The node features include oil temperature of 75°C, load rate of 82%, and defect markers; graph attention layer: calculates the attention weights between the main transformer node and the sensor node, and generates a spatial feature vector after aggregation; temporal convolutional network: takes the spatial feature sequence of the past 10 time points as input, extracts the temporal pattern through causal convolution, and outputs a temporal embedding vector; regression layer: inputs the temporal embedding vector into the fully connected layer, and outputs a health index HI=0.83.

[0048] In this embodiment, the source of the fault is traced along the electrical connections and device topology relationships in the knowledge graph to locate the root cause and calculate the confidence probability, including the following steps: When an anomaly is detected, an anomaly evidence set is extracted from the knowledge graph, and the fault propagation subgraph is obtained by traversing around the anomaly node. The fault propagation subgraph is mapped to a Bayesian network, where nodes represent device status or fault phenomena, and edges represent fault propagation relationships. The confidence probability of each potential fault root cause is calculated based on Bayes' theorem. The potential fault root causes are sorted according to their confidence probability, and the root cause with the highest confidence probability and its tracing path are highlighted and visualized.

[0049] In this embodiment, the fault propagation subgraph is a time-series knowledge graph. The source tracing process includes analyzing the order of occurrence and propagation delay of abnormal evidence in the time-series graph, and matching and verifying it with the expected propagation characteristics of the fault physical model to correct the confidence probability. The confidence probability of each potential fault root cause is calculated based on Bayes' theorem as follows: In the formula, Root cause of the fault The prior probability, Let be the likelihood probability. Let be the confidence probability, representing the probability of finding the evidence set in the given evidence. Under these conditions, the root cause of the failure The posterior probability of occurrence; For the evidence set The marginal probability.

[0050] Step S40: Based on the confidence probability, risk level assessment and processing priority ranking are performed to obtain operation and maintenance diagnosis work orders containing execution time sequence. After the work order is processed, the maintenance results are fed back to the knowledge graph to optimize the status assessment rules and fault propagation path model.

[0051] For example, when the knowledge graph-based automated operation and maintenance method for substation equipment in this embodiment is applied to the No. 1 main transformer of a 110kV smart substation as the target equipment, in the knowledge graph construction stage, the static data of the No. 1 main transformer is first obtained from the PMS system, and real-time data of oil temperature, load rate and dissolved gas in oil are collected through the Internet of Things platform. Historical data of maintenance records in the past three years are extracted. For example, from the static data of the No. 1 main transformer with model S11-50000 / 110, capacity of 50MVA and commissioning date of 2018-09 in the PMS system, the real-time data collected shows that the oil temperature is 75℃, the load rate is 82%, and the dissolved gases in oil are: H2, 158μL / L and C2H2, 3.2μL / L. In the extracted historical data, the submersible oil pump was replaced in 2022 and bushing oil leakage was treated in 2023.

[0052] Knowledge extraction was performed on the collected data. The BERT-BiLSTM model was used to extract entity relationships from the inspection report of abnormal oil temperature rise in main transformer #1, labeled as: #1 main transformer, hasSymptom, abnormal oil temperature rise; then, a topology containing key nodes was constructed in the graph database Neo4j. (#1 Main Transformer) - [Connection] -> (110kV Busbar); (#1 Main Transformer) - [Installation Location] -> (Transformer Room No. 1); (#1 Main Transformer) - [Monitored] -> (Oil Temperature Sensor T001); (Oil temperature sensor T001) - [Real-time value] -> (75℃).

[0053] Based on the collected data and topology, a transformer thermal model was established. Using historical data from the last 30 days, the least squares method was used to correct the model and obtain the optimized parameter τ=192min. The model prediction error was reduced from ±3℃ to ±1.2℃. During bidirectional data channel operation, the knowledge graph published real-time oil temperature data through an MQTT topic. The digital twin subscribed to this topic, driving the model to output a predicted oil temperature of 76.3℃. The digital twin fed back the health index HI=0.83 to the knowledge graph through the topic, updating the health index attribute of the No. 1 main transformer node.

[0054] Then, the data from 12 sensors associated with the No. 1 main transformer, 3 historical defect records, and the operating data of the No. 2 main transformer at the same station are extracted to form a feature vector: V=[75,0.82,158,3.2,0,0.85,0.91] (temperature, load, H2, C2H2, defect flag, self-comparison, lateral comparison); then the equipment topology subgraph is input into a 3-layer graph convolutional network, which outputs a health status embedding vector, and finally obtains HI=0.83 through a fully connected layer.

[0055] If the system detects that the HI value of main transformer No. 1 drops to 0.63 and the C2H2 content increases to 5.8 μL / L, an abnormal alarm is triggered. During risk rating, based on a confidence probability of 0.76 and the equipment's criticality level (Class A), a risk level of "High Risk" is generated. After implementing this embodiment, the fault warning time for main transformer No. 1 is reduced from an average of 72 hours to 240 hours, the fault location accuracy rate increases from 65% to 92%, and maintenance decision-making efficiency is improved by approximately 3 times. Through continuous learning of the knowledge graph, the system's diagnostic confidence probability for similar faults increases by approximately 15% after 6 months of operation.

[0056] In some embodiments, the present invention also provides a knowledge graph-based automated operation and maintenance system for plant equipment, comprising: The data perception and access layer is configured to collect static ledger data, real-time operation data, and historical operation and maintenance data of plant equipment from the plant's monitoring and data acquisition system, online monitoring devices, inspection robots, production management system, and asset management system. The knowledge center layer, connected to the data perception and access layer, includes a knowledge graph management module and a digital twin engine module. The knowledge graph management module is used to store and manage a multi-dimensional associated knowledge graph centered on equipment entities, integrating equipment topology relationships, real-time monitoring data streams, and historical operation and maintenance knowledge. The digital twin engine module is used to create and run digital twins for target equipment within the plant. The intelligent reasoning and analysis layer interacts with the knowledge center layer and includes a status assessment module and a fault diagnosis module. The status assessment module is configured to perform real-time health status assessment of the target device based on the status assessment rules preset in the knowledge graph to generate a health index. The fault diagnosis module is configured to perform source tracing reasoning using the fault propagation path model in the knowledge graph when an anomaly is detected, locate the root cause of the fault, and calculate the confidence probability. The operation and maintenance decision and service platform, connected to the intelligent reasoning and analysis layer, includes a decision generation module and a service interface module. The decision generation module is configured to assess risk levels and prioritize processing based on the confidence probability, and generate an operation and maintenance diagnostic work order containing the execution sequence. The service interface module is used to push the work order to the operation and maintenance management system and feed back the inspection results to the knowledge graph. The human-computer interaction layer connects to the operation and maintenance decision-making and service platform and the intelligent reasoning and analysis layer, providing operation and maintenance personnel with a graphical operation interface for system configuration, status visualization, early warning information display, diagnostic result query and work order processing.

[0057] In this embodiment, in the knowledge center layer, the knowledge graph management module is built on a time-series graph database and is used to store and query the sequence of entities, relationships and their attributes changing over time. The digital twin engine module and the knowledge graph management module interact bidirectionally through a message bus and an application programming interface to achieve state synchronization and model parameter correction between the digital twin and the corresponding device entity nodes in the knowledge graph.

[0058] In this embodiment, the data perception and access layer includes a data quality assessment unit, configured to calculate a quality score for the collected real-time data and perform dynamic credibility-weighted fusion of multi-source monitoring data.

[0059] The state assessment module in the intelligent reasoning and analysis layer integrates a graph neural network model. It is configured to take a local knowledge subgraph centered on the target device as input, learn the device's health status embedding vector through message passing and node feature aggregation, and output a health index.

[0060] This invention constructs a multi-dimensional knowledge graph, unifying and associating the static attributes, dynamic states, spatial topology, and historical knowledge of equipment using a graph structure. This breaks down system barriers. Furthermore, by introducing dynamic credibility weighted fusion based on data quality scoring and historical accuracy, it effectively eliminates the influence of abnormal and low-quality data, creating a digital twin of critical equipment that is linked to the knowledge graph in real time. High-fidelity simulations, such as transformer thermal models, are performed using mechanistic models to achieve perception of the internal state of the equipment. Learning from the equipment-centric knowledge subgraph can comprehensively associate multi-dimensional information of the equipment, generating a more comprehensive and accurate health index, avoiding the limitations of a single criterion. Once an anomaly is detected, the system can automatically reason along the electrical connections and topological relationships in the knowledge graph, mapping the fault propagation subgraph to a Bayesian network, highlighting the fault root cause with the highest confidence probability and its complete propagation path, helping maintenance personnel quickly understand the fault mechanism, greatly shortening fault troubleshooting time, and reducing over-reliance on expert experience. This invention effectively solves the core pain points in the operation and maintenance of traditional power plant equipment, such as reliance on manual labor, information fragmentation, delayed early warning, and difficulty in decision-making. It significantly improves the equipment status control capability, fault early warning capability, and intelligent operation and maintenance management level, and has great value for ensuring the safe, stable, and economical operation of the power system.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge graph-based substation equipment automation operation and inspection method, characterized in that, The method comprises the following steps: Collecting static account data, real-time operation data and historical operation and maintenance data of the plant station equipment, constructing a multi-dimensional associated knowledge graph taking equipment entities as the core, and integrating equipment topology relationship, real-time monitoring data flow and historical operation and maintenance knowledge; Creating a digital twin for the target equipment in the plant station, establishing a bidirectional data channel between the digital twin and the corresponding equipment entity node in the knowledge graph, driving the digital twin state update through real-time data flow, and synchronously mapping and updating the attributes of the corresponding entity node in the knowledge graph according to the simulation and prediction state of the digital twin; Based on the state evaluation rules and fault propagation path model preloaded in the knowledge graph, the knowledge graph is traversed in real time, the target equipment node and the associated sensor data node and historical defect node are taken as input, the state evaluation rules and fault propagation path model are used to evaluate the real-time health state of the target equipment to generate a health index, and when an anomaly is monitored, the fault root cause is located and the confidence probability is calculated by tracing along the electrical connection and equipment topology relationship in the knowledge graph; According to the confidence probability, risk level evaluation and processing priority sorting are performed to obtain an operation and maintenance diagnosis work order containing an execution time sequence, and the repair result is fed back to the knowledge graph after the execution of the work order, so that the state evaluation rules and fault propagation path model are optimized.

2. The knowledge graph-based substation equipment automation inspection method of claim 1, wherein, The construction of the multi-dimensional associated knowledge graph comprises the following steps: Timestamps and device spatial coordinate information are added to the real-time operation data, and the quality score of each data point is calculated, wherein the quality score is dynamically calculated based on signal strength, signal-to-noise ratio and sensor health state; Entity alignment is performed based on space-time constraints and entity attribute similarity, and the collected multi-source monitoring data is dynamically and credibly weighted and fused, wherein a weighting algorithm based on data quality and historical credibility is used in the fusion process; The entities and relationships of device name, defect phenomenon and repair action are extracted from unstructured text, and entity extraction is realized by using a natural language processing model.

3. The knowledge graph-based substation equipment automation inspection method of claim 2, wherein, When calculating the quality score of each data point, the calculation formula is: wherein, is a quality score of the th data point, is a signal strength of the th data point, is a signal-to-noise ratio of the th data point, is a sensor self-health status corresponding to the th data point, is a credibility decay factor based on historical abnormal frequency, , is a decay coefficient, is a recent abnormal number of the sensor; When the collected multi-source monitoring data is dynamically and credibly weighted and fused, dynamic weighted fusion based on data quality, historical credibility and real-time environmental factors is adopted, and the formula is: wherein, wherein, is a fusion value, is a sensor reading for the nth data point, is a quality score for the nth data point, is a dynamic trustworthiness, is a historical accuracy rate for the nth data point, is an environmental factor weight, , , and are weight balancing coefficients that control the real-time quality, historical accuracy rate, and environmental factor weight, , is a correction term based on real-time environmental data.​​ 4. The knowledge graph-based substation equipment automation inspection method of claim 1, wherein, The creation of the digital twin for the target equipment in the plant station comprises the following steps: Based on the keyness analysis of the equipment in the knowledge graph, the target equipment is selected, and the modeling target of the digital twin is determined, wherein the keyness analysis calculates the keyness score through the equipment fault history, power grid topology center degree and repair cost, and the equipment with a score higher than a threshold value is preferentially modeled; In combination with physical mechanism and data driving, a multi-modal simulation model of the target device is constructed, wherein a physical mechanism model in the multi-modal simulation model comprises a transformer thermal model, which is in a form of: ; wherein, is a top layer oil temperature, as a state variable of the model; is an ambient temperature, as an input variable of the model; is a rated oil temperature difference obtained from a device account in a knowledge graph; is a load rate from real-time data, as an input variable of the model; is a thermal time constant, is a heat dissipation index, and are model parameters, wherein, is 1.6-1.

8. The parameters of the multi-modal simulation model are initialized and calibrated by using the static parameters and historical operation data of the target equipment in the knowledge graph, wherein the optimal parameter combination is searched to minimize the model prediction error during calibration; The calibrated model is service-encapsulated, and a unique identifier is allocated to it, which is associated with the corresponding equipment entity node in the knowledge graph.

5. The knowledge graph-based substation equipment automation inspection method of claim 1, wherein, The bidirectional data channel is used for inputting updated device parameters, adjacent device states and topological change information in the knowledge graph into the digital twin to initialize or dynamically adjust parameters of a physical mechanism model of the digital twin; and the device health index, the key state quantity prediction value and the residual life assessment result simulated and output by the digital twin based on real-time data are written reversely into a dynamic attribute field of a corresponding device node in the knowledge graph.

6. The knowledge graph-based substation equipment automation inspection method of claim 5, wherein, The bidirectional data channel adopts a message queue telemetry transport protocol to realize data interaction, and in the bidirectional data channel, a hybrid of strong tracking Kalman filtering and deep learning is adopted to correct model parameters of the digital twin online, and a parameter recursive formula is as follows: in, for The model parameter vector at time step, for The actual observed value at time , For based on The model prediction values ​​of the time parameters, Here is the Kalman gain matrix. For deep learning weights, This is the feature extraction output of a convolutional neural network for recent observation sequences.

7. The knowledge graph-based substation equipment automation operation inspection method of claim 1, wherein, The target device is subjected to real-time health state assessment by using a state assessment rule and a fault propagation path model to generate a health index, wherein the generation of the health index comprises processing a local subgraph centered on the target device by using a graph neural network model, including the following steps: Taking the target device node as a root node, the knowledge graph is traversed to obtain associated real-time data nodes, historical defect nodes, same-type device nodes and environment nodes, and time alignment and feature extraction are performed to form a comprehensive state vector of the target device; A multi-factor weighted fusion is adopted to calculate the health index of the target device, wherein the health index is obtained by processing a local subgraph by using a pre-trained graph neural network model, the graph neural network model integrates a graph attention mechanism and a time sequence convolution network, and the influence of key neighbor nodes is highlighted by an attention weight, and a health state embedding vector is output, and then a health index is output by a regression layer; The health index is compared with a preset threshold to generate a device health state grade, and visual display and early warning are performed. Wherein, when calculating the health index, a local subgraph centered on the target device extracted from the knowledge graph is input into a pre-trained graph neural network model, a health state embedding vector of the target device is learned by message passing and node feature aggregation, and then a health index is output by a regression layer.

8. The knowledge graph-based substation equipment automation operation inspection method of claim 1, wherein, The fault root is located and a confidence probability is calculated by tracing along an electrical connection and a device topological relationship in the knowledge graph, including the following steps: When an anomaly is monitored, an abnormal evidence set is extracted from the knowledge graph, and a fault propagation subgraph is obtained by traversing around an abnormal node; The fault propagation subgraph is mapped into a Bayesian network, wherein a node represents a device state or a fault phenomenon, and an edge represents a fault propagation relationship; Based on a Bayesian formula, a confidence probability of each potential fault root is calculated, the potential fault roots are sorted according to the confidence probability, and the fault root with the highest confidence probability and its tracing path are highlighted and visualized.

9. The knowledge graph-based substation equipment automation operation inspection method of claim 8, wherein, The fault propagation subgraph is a time sequence knowledge graph, and the tracing process includes analyzing an appearance order and a propagation delay of the abnormal evidence in the time sequence graph, and matching and verifying with expected propagation characteristics of a fault physical model to correct the confidence probability; wherein, the confidence probability of each potential fault root is calculated based on a Bayesian formula as follows: In the formula, Root cause of the fault The prior probability, Let be the likelihood probability. Let be the confidence probability, representing the probability of finding the evidence set in the given evidence. Under these conditions, the root cause of the failure The posterior probability of occurrence; For the evidence set The marginal probability.

10. A knowledge graph-based substation equipment automation operation and inspection system, characterized in that, The system is used for executing steps of the knowledge graph-based power station device automatic operation and inspection method according to any one of claims 1-9, and comprises: The data perception and access layer is configured to collect static account data, real-time operation data and historical operation and maintenance data of power station equipment from a supervisory control and data acquisition system, an online monitoring device, a patrol robot, a production management system and an asset management system of the power station; The knowledge center layer is connected with the data perception and access layer and includes a knowledge graph management module and a digital twin engine module; the knowledge graph management module is configured to store and manage a multi-dimensional associated knowledge graph with equipment entities as cores, integrated with equipment topology relationships, real-time monitoring data streams and historical operation and maintenance knowledge; and the digital twin engine module is configured to create and run a digital twin for a target equipment in the power station; The intelligent reasoning and analysis layer interacts with the knowledge center layer and includes a state evaluation module and a fault diagnosis module; the state evaluation module is configured to perform real-time health state evaluation on the target equipment based on pre-stored state evaluation rules in the knowledge graph to generate a health index; and the fault diagnosis module is configured to perform traceability reasoning, locate a fault root cause and calculate a confidence probability by using a fault propagation path model in the knowledge graph when an abnormality is detected; The operation and maintenance decision and service platform is connected with the intelligent reasoning and analysis layer and includes a decision generation module and a service interface module; the decision generation module is configured to perform risk level evaluation and processing priority sorting according to the confidence probability to generate an operation and maintenance diagnosis work order containing an execution time sequence; and the service interface module is configured to push the work order to an operation and maintenance management system and feed back a repair result to the knowledge graph; The human-computer interaction layer is connected with the operation and maintenance decision and service platform and the intelligent reasoning and analysis layer and provides a graphical operation interface for operation and maintenance personnel to configure a system, visualize a state, display early warning information, query a diagnosis result and process a work order.

Images