Health management method of thermal power generation equipment based on knowledge graph and mechanism model

Abstract

The present invention provides a thermal power equipment health management method based on knowledge graph and mechanism model, which collects equipment operation data in real time or regularly; processes and analyzes data to identify the health status and potential problems of equipment; uses models such as local recursive global feedforward neural network to train models based on historical data to predict future faults; arranges regular inspections and necessary component replacements according to the prediction results, and when an abnormality is detected, uses wavelet lifting algorithm to analyze residual signals, extracts fault features, and compares with historical data; identifies and locates equipment faults by analyzing monitoring data and combining knowledge graph tools, and accurately determines the type and severity of faults; develops decision support systems in combination with existing mature management methods and operating procedures; builds a fault operation and maintenance knowledge graph framework to realize visual management of knowledge and a question-answering system based on knowledge graphs. Reduce manual intervention and build a humanized power plant that is energy-efficient, safe and reliable.

Description

Technical Field

[0001] The present invention relates to the field of health management of thermal power generation equipment, and in particular to a health management method of thermal power generation equipment based on a knowledge graph and a mechanism model. Background Art

[0002] Smart thermal power plants are based on the development of digital and information technology, integrating advanced sensor measurement, information communication, automatic control, artificial intelligence, cloud computing, big data processing, 3D visualization and other information technology means. These technologies are combined with industrialized technologies in the traditional power generation process, focusing on building an integrated data flow system and a comprehensive linkage application platform for multiple systems, aiming to create the digital management center of thermal power plants - the "digital brain".

[0003] Through this platform, smart thermal power plants can promote seamless information exchange, pre-judge and handle potential problems, enable data to evolve on its own, support automation and linkage of control, and further integrate with the physical infrastructure of power plants. Its goal is to improve equipment reliability, optimize operating performance and power generation efficiency, reduce energy consumption and environmental pollution, enhance adaptability to the power market, and reduce overall operating costs. Ultimately, smart thermal power plants will demonstrate advanced features in terms of intelligence, safety, economic benefits, and environmental protection, forming a new type of modern thermal power generation model.

[0004] In contrast, traditional thermal power equipment health management mainly relies on manual monitoring and regular maintenance, which has many disadvantages. On the one hand, it is inefficient and difficult to meet the high requirements of modern power production; on the other hand, it is impossible to achieve real-time fault prediction and prevention, which increases the risk of equipment failure and may lead to production interruptions and economic losses. The application of smart thermal power plant technology has greatly made up for these shortcomings and improved the overall operation level of the power plant. Summary of the invention

[0005] The main purpose of the present invention is to provide a thermal power equipment health management method based on knowledge graph and mechanism model, so as to solve the problem that the traditional thermal power equipment health management method mainly relies on manual real-time monitoring and regular maintenance plans. This method is not only inefficient, but also has obvious deficiencies in fault prediction and prevention due to the lack of real-time monitoring means.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a health management method for thermal power generation equipment based on knowledge graph and mechanism model, the method comprising:

[0007] S1. Collect equipment operation data in real time or regularly;

[0008] S2, process and analyze data to identify equipment health status and potential problems;

[0009] S3, using models such as local recursive global feedforward neural networks to train models based on historical data to predict future failures;

[0010] S4. Arrange regular inspections and necessary parts replacements based on forecast results to avoid unplanned downtime;

[0011] S5. When an abnormality is detected, the residual signal is analyzed using the wavelet lifting algorithm, the fault features are extracted, and compared with the historical data;

[0012] By analyzing monitoring data and combining knowledge graph tools, equipment faults can be identified and located, and the type and severity of the faults can be accurately determined;

[0013] S6. Develop a decision support system based on existing mature management methods and operating procedures;

[0014] S7. Build a knowledge graph framework for fault operation and maintenance to realize visual management of knowledge and a question-answering system based on the knowledge graph.

[0015] In the preferred embodiment, the specific method of step S1 is:

[0016] Deploy temperature sensors, vibration sensors, pressure sensors, and current sensors on each device and connect them to the central data acquisition system through IoT technology to collect device operation data in real time or regularly. The collected data will be directly used for the next step of data analysis and processing;

[0017] Use MQTT or CoAP lightweight protocols for IoT environments;

[0018] Use Apache Kafka or Amazon Kinesis to process large amounts of sensor data.

[0019] In the preferred embodiment, the specific method of step S2 is:

[0020] A1. Organize and store the data collected in step S1 with the help of big data technology, and use data analysis and artificial intelligence to deeply process and analyze the data;

[0021] Among them, the collected data is first cleaned and standardized to remove outliers and noise;

[0022] The box plot method was used to detect outliers, and data points that exceeded the range of 1.5 times the interquartile range of the upper and lower quartiles were treated as outliers;

[0023] Then the data is standardized so that data with different characteristics have the same scale. The Z-score standardization method is used to make data with different dimensions comparable, which is convenient for subsequent data analysis and algorithm processing.

[0024] A2. Through multi-source information fusion, vector database, SQL database and knowledge graph tools are used to integrate data and knowledge.

[0025] In the preferred embodiment, the specific method of step A2 is:

[0026] A21. Multi-source information fusion: Extract vector feature data from the vector database, obtain structured data from the SQL database, extract semantic relationship data from the knowledge graph, and use a weighted fusion method to fuse multi-source data. Suppose the fused data is: ,in:

[0027] ;

[0028] is the weight coefficient, satisfying ;

[0029] in, , , Represents the number of data feature dimensions extracted from the vector database, SQL database, and knowledge graph respectively;

[0030] By weighted fusion of data from different sources, the advantages of each data source can be fully utilized to provide more comprehensive information for subsequent analysis;

[0031] A22. Use the principal component analysis method to reduce the dimension and extract features of the fused data. Suppose the fused data matrix is Calculate its covariance matrix ;

[0032] Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue and the corresponding eigenvector , select the first The eigenvectors form the projection matrix , project the original data into the low-dimensional space to obtain the reduced-dimensional data ;

[0033] in, The covariance matrix is ​​used to measure the correlation of data, and the eigenvectors obtained by eigenvalue decomposition are used to construct a projection matrix to project high-dimensional data into a low-dimensional space, achieve dimensionality reduction and feature extraction, reduce data complexity, and improve the efficiency of subsequent analysis;

[0034] A23. Introduce the autoencoder in deep learning to further extract features and detect anomalies in the data after dimensionality reduction;

[0035] The autoencoder consists of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent space, and the decoder reconstructs the data in the low-dimensional latent space into the original data. The input data is , the mapping function of the encoder is ,in is the activation function, is the weight matrix, is the bias vector;

[0036] The mapping function of the decoder is , by training the autoencoder to minimize the reconstruction error, that is ;

[0037] in, is the weight matrix, is the bias vector;

[0038] in, The encoder maps the data to a low-dimensional latent space and extracts the key features of the data. The decoder reconstructs the data and trains the autoencoder by minimizing the reconstruction error to achieve feature extraction and anomaly detection of the data. Abnormal data will produce large errors during reconstruction.

[0039] In the preferred embodiment, the specific method of step S3 is:

[0040] B1. Use LRGF neural network to establish a dynamic process model and apply the processed data analysis to the health management system. LRGF is a local recursive global feedforward neural network.

[0041] Through training mode and fault diagnosis mode, the equipment status is monitored in real time to identify the health status and potential problems of the equipment;

[0042] B2. Establish a fault prediction model, combine the analyzed multi-source data to identify possible equipment failures in advance, and predict the time and cause of their occurrence.

[0043] In the preferred embodiment, the specific method of steps B1-B2 is:

[0044] B01, LRGF neural network consists of local recursive layers and global feedforward layers. The local recursive layers are used to capture short-term dynamic patterns in time series data, and the global feedforward layers are used to integrate long-term trends and multi-source data information.

[0045] The state update formula of the local recursive layer is: ,in is the local recursive layer at time Status, is the activation function, is the local recursive weight matrix, is the input weight matrix, is the bias vector, Indicates at time Input data;

[0046] The output formula of the global feedforward layer is: ,in is the global feed-forward layer at time The output, is the global feedforward weight matrix, is the bias vector;

[0047] The state update formula of the local recursive layer is used to capture the short-term dynamics of time series data. By recursively updating the state, past information can be remembered;

[0048] The output formula of the global feedforward layer integrates the states of the local recursive layers and outputs the prediction result for the current time point;

[0049] B02. Training LRGF neural network:

[0050] Define the loss function and use the mean square error loss function;

[0051] Update network parameters using back-propagation algorithm and gradient descent method;

[0052] The loss function is used to measure the difference between the network output and the true value. The gradient is calculated through the back-propagation algorithm, and then the gradient descent method is used to update the network parameters, so that the network gradually learns the patterns in the data and improves the prediction accuracy.

[0053] B03. Based on the trained LRGF neural network, evaluate the health status of the equipment by analyzing the characteristics of the network output and input data and define health indicators , the formula is: ,in It is a function designed according to specific application scenarios, which is used to comprehensively consider the characteristics of network output and input data;

[0054] Health indicators are used to quantify the health status of the device. By comprehensively considering the characteristics of network output and input data, the operation of the device can be more comprehensively evaluated;

[0055] Health indicators can be determined based on the degree of deviation of the output value from the normal range and the changing trend of the input data;

[0056] B04. Use the predictive capabilities of the LRGF neural network and combine it with multi-source data analysis to identify possible equipment failures in advance;

[0057] For time series data, it is possible to determine whether a failure is likely to occur by predicting the output values ​​at several future time points;

[0058] Predicting the future The output at each time point is ,If the predicted value exceeds the normal range or shows a clear abnormal trend, it is considered that the equipment may fail;

[0059] At the same time, other features in multi-source data, including equipment operating parameters and environmental factors, are analyzed and combined with the prediction results of the LRGF neural network to further improve the accuracy of fault prediction;

[0060] Establish a comprehensive judgment model, the formula is: ,in represents the fault prediction result, H is the health index, is the feature vector of multi-source data, It is a function designed according to specific circumstances.

[0061] Predicting the output value at a future time point can detect possible equipment failures in advance. The comprehensive judgment model combines health indicators and multi-source data characteristics to improve the accuracy and reliability of fault prediction.

[0062] In the preferred embodiment, the specific method of step S5 is:

[0063] C1. The wavelet lifting algorithm includes two processes: decomposition and reconstruction, followed by prediction and update operations. The prediction operation formula is: ,in is the detail coefficient, which is the high frequency part, is the prediction operator;

[0064] The update operation formula is: ,in is the approximate coefficient, i.e. the low-frequency part;

[0065] During the reconstruction process, the inverse update operation formula is: , the inverse prediction operation formula is: ;

[0066] The prediction operation formula and the update operation formula are used to perform wavelet lifting decomposition and reconstruction on the signal. The different frequency components of the signal can be effectively extracted through the prediction and update operations.

[0067] C2. Wavelet lifting algorithm is applied to equipment residual signal analysis:

[0068] Collect the output signal of the equipment during normal operation as a reference signal , when the device is abnormal, collect the current output signal , calculate the residual signal ;

[0069] For the residual signal Apply wavelet lifting algorithm to decompose and select appropriate prediction operator and update operator , according to the above formula, multi-scale decomposition is performed to obtain detail coefficients and approximation coefficients at different scales.

[0070] Analyze the characteristics of detail coefficients and calculate the energy of detail coefficients ,If the energy exceeds a certain threshold, it indicates that a fault may exist;

[0071] C3. Combine knowledge graph to locate faults and determine their types:

[0072] Building a knowledge graph of equipment failures , where V is a node set, including equipment components, fault types, and symptom nodes; E is an edge set, representing the relationship between nodes;

[0073] When a possible fault is detected, the feature vector of the current monitoring data is extracted ,in represents a certain eigenvalue;

[0074] Calculate the similarity between the feature vector and each fault type node in the knowledge graph. The eigenvector of , the similarity calculation formula is:

[0075] ;

[0076] Based on the similarity results, the most likely fault type and the location where the fault occurred are determined; if the similarity exceeds a certain threshold, the fault type is considered to match the current device status;

[0077] C4. Determine the severity of the fault:

[0078] Define the fault severity index S1, which is determined based on multiple factors;

[0079] For a specific fault type, determine the weight vector of factors affecting the severity of the fault based on historical data and expert experience ;

[0080] Calculate the severity index of the current fault, the formula is: ,in It is Quantitative value of a factor.

[0081] In the preferred solution, the specific method of developing a decision support system in step S6 is:

[0082] D1. Collect operation monitoring data. At the same time, organize existing mature management methods and operation procedure documents, extract key information to form a rule set, and perform principal component analysis and dimensionality reduction on the operation monitoring data;

[0083] D2. Define equipment health status indicators, which are determined based on the reduced-dimensional operation monitoring data and preset thresholds; combine the rule set and equipment health status indicators to build a decision tree model;

[0084] Each node of the decision tree is divided according to a rule until a leaf node is reached, which represents a specific maintenance or repair decision;

[0085] D3. Use the semantic information provided by the knowledge graph to associate the operation monitoring data with equipment components and fault type entities;

[0086] The reinforcement learning algorithm is used to optimize the operation and maintenance strategy. The state space S2 is defined as various state combinations of the equipment, and the action space A is the possible operation and maintenance operations. At each time step t, the agent performs operations according to the current state. Select an action , get reward after executing the action ;

[0087] By continuously interacting with the environment, the agent learns the optimal policy function , which maximizes the long-term cumulative reward, Represents the agent in state Specific operation and maintenance actions taken when

[0088] In the preferred embodiment, the specific method of step S7 is:

[0089] E1. Knowledge graph construction: Determine the entities and relationships of the knowledge graph; entities include equipment parts, fault types, and maintenance operations, and relationships include "belong to", "may cause", and "need to be performed";

[0090] Extract entity and relationship information from operation monitoring data, maintenance records, and expert experience data sources;

[0091] Use Neo4j graph database to store knowledge graph;

[0092] E2. Visual management: Use force-directed layout algorithm to visualize the knowledge graph. The location is , the attraction formula between nodes is ,in is the attraction coefficient, Is a node and The distance between

[0093] The formula for the repulsive force between nodes is: , where c is the repulsive force coefficient. By continuously iteratively calculating the position of the nodes, the entire knowledge graph reaches a balanced state;

[0094] Add visual attributes such as color and size to nodes and edges to highlight different types of entities and relationships;

[0095] E3. Build a question-answering system:

[0096] Build a question-answering system framework based on knowledge graph, including question parsing module, query generation module and answer generation module;

[0097] The question parsing module converts natural language questions into structured query statements;

[0098] The query generation module sends the structured query statement to the graph database for query and returns the query results;

[0099] The answer generation module converts query results into natural language answers.

[0100] The present invention provides a health management method for thermal power generation equipment based on knowledge graph and mechanism model, which realizes real-time health monitoring and fault prediction of thermal power generation equipment, can timely discover potential problems of equipment, avoid sudden occurrence of faults, and ensure stable operation of equipment.

[0101] Intelligent means are used to reduce maintenance costs, unnecessary manual inspections and regular maintenance costs, while improving the pertinence and effectiveness of maintenance.

[0102] It improves the reliability and service life of the equipment. By predicting failures in advance and taking corresponding maintenance measures, it reduces equipment damage and wear and extends the service life of the equipment.

[0103] It has enhanced the adaptability of the electricity market, and can flexibly adjust power generation plans according to market demand and equipment status, thereby improving the competitiveness of enterprises in the electricity market.

[0104] It improves the economic benefits of enterprise operations, reduces maintenance costs, improves equipment reliability and power generation efficiency, and thus increases enterprise profits. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0106] Figure 1It is a flow chart of the health management method of thermal power generation equipment of the present invention;

[0107] Figure 2 It is a flow chart of the integration of vector database, SQL database and knowledge graph data of the present invention. DETAILED DESCRIPTION

[0108] Example 1

[0109] like Figure 1-2 As shown, a thermal power generation equipment health management method based on knowledge graph and mechanism model includes:

[0110] S1. Collect equipment operation data in real time or regularly;

[0111] S2, process and analyze data to identify equipment health status and potential problems;

[0112] S3, using models such as local recursive global feedforward neural networks to train models based on historical data to predict future failures;

[0113] S4. Arrange regular inspections and necessary parts replacements based on forecast results to avoid unplanned downtime;

[0114] S5. When an abnormality is detected, the residual signal is analyzed using the wavelet lifting algorithm, the fault features are extracted, and compared with the historical data;

[0115] By analyzing monitoring data and combining knowledge graph tools, equipment faults can be identified and located, and the type and severity of the faults can be accurately determined;

[0116] S6. Develop a decision support system based on existing mature management methods and operating procedures;

[0117] S7. Build a knowledge graph framework for fault operation and maintenance to realize visual management of knowledge and a question-answering system based on the knowledge graph.

[0118] Example 2

[0119] Further described in conjunction with Example 1, the specific method of step S1 is:

[0120] Deploy temperature sensors, vibration sensors, pressure sensors, and current sensors on each device and connect them to the central data acquisition system through IoT technology to collect device operation data in real time or regularly. The collected data will be directly used for the next step of data analysis and processing;

[0121] Use MQTT or CoAP lightweight protocols for IoT environments;

[0122] Use Apache Kafka or Amazon Kinesis to process large amounts of sensor data.

[0123] In the preferred embodiment, the specific method of step S2 is:

[0124] A1. Organize and store the data collected in step S1 with the help of big data technology, and use data analysis and artificial intelligence to deeply process and analyze the data;

[0125] Among them, the collected data is first cleaned and standardized to remove outliers and noise;

[0126] The box plot method was used to detect outliers, and data points that exceeded the range of 1.5 times the interquartile range of the upper and lower quartiles were treated as outliers;

[0127] formula: ,in is the lower quartile, is the upper quartile, . Use: Used to determine the normal range of data in order to identify and handle outliers.

[0128] Then the data is standardized so that data with different characteristics have the same scale. The Z-score standardization method is used to make data with different dimensions comparable, which is convenient for subsequent data analysis and algorithm processing.

[0129] A2. Through multi-source information fusion, vector database, SQL database and knowledge graph tools are used to integrate data and knowledge.

[0130] In the preferred embodiment, the specific method of step A2 is:

[0131] A21. Multi-source information fusion: Extract vector feature data from the vector database, obtain structured data from the SQL database, extract semantic relationship data from the knowledge graph, and use a weighted fusion method to fuse multi-source data. Suppose the fused data is: ,in:

[0132] ;

[0133] is the weight coefficient, satisfying ;

[0134] in, , , Represents the number of data feature dimensions extracted from the vector database, SQL database, and knowledge graph respectively;

[0135] By weighted fusion of data from different sources, the advantages of each data source can be fully utilized to provide more comprehensive information for subsequent analysis;

[0136] A22. Use the principal component analysis method to reduce the dimension and extract features of the fused data. Suppose the fused data matrix is Calculate its covariance matrix ;

[0137] Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue and the corresponding eigenvector , select the first The eigenvectors form the projection matrix , project the original data into the low-dimensional space to obtain the reduced-dimensional data ;

[0138] in, The covariance matrix is ​​used to measure the correlation of data, and the eigenvectors obtained by eigenvalue decomposition are used to construct a projection matrix to project high-dimensional data into a low-dimensional space, achieve dimensionality reduction and feature extraction, reduce data complexity, and improve the efficiency of subsequent analysis;

[0139] A23. Introduce the autoencoder in deep learning to further extract features and detect anomalies in the data after dimensionality reduction;

[0140] The autoencoder consists of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent space, and the decoder reconstructs the data in the low-dimensional latent space into the original data. The input data is , the mapping function of the encoder is ,in is the activation function, is the weight matrix, is the bias vector;

[0141] The mapping function of the decoder is , by training the autoencoder to minimize the reconstruction error, that is ;

[0142] in, is the weight matrix, is the bias vector;

[0143] in, The encoder maps the data to a low-dimensional latent space and extracts the key features of the data. The decoder reconstructs the data and trains the autoencoder by minimizing the reconstruction error to achieve feature extraction and anomaly detection of the data. Abnormal data will produce large errors during reconstruction.

[0144] is the weight matrix, which determines the linear transformation relationship from the low-dimensional latent space to the reconstructed space. In the process of mapping the data in the low-dimensional latent space back to the original data space, For the input low-dimensional data When performing weighted operations, different weight values ​​will change the way the data is scaled and combined in each dimension, thus affecting the final reconstruction result. For example, if the weight on a certain dimension is large, the feature corresponding to that dimension will be more prominent during reconstruction; conversely, if the weight is small, its impact on the reconstruction result will be relatively weak.

[0145] is the bias vector, which plays a role of translation. In the linear transformation process of data, the bias vector A fixed offset is provided for the reconstruction result. This makes the reconstruction process not only rely on the linear combination of the input data, but also can be adjusted in different directions according to the offset value, which helps to better fit the distribution characteristics of the original data and improve the accuracy of the reconstruction.

[0146] By adjusting and The value of , makes the reconstruction error Minimize the error, thereby achieving effective reconstruction and feature extraction of the data, and can use the reconstruction error to detect anomalies in the data, because abnormal data often produce large errors when reconstructed.

[0147] Example 3

[0148] Further described in conjunction with Example 1, the specific method of step S3 is:

[0149] B1. Use LRGF neural network to establish a dynamic process model and apply the processed data analysis to the health management system. LRGF is a local recursive global feedforward neural network.

[0150] Through training mode and fault diagnosis mode, the equipment status is monitored in real time to identify the health status and potential problems of the equipment;

[0151] B2. Establish a fault prediction model, combine the analyzed multi-source data to identify possible equipment failures in advance, and predict the time and cause of their occurrence.

[0152] In the preferred embodiment, the specific method of steps B1-B2 is:

[0153] B01, LRGF neural network consists of local recursive layers and global feedforward layers. The local recursive layers are used to capture short-term dynamic patterns in time series data, and the global feedforward layers are used to integrate long-term trends and multi-source data information.

[0154] Assume that the input time series data is ,in Indicates at time of -dimensional feature vector.

[0155] The state update formula of the local recursive layer is: ,in is the local recursive layer at time Status, is the activation function, is the local recursive weight matrix, is the input weight matrix, is the bias vector, Indicates at time Input data;

[0156] Indicates at time Input data: It is the external information received by the neural network at the current time step. These input data carry relevant feature information such as the operating status of the equipment, and play a key role in capturing short-term dynamic patterns in time series data in the local recursive layer. By inputting the weight matrix With the local recursive weight matrix , the previous moment state and the bias vector Working together, through the activation function After processing, update to get the current time The local recursive layer state of , so that the network can continuously learn and adapt to the changing patterns of input data, and realize effective monitoring of equipment operating status and identification of potential problems.

[0157] The output formula of the global feedforward layer is: ,in is the global feed-forward layer at time The output, is the global feedforward weight matrix, is the bias vector;

[0158] formula: , . Use: The state update formula of the local recursive layer is used to capture the short-term dynamics of time series data. By recursively updating the state, past information can be remembered. The output formula of the global feedforward layer integrates the state of the local recursive layer and outputs the prediction result for the current time point.

[0159] B02. Training LRGF neural network:

[0160] Define the loss function and use the mean square error loss function; the formula is: ,in is the true value.

[0161] Use the back propagation algorithm and gradient descent method to update the network parameters; calculate the gradient of the loss function with respect to each parameter, that is, , , , , , and then update the parameters according to the gradient, the formula is: , , , , ,in is the learning rate.

[0162] The loss function is used to measure the difference between the network output and the true value. The gradient is calculated through the back-propagation algorithm, and then the gradient descent method is used to update the network parameters, so that the network gradually learns the patterns in the data and improves the prediction accuracy.

[0163] B03. Based on the trained LRGF neural network, evaluate the health status of the equipment by analyzing the characteristics of the network output and input data and define health indicators , the formula is: ,in It is a function designed according to specific application scenarios, which is used to comprehensively consider the characteristics of network output and input data;

[0164] Health indicators are used to quantify the health status of the device. By comprehensively considering the characteristics of network output and input data, the operation of the device can be more comprehensively evaluated;

[0165] Health indicators can be determined based on the degree of deviation of the output value from the normal range and the changing trend of the input data;

[0166] B04. Use the predictive capabilities of the LRGF neural network and combine it with multi-source data analysis to identify possible equipment failures in advance;

[0167] For time series data, it is possible to determine whether a failure is likely to occur by predicting the output values ​​at several future time points;

[0168] Predicting the future The output at each time point is ,If the predicted value exceeds the normal range or shows a clear abnormal trend, it is considered that the equipment may fail;

[0169] At the same time, other features in multi-source data, including equipment operating parameters and environmental factors, are analyzed and combined with the prediction results of the LRGF neural network to further improve the accuracy of fault prediction;

[0170] Establish a comprehensive judgment model, the formula is: ,in represents the fault prediction result, H is the health index, is the feature vector of multi-source data, It is a function designed according to specific circumstances.

[0171] Predicting the output value at a future time point can detect possible equipment failures in advance. The comprehensive judgment model combines health indicators and multi-source data characteristics to improve the accuracy and reliability of fault prediction.

[0172] Example 4

[0173] Further described in conjunction with Example 1, the specific method of step S5 is:

[0174] Wavelet lifting algorithm is a modern wavelet transform implementation method, which is flexible and efficient.

[0175] C1. Wavelet lifting algorithm includes two processes: decomposition and reconstruction. Suppose the original signal is In the decomposition process, the splitting operation is first performed to divide the signal into an even sample sequence and odd sample sequence ,Right now , .

[0176] Then perform prediction and update operations. The prediction operation formula is: ,in is the detail coefficient, which is the high frequency part, is the prediction operator;

[0177] The update operation formula is: ,in is the approximate coefficient, i.e. the low-frequency part;

[0178] During the reconstruction process, the inverse update operation formula is: , the inverse prediction operation formula is: ;

[0179] The prediction operation formula and the update operation formula are used to perform wavelet lifting decomposition and reconstruction on the signal. The different frequency components of the signal can be effectively extracted through the prediction and update operations.

[0180] formula:

[0181] .

[0182] These formulas are used to perform wavelet lifting decomposition and reconstruction on the signal. The different frequency components of the signal can be effectively extracted through prediction and update operations. The detail coefficient can reflect the high-frequency changes of the signal, and the approximate coefficient can reflect the low-frequency trend of the signal. It is very useful for analyzing the residual signal output by the device.

[0183] C2. Wavelet lifting algorithm is applied to equipment residual signal analysis:

[0184] Collect the output signal of the equipment during normal operation as a reference signal , when the device is abnormal, collect the current output signal , calculate the residual signal ;

[0185] For the residual signal Apply wavelet lifting algorithm to decompose and select appropriate prediction operator and update operator , according to the above formula, multi-scale decomposition is performed to obtain detail coefficients and approximation coefficients at different scales.

[0186] Analyze the characteristics of detail coefficients and calculate the energy of detail coefficients ,If the energy exceeds a certain threshold, it indicates that a fault may exist;

[0187] formula: .

[0188] The calculation of the residual signal is used to highlight the changes when the equipment is abnormal. The calculation of the detail coefficient energy can be used as an indicator for fault detection. When the energy exceeds the threshold, it indicates that a fault may exist.

[0189] C3. Combine knowledge graph to locate faults and determine their types:

[0190] Building a knowledge graph of equipment failures , where V is a node set, including equipment components, fault types, and symptom nodes; E is an edge set, representing the relationship between nodes;

[0191] For example, equipment component node Node with fault type The edges between them represent the possible fault types of the component.

[0192] When a possible fault is detected, the feature vector of the current monitoring data is extracted ,in represents a certain eigenvalue;

[0193] Calculate the similarity between the feature vector and each fault type node in the knowledge graph. The eigenvector of , the similarity calculation formula is:

[0194] ;

[0195] Based on the similarity results, the most likely fault type and the location where the fault occurred are determined; if the similarity exceeds a certain threshold, the fault type is considered to match the current device status;

[0196] The similarity calculation formula is used to calculate the similarity between the monitoring data feature vector and the fault type node feature vector in the knowledge graph. By comparing the similarity, the most likely fault type and location can be determined. Combined with the knowledge graph, the existing equipment fault knowledge can be fully utilized to improve the accuracy of fault diagnosis.

[0197] C4. Determine the severity of the fault:

[0198] The fault severity index S1 is defined and determined based on a combination of multiple factors; for example: the impact of the fault on equipment performance, the frequency of the fault, etc.

[0199] For a specific fault type, determine the weight vector of factors affecting the severity of the fault based on historical data and expert experience ;

[0200] Calculate the severity index of the current fault, the formula is: ,in It is Quantitative value of a factor.

[0201] formula It is used to calculate the severity index of the fault. By comprehensively considering multiple factors and performing weighted summation according to the weight vector, the severity of the fault can be judged more accurately, providing a basis for subsequent maintenance decisions.

[0202] Example 5

[0203] Further illustrate with reference to Example 1, Figure 1-2 The structure shown in FIG. 1 is as follows: Step S6 is a specific method for developing a decision support system:

[0204] D1. Collect operation monitoring data. At the same time, organize existing mature management methods and operation procedure documents, extract key information to form a rule set, and perform principal component analysis and dimensionality reduction on the operation monitoring data;

[0205] Collecting operational monitoring data ,in Indicates At the same time, we sort out the existing mature management methods and operating procedures, extract key information and form a set of rules. .

[0206] The principal component analysis (PCA) is used to reduce the dimension of the operation monitoring data. Assume that the data matrix is , calculate the covariance matrix . Perform eigenvalue decomposition on the covariance matrix and obtain the eigenvalue and the corresponding eigenvector . Select the first The eigenvectors form the projection matrix , project the original data into the low-dimensional space to obtain the reduced-dimensional data .

[0207] formula: The covariance matrix is ​​used to measure the correlation of the data, and the eigenvectors obtained by eigenvalue decomposition are used to construct the projection matrix, which projects the high-dimensional data into a low-dimensional space, reducing the complexity of the data and facilitating subsequent analysis.

[0208] D2. Define equipment health status indicators, which are determined based on the reduced-dimensional operation monitoring data and preset thresholds; combine the rule set and equipment health status indicators to build a decision tree model;

[0209] Each node of the decision tree is divided according to a rule until a leaf node is reached, which represents a specific maintenance or repair decision;

[0210] D3. Use the semantic information provided by the knowledge graph to associate the operation monitoring data with equipment components and fault type entities;

[0211] The reinforcement learning algorithm is used to optimize the operation and maintenance strategy. The state space S2 is defined as various state combinations of the equipment, and the action space A is the possible operation and maintenance operations. At each time step t, the agent performs operations according to the current state. Select an action , get reward after executing the action ;

[0212] By continuously interacting with the environment, the agent learns the optimal policy function , maximizing the long-term cumulative rewards.

[0213] In the context of reinforcement learning, the agent follows a policy function In a given state Next select action Here Represents the agent in state Specific operation and maintenance actions taken when

[0214] In the thermal power equipment health management system, action space It includes a series of possible operation and maintenance operations, such as arranging equipment inspections, replacing specific parts, adjusting equipment operating parameters, etc. The agent learns the policy function and makes decisions in different equipment states (i.e., state space). Different states in ) and select the appropriate action , and continuously optimize the strategy based on the feedback (rewards) after executing the action to maximize the long-term cumulative rewards, thereby achieving the goals of optimizing equipment operation and maintenance strategies, improving equipment reliability and operating efficiency, etc.

[0215] The specific detailed method is: using the semantic information provided by the knowledge graph, the operation monitoring data is associated with entities such as equipment components and fault types. Suppose the knowledge graph is ,in is a collection of nodes, is an edge set. For monitoring data points , find its corresponding node in the knowledge graph , and obtain relevant semantic information through edge relationships.

[0216] Formula: State Value Function ,in is the discount factor. The state value function is used to evaluate the long-term value of being in a certain state under a specific strategy. By optimizing the state value function, the agent can learn the optimal operation and maintenance strategy.

[0217] In the preferred embodiment, the specific method of step S7 is:

[0218] E1. Knowledge graph construction: Determine the entities and relationships of the knowledge graph; entities include equipment parts, fault types, and maintenance operations, and relationships include "belong to", "may cause", and "need to be performed";

[0219] Extract entity and relationship information from operation monitoring data, maintenance records, and expert experience data sources; for example, for fault data points ,If its characteristics indicate that it is a fault of a specific equipment component, the equipment component entity and the fault type entity are created in the knowledge graph, and a “belongs to” relationship is established.

[0220] Use Neo4j graph database to store knowledge graph; knowledge graph can be represented as , where V is the set of nodes, E is the set of edges, and L is the set of labels for nodes and edges.

[0221] E2. Visual management: Use force-directed layout algorithm to visualize the knowledge graph. The location is , the attraction formula between nodes is ,in is the attraction coefficient, Is a node and The distance between

[0222] The formula for the repulsive force between nodes is: , where c is the repulsive force coefficient. By continuously iteratively calculating the position of the nodes, the entire knowledge graph reaches a balanced state;

[0223] Add visual attributes such as color and size to nodes and edges to highlight different types of entities and relationships;

[0224] formula:

[0225] The attraction formula and repulsion formula are used to calculate the forces between nodes. The force-directed layout algorithm can make the visualization of the knowledge graph clearer and easier for operation and maintenance personnel to understand and analyze.

[0226] E3. Build a question-answering system:

[0227] Build a question-answering system framework based on knowledge graph, including question parsing module, query generation module and answer generation module.

[0228] The question parsing module converts natural language questions into structured query statements. For example, for the question "What is the cause of the failure of device X?", it is parsed into the query statement "MATCH (d: device part {name: 'device X'})-[r: may cause]->(f: failure type) RETURN f".

[0229] The query generation module sends the structured query statement to the graph database for query and returns the query results.

[0230] The answer generation module converts the query result into a natural language answer. For example, if the query result is the fault type node "circuit fault", the answer is "the cause of the fault of device X may be circuit fault".

[0231] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A thermal power equipment health management method based on knowledge graph and mechanism model, characterized by: The method includes: S1. Collect equipment operation data in real time or regularly; S2, process and analyze data to identify equipment health status and potential problems; S3, using a local recursive global feedforward neural network model, training the model based on historical data to predict future faults, step S3 method is: B01, LRGF neural network consists of local recursive layers and global feedforward layers. The local recursive layers are used to capture short-term dynamic patterns in time series data, and the global feedforward layers are used to integrate long-term trends and multi-source data information. The state update formula of the local recursive layer is: ,in is the local recursive layer at time Status, is the activation function, is the local recursive weight matrix, is the input weight matrix, is the bias vector, Indicates at time Input data; The output formula of the global feedforward layer is: ,in is the global feed-forward layer at time The output, is the global feedforward weight matrix, is the bias vector; The state update formula of the local recursive layer is used to capture the short-term dynamics of time series data. By recursively updating the state, past information can be remembered; The output formula of the global feedforward layer integrates the states of the local recursive layers and outputs the prediction result for the current time point; B02. Training LRGF neural network: Define the loss function and use the mean square error loss function; Update network parameters using back-propagation algorithm and gradient descent method; The loss function is used to measure the difference between the network output and the true value. The gradient is calculated through the back-propagation algorithm, and then the gradient descent method is used to update the network parameters. B03. Based on the trained LRGF neural network, evaluate the health status of the equipment by analyzing the characteristics of the network output and input data and define health indicators , the formula is: ,in It is a function designed according to specific application scenarios, which is used to comprehensively consider the characteristics of network output and input data; Health indicators are used to quantify the health status of the device; B04. Use the predictive capabilities of the LRGF neural network and combine it with multi-source data analysis to identify possible equipment failures in advance; For time series data, it is possible to determine whether a failure is likely to occur by predicting the output values ​​at several future time points; Predicting the future The output at each time point is ,If the predicted value exceeds the normal range or shows a clear abnormal trend, it is considered that the equipment may fail; At the same time, other features in multi-source data, including equipment operating parameters and environmental factors, are analyzed and combined with the prediction results of the LRGF neural network to further improve the accuracy of fault prediction; Establish a comprehensive judgment model, the formula is: ,in represents the fault prediction result, H is the health index, is the feature vector of multi-source data, It is a function designed according to specific circumstances; S4. Arrange regular inspections and necessary parts replacements based on forecast results to avoid unplanned downtime; S5. When an abnormality is detected, the residual signal is analyzed using the wavelet lifting algorithm, the fault features are extracted, and compared with the historical data; By analyzing monitoring data and combining knowledge graph tools, equipment faults can be identified and located, and the type and severity of the faults can be accurately determined; S6. Develop a decision support system based on existing mature management methods and operating procedures; S7. Build a knowledge graph framework for fault operation and maintenance to realize visual management of knowledge and a question-answering system based on the knowledge graph.

2. According to claim 1, a thermal power generation equipment health management method based on knowledge graph and mechanism model is characterized by: The specific method of step S1 is: Deploy temperature sensors, vibration sensors, pressure sensors, and current sensors on each device and connect them to the central data acquisition system through IoT technology to collect device operation data in real time or regularly. The collected data will be directly used for the next step of data analysis and processing; Use MQTT or CoAP lightweight protocols for IoT environments; Use Apache Kafka or Amazon Kinesis to process large amounts of sensor data.

3. According to claim 1, a thermal power generation equipment health management method based on knowledge graph and mechanism model is characterized by: The specific method of step S2 is: A1. Organize and store the data collected in step S1 with the help of big data technology, and use data analysis and artificial intelligence to deeply process and analyze the data; Among them, the collected data is first cleaned and standardized to remove outliers and noise; The box plot method was used to detect outliers, and data points that exceeded the range of 1.5 times the interquartile range of the upper and lower quartiles were treated as outliers; Then the data is standardized so that data with different characteristics have the same scale. The Z-score standardization method is used to make data with different dimensions comparable, which is convenient for subsequent data analysis and algorithm processing. A2. Through multi-source information fusion, vector database, SQL database and knowledge graph tools are used to integrate data and knowledge.

4. According to claim 3, a thermal power generation equipment health management method based on knowledge graph and mechanism model is characterized by: The specific method of step A2 is: A21. Multi-source information fusion: Extract vector feature data from the vector database, obtain structured data from the SQL database, extract semantic relationship data from the knowledge graph, and use a weighted fusion method to fuse multi-source data. Suppose the fused data is: ,in: ； is the weight coefficient, satisfying ; in, , , Represents the number of data feature dimensions extracted from the vector database, SQL database, and knowledge graph respectively; By weighted fusion of data from different sources, the advantages of each data source can be fully utilized to provide more comprehensive information for subsequent analysis; A22. Use the principal component analysis method to reduce the dimension and extract features of the fused data. Suppose the fused data matrix is Calculate its covariance matrix ; Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue and the corresponding eigenvector , select the first The eigenvectors form the projection matrix , project the original data into the low-dimensional space to obtain the reduced-dimensional data ; in, The covariance matrix is ​​used to measure the correlation of data, and the eigenvectors obtained by eigenvalue decomposition are used to construct a projection matrix to project high-dimensional data into a low-dimensional space, achieve dimensionality reduction and feature extraction, reduce data complexity, and improve the efficiency of subsequent analysis; A23. Introduce the autoencoder in deep learning to further extract features and detect anomalies in the data after dimensionality reduction; The autoencoder consists of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent space, and the decoder reconstructs the data in the low-dimensional latent space into the original data. The input data is , the mapping function of the encoder is ,in is the activation function, is the weight matrix, is the bias vector; The mapping function of the decoder is , by training the autoencoder to minimize the reconstruction error, that is ; in, is the weight matrix, is the bias vector; in, The encoder maps the data to a low-dimensional latent space and extracts the key features of the data. The decoder reconstructs the data and trains the autoencoder by minimizing the reconstruction error to achieve feature extraction and anomaly detection of the data. Abnormal data will produce large errors during reconstruction.

5. According to claim 1, a thermal power generation equipment health management method based on knowledge graph and mechanism model is characterized by: The specific method of step S5 is: C1. The wavelet lifting algorithm includes two processes: decomposition and reconstruction, followed by prediction and update operations. The prediction operation formula is: ,in is the detail coefficient, which is the high frequency part, is the prediction operator; The update operation formula is: ,in is the approximate coefficient, i.e. the low-frequency part; During the reconstruction process, the inverse update operation formula is: , the inverse prediction operation formula is: ; The prediction operation formula and the update operation formula are used to perform wavelet lifting decomposition and reconstruction on the signal. The different frequency components of the signal can be effectively extracted through the prediction and update operations. C2. Wavelet lifting algorithm is applied to equipment residual signal analysis: Collect the output signal of the equipment during normal operation as a reference signal , when the device is abnormal, collect the current output signal , calculate the residual signal ; For the residual signal Apply wavelet lifting algorithm to decompose and select appropriate prediction operator and update operator , perform multi-scale decomposition according to the above formula to obtain detail coefficients and approximate coefficients at different scales; Analyze the characteristics of detail coefficients and calculate the energy of detail coefficients ,If the energy exceeds a certain threshold, it indicates that a fault may exist; C3. Combine knowledge graph to locate faults and determine their types: Building a knowledge graph of equipment failures , where V is a node set, including equipment components, fault types, and symptom nodes; E is an edge set, representing the relationship between nodes; When a possible fault is detected, the feature vector of the current monitoring data is extracted ,in represents a certain eigenvalue; Calculate the similarity between the feature vector and each fault type node in the knowledge graph. The eigenvector of , the similarity calculation formula is: ； Based on the similarity results, determine the most likely fault type and the location where the fault occurs; If the similarity exceeds a certain threshold, the fault type is considered to match the current device status; C4. Determine the severity of the fault: Define the fault severity index S1, which is determined based on multiple factors; For a specific fault type, determine the weight vector of factors affecting the severity of the fault based on historical data and expert experience ; Calculate the severity index of the current fault, the formula is: ,in It is Quantitative value of a factor.

6. According to the thermal power equipment health management method based on knowledge graph and mechanism model as described in claim 1, The characteristic is that: the specific method of developing the decision support system in step S6 is: D1. Collect operation monitoring data. At the same time, organize existing mature management methods and operation procedure documents, extract key information to form a rule set, and perform principal component analysis and dimensionality reduction on the operation monitoring data; D2. Define equipment health status indicators, which are determined based on the reduced-dimensional operation monitoring data and preset thresholds; combine the rule set and equipment health status indicators to build a decision tree model; Each node of the decision tree is divided according to a rule until a leaf node is reached, which represents a specific maintenance or repair decision; D3. Use the semantic information provided by the knowledge graph to associate the operation monitoring data with equipment components and fault type entities; The reinforcement learning algorithm is used to optimize the operation and maintenance strategy. The state space S2 is defined as various state combinations of the equipment, and the action space A is the possible operation and maintenance operations. At each time step t, the agent performs operations according to the current state. Select an action , get reward after executing the action ; By continuously interacting with the environment, the agent learns the optimal policy function , which maximizes the long-term cumulative reward, Represents the agent in state Specific operation and maintenance actions taken when 7. According to claim 1, a thermal power generation equipment health management method based on knowledge graph and mechanism model is characterized by: The specific method of step S7 is: E1. Knowledge graph construction: determine the entities and relationships of the knowledge graph; entities include equipment components, fault types, and maintenance operations, and relationships include "belong to", "may cause", and "need to be performed"; Extract entity and relationship information from operation monitoring data, maintenance records, and expert experience data sources; Use Neo4j graph database to store knowledge graph; E2. Visual management: Use force-directed layout algorithm to visualize the knowledge graph. The location is , the attraction formula between nodes is ,in is the attraction coefficient, Is a node and The distance between The formula for the repulsive force between nodes is: , where c is the repulsive force coefficient. By continuously iteratively calculating the position of the nodes, the entire knowledge graph reaches a balanced state; Add color and size visualization attributes to nodes and edges to highlight different types of entities and relationships; E3. Build a question-answering system: Build a question-answering system framework based on knowledge graph, including question parsing module, query generation module and answer generation module; The question parsing module converts natural language questions into structured query statements; The query generation module sends the structured query statement to the graph database for query and returns the query results; The answer generation module converts query results into natural language answers.

Images