Power generation equipment fault diagnosis method based on large language model

Abstract

The invention discloses a power generation equipment fault diagnosis method based on a large language model. According to the method, firstly, an expert knowledge base is converted into a fault knowledge base composed of criterion nodes and edges, and fault retrieval is completed through the knowledge base; synthesis of natural language texts is completed by using input based on text function segmentation, cue word templates are constructed based on calculation types of different criteria, and a large model is promoted to complete calculation reasoning tasks of the criteria; multi-dimensional equipment state monitoring time sequence data are input, a cue word set at the current time is automatically generated in combination with the cue word template, the cue word set is input into a large language model for computational reasoning, and a current state set of equipment is generated; using a knowledge base to retrieve the state set to obtain a fault diagnosis result; and then a visualization system is designed to collect and visualize a large model output result and a fault diagnosis result, and input and interaction interfaces are provided for cue word template construction. The fault diagnosis completed by the method has good practicability and expansibility.

Description

Technical Field

[0001] The invention belongs to the field of information technology, and in particular relates to a method for diagnosing faults of power generation equipment based on a large language model. Background Art

[0002] With the development of information technology and the popularization of sensors, power plants can obtain and collect massive amounts of multi-dimensional equipment monitoring time series data, such as motor current, steam temperature, and oil tank temperature. In order to improve the efficiency of fault detection and diagnostic analysis, researchers have proposed traditional data-based methods, deep learning-based methods, and model-based methods. However, traditional data-based methods require domain expertise and a lot of effort to support the feature extraction process. Deep learning-based methods rely heavily on expertise in signal processing and deep learning, and also bring high costs for training models.

[0003] Therefore, in actual production applications, most smart power plants adopt a model-based automated fault detection method. This method solidifies expert knowledge into computer-executable software code and obtains judgment results by inputting multi-dimensional time series data. Among them, expert knowledge refers to a series of rules and knowledge summarized by experienced staff during equipment maintenance and inspection. The criteria in the criterion knowledge formed by expert knowledge are manually programmed into computer-recognizable codes and embedded in the fault diagnosis system. Although this method can efficiently perform preset specific diagnostic tasks through code, it still has some defects. First, the model-based method is highly dependent on professional knowledge in terms of performance (WJ Li, HL, S. Gu, T. Chen, Process fault diagnosis with model- and knowledge-basedapproaches: Advances and opportunities, Control Engineering Practice. 105(11) (2020) 104637). It takes a lot of manual work to accurately correspond and solidify expert knowledge into code. Secondly, the update of the knowledge base increases the difficulty of code maintenance. When a certain point of the expert criterion is adjusted or added, such as: changes in the threshold judgment range, changes in coefficients, and relationship combination patterns of faults, it is necessary to find the corresponding code for modification. This consumes a lot of manpower and will cause the fault diagnosis system to be updated or even shut down, resulting in possible huge losses. In addition, even if the equipment is the same in different power plants, the expert knowledge base will be different under different production conditions, and code migration will not be possible, which will also cause the above-mentioned code conversion and maintenance problems. These problems limit the scalability and practicality of power equipment fault diagnosis. Summary of the invention

[0004] The purpose of the present invention is to provide a method for diagnosing faults of power generation equipment based on a large language model.

[0005] The method of the present invention comprises the following steps: Step (1) converting the existing power plant expert knowledge base into a fault knowledge base, the fault knowledge base includes multiple fault complete graphs, each fault complete graph is represented as a fault, and is visualized in the form of a fault complete graph, with different colors distinguishing the calculation type of each criterion node; Step (2) sets corresponding prompt word templates according to different computing types to automatically generate prompt words; the prompt word template includes four parts: The head is used to describe the task requirements and specify the output of the large model; The external knowledge part, knowledge, contains the preliminary knowledge needed to perform the task; The example section contains multiple examples of performing tasks. Synthesize some of the formats and connect other parts and input data using natural language; Step (3) setting a tag matching module, wherein the judgment status of the input data includes normal false, abnormal true, and error error, and the tag matching module synthesizes the data with the judgment status of abnormal true into an abnormal status set; Step (4) connects the prompt word template to the front end of the large language model, connects the tag matching module to the back end of the large language model, and the prompt word template, the large language model and the tag matching module constitute an equipment state reasoning module; inputs the multi-dimensional equipment state monitoring time series data into the equipment state reasoning module, the prompt word template generates a prompt word set at the current moment, the large language model calculates and infers the prompt word set, generates the state of the power plant power generation equipment at the current moment, and expresses it in the form of a tag language combined with a criterion state; the tag matching module synthesizes the abnormal state set at the current moment according to the criterion state, and expresses it as a set of multiple criterion nodes; Step (5) inputting the abnormal state set at the current moment into the fault knowledge base, performing fault retrieval, and obtaining the fault diagnosis result; Step (6) constructs a power plant equipment fault diagnosis visualization analysis system based on the equipment state reasoning module and the fault knowledge base; the system interface includes the following parts: (6.1) Line graph of multi-dimensional time series data: When multi-dimensional time series data enters the system, it is classified by its attributes and plotted on different line graphs; (6.2) Fault knowledge graph: The fault knowledge base is visualized in the form of a knowledge graph. Different colors of criterion nodes represent different calculation types, and detailed information of criterion nodes can be viewed by hovering. (6.3) Fault diagnosis view: This is a flow view that allows users to view and analyze the abnormalities that occur at each moment in the device. The abnormality criteria that occur at each moment are represented by rectangles and stacked, and different colors are used to distinguish the calculation types. (6.4) Prompt word template building component: provides multiple input windows corresponding to different segmented parts. After the user completes all parts, a complete prompt word is formed and displayed; (6.5) Prompt word evaluation component: Provides prompt word performance feedback to users to help them optimize the prompt word. The component consists of text preview, bidirectional bar chart, reply projection view, reply detail display, and comparison view.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention integrates technologies such as a large language model, a knowledge base, visualization, and human-computer interaction, manages and maintains expert knowledge in the form of a knowledge base, and is conducive to updating and expanding the fault model; an end-to-end large language model is used to undertake the computational reasoning tasks of the judgment criteria, thereby reducing the need for signal processing or machine learning expertise and data sets; prompt word templates are used to ensure the efficiency and accuracy of the large language model output, adapt to most fault detection tasks, and have strong portability between power plants with structured or semi-structured expert knowledge; the original time series data, the knowledge base, and the collected large model responses are visualized and human-computer interaction is provided, which is convenient for viewing fault details and making fault-tolerant responses; the use of interactive input templates is conducive to the construction of prompt word templates; the use of richly designed prompt word evaluation templates is conducive to the optimization of prompt word templates. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the workflow of the present invention; Figure 2 is a schematic diagram of a visualization system of the present invention; Figure 3 is a schematic diagram of the quantitative evaluation results of the present invention; Figure 4 It is a schematic diagram of an application example of the present invention. DETAILED DESCRIPTION

[0008] The power generation equipment fault diagnosis method based on the large language model of the present invention is further described below in conjunction with the accompanying drawings.

[0009] like Figure 1 As shown, the specific process includes the following steps: Step (1) converts the existing power plant expert knowledge base into a fault knowledge base, which includes multiple fault complete graphs, each of which is represented as a fault, and is visualized in the form of a fault complete graph, with different colors distinguishing the calculation type of each criterion node.

[0010] The complete fault graph includes one or more nodes. The node information contains the criterion itself and its corresponding calculation type, data encoding and fault information. The edge information contains the association relationship between the criteria. A complete fault graph is represented by a single criterion or multiple criterion nodes and the association relationships represented by the edges.

[0011] Each sub-criteria in the complete fault graph obtains the same attributes as the original criterion and is connected with edges, and the criteria are classified according to the calculation type. According to the fault definition, multiple criteria corresponding to the fault are connected with edges.

[0012] Step (2) sets corresponding prompt word templates according to different calculation types to automatically generate prompt words.

[0013] The automatic generation of prompt words is constructed from the problem of natural language synthesis. The entire prompt word template is segmented based on the function of text content, and natural language is synthesized by splicing different contents. The prompt word template consists of four parts:

[0014] The head is used to describe the task requirements and specify the output of the large model; The external knowledge part, knowledge, contains the preliminary knowledge needed to perform the task. Although the large model may already contain such knowledge, inputting it into the large model in advance through prompt words can guide it to complete the task faster and reduce the response time. The example section contains multiple examples of performing tasks. Synthesize some of the format and connect other parts and input data using natural language.

[0015] Each of the above parts and input data is represented as "{{tag name}}", such as "{{head}}", which is convenient for the program to retrieve and replace the content through strings. Then, the format part connects these tags with natural language. Finally, the program automatically fills in the content according to the tags in the format to complete the automatic generation of prompt words.

[0016] Step (3) sets a tag matching module. The judgment status of the input data includes normal false, abnormal true, and error error. The tag matching module synthesizes the data with the judgment status of abnormal true into an abnormal status set.

[0017] Step (4) connects the prompt word template to the front end of the large language model, and the tag matching module to the back end of the large language model. The prompt word template, the large language model and the tag matching module constitute the equipment state reasoning module. The multi-dimensional equipment state monitoring time series data is input into the equipment state reasoning module. The prompt word template generates a prompt word set at the current moment. The large language model calculates and infers the prompt word set to generate the state of the power plant power generation equipment at the current moment, which is expressed in the form of a tag language combined with a criterion state. The tag matching module synthesizes the abnormal state set at the current moment according to the criterion state, which is expressed as a set of multiple criterion nodes.

[0018] Step (5) inputs the abnormal state set at the current moment into the fault knowledge base, performs fault retrieval, and obtains the fault diagnosis result.

[0019] The judgment node in the abnormal state set is traversed through each fault complete graph in the fault knowledge base. If the judgment node is a node connected by an edge, the other node of the edge traverses the abnormal state set, otherwise the judgment node is deleted from the abnormal state set; if the other node of the edge is not any other judgment node in the abnormal state set, the judgment node is deleted from the abnormal state set, otherwise the judgment node is retained. After each judgment node in the abnormal state set performs the above operations, an updated abnormal state set is formed, and the judgment nodes in the updated abnormal state set are the fault diagnosis results.

[0020] Step (6) constructs a power plant equipment fault diagnosis visualization analysis system based on the equipment status reasoning module and the fault knowledge base.

[0021] The system supports users to analyze original multi-dimensional time series data and criterion knowledge, use large models for fault diagnosis, and perform interactive prompt word construction and optimization. The system interface includes the following parts:

[0022] (6.1) Multidimensional time series data line chart: To avoid visual clutter caused by drawing a large number of lines in one chart, draw multiple line charts based on the detection target and data unit, such as Figure 2 Part A. When multidimensional time series data enters the system, it is classified by its attributes and then plotted on different line graphs, making it easier to clearly observe the changing trends of multidimensional time series data and locate and compare data with the same units.

[0023] (6.2) Fault knowledge graph: The fault knowledge base is visualized in the form of a knowledge graph. Different colors of criterion nodes represent different calculation types, and detailed information of criterion nodes can be viewed by hovering, such as Figure 2 Part B. Users can clearly analyze the composition of the judgment criteria for a single fault through the knowledge graph.

[0024] (6.3) Fault diagnosis view: It is a flow view that allows users to view and analyze the abnormalities that occur at each moment in the device, such as Figure 2 Part C. The abnormal criteria that appear at each moment are represented by rectangles and stacked, and different colors are used to distinguish the calculation types.

[0025] When a fault occurs, the background of the rectangle will turn red to prompt the user. Users can view fault information by hovering, which is convenient for them to respond. In the upper right corner of the chart, a timeline is also provided for users to view historical data. When the timeline status is "Following", the chart will be updated at any time to display the latest results of fault diagnosis. Users change the state of the timeline to "Stop" by clicking or dragging, and then select the time range to update the chart to view historical fault results. In addition, by clicking to highlight the criterion rectangle, the data location in the Time Series and the corresponding node in the knowledge graph. By hovering the mouse over the rectangle, the user can check the detailed text content of the large model response. In addition, since the illusion problem of large models is inevitable, an interactive method is also provided to support users to revise the diagnosis results, and the diagnosis results can be modified by double-clicking the rectangle.

[0026] (6.4) Prompt word template building component: Provide multiple input windows corresponding to different segmented parts. After the user completes all parts, a complete prompt word is formed and displayed, such as Figure 2 Part D. Provides a prompt word template input component based on text content function segmentation to meet the task and design requirements of automatic prompt word generation.

[0027] (6.5) Prompt word evaluation component: Provides prompt word performance feedback to users to help them optimize the prompt word. The component consists of text preview, bidirectional bar chart, reply projection view, reply detail display, and comparison view.

[0028] Text preview is a preview of the evaluation data set and prompt words. Data sample preview allows users to enter the number of samples and view the data. Prompt word preview generates complete prompt words based on data and prompt word templates for users to view, making it easier for users to find prompt word errors in natural language text, such as Figure 2 Part E.

[0029] The two-way bar chart shows the statistical results of the consistency and accuracy of different versions of the prompt word template on the large model's answers. The gray bars represent the number of inaccuracies or inconsistencies, and are stacked on the corresponding bars. Users can filter the results by clicking on the gray bars, such as Figure 2 Part F.

[0030] Reply projection view, embed the reply of the large model, and form a scatter plot after dimensionality reduction through the T-sne method, such as Figure 2 Part G in the middle. Different versions of prompt words are distinguished by different colors. Users use projection to analyze the consistency of the large model's replies under a prompt word template, and the projection shows whether the points of the same color form a tight cluster. Users use projection to find outliers and inspire them to optimize the prompt words. By clicking a point on the projection view, the reply details display will filter out other replies.

[0031] The reply details display shows the specific reply content of the large model and distinguishes whether the answer is consistent or correct by background color, such as Figure 2 Middle H part.

[0032] Comparison view, through the radar chart to intuitively compare the performance of the prompt word template, such as Figure 2 Part I. First, in order to help users find the best prompt word template of the same type, the version comparison uses accuracy, consistency and running time as the axis to compare different versions of prompt words of the same type. The template comparison is evaluated on the complete dataset and uses different calculation logic types as the axis to compare the performance of different types of prompt word templates.

[0033] The effectiveness of this method is evaluated as follows: The different capabilities of the prompt word templates for fault detection are evaluated through accuracy, consistency rate, and running time.

[0034] The accuracy rate indicates the accuracy of the prompt word template for the fault detection task. , Indicates the number of correct output results. Indicates the total number of data.

[0035] The consistency rate indicates the stability of the prompt word template. , Indicates the number of consistent output results. Indicates the total number of data.

[0036] Runtime Description The response speed performance of the prompt word template, run time , Indicates the end time of the operation. Indicates the start time of the run.

[0037] GLM-4, Baichuan2-Turbo and Qwen-Turbo were used to evaluate the constructed prompt word templates. The evaluation results are as follows: Figure 3 As shown, the fault diagnosis method proposed in the present invention can obtain a high accuracy diagnosis result, with an accuracy score of 86% to 100%.

[0038] The practicality of this method was tested by actually operating it, and by updating the knowledge base, the changes in the fault diagnosis results were checked. Figure 4 As shown in the figure, when the data at time "00:17" entered the system, a fault occurred, and the fault information was "Coal mill foreign body alarm". When "A mill current is greater than 40A" was changed to "A mill current is greater than 50A", the fault at "00:17" no longer occurred. After deleting the criterion of "A mill current is greater than 50A", the fault at "00:17" appeared again. When a new criterion "The steam temperature above the outlet drops by more than 5°C in 1 minute" and its information were added. At time "00:33", a new fault "A coal feeder abnormal alarm" was reported.

Claims

1. A method for fault diagnosis of power generation equipment based on a large language model, characterized in that: Step (1) converting the existing power plant expert knowledge base into a fault knowledge base, the fault knowledge base includes multiple fault complete graphs, each fault complete graph is represented as a fault, and is visualized in the form of a fault complete graph, with different colors distinguishing the calculation type of each criterion node; Step (2) setting corresponding prompt word templates according to different computing types to automatically generate prompt words; The prompt word template consists of four parts: The head is used to describe the task requirements and specify the output of the large model; The external knowledge part, knowledge, contains the preliminary knowledge needed to perform the task; The example section contains multiple examples of performing tasks. Synthesize some of the formats and connect other parts and input data using natural language; Step (3) setting a tag matching module, wherein the judgment status of the input data includes normal false, abnormal true, and error error, and the tag matching module synthesizes the data with the judgment status of abnormal true into an abnormal status set; Step (4) connecting the prompt word template to the front end of the large language model, connecting the tag matching module to the back end of the large language model, and the prompt word template, the large language model and the tag matching module constitute a device state inference module; The multi-dimensional equipment status monitoring time series data is input into the equipment status reasoning module, the prompt word template generates the prompt word set at the current moment, the large language model calculates and infers the prompt word set, generates the state of the power plant power generation equipment at the current moment, and expresses it in the form of a markup language combined with a criterion state; the tag matching module synthesizes the abnormal state set at the current moment according to the criterion state, and expresses it as a set of multiple criterion nodes; Step (5) inputting the abnormal state set at the current moment into the fault knowledge base, performing fault retrieval, and obtaining the fault diagnosis result; Step (6) constructs a power plant equipment fault diagnosis visualization analysis system based on the equipment state reasoning module and the fault knowledge base; the system interface includes the following parts: (6.1) Line graph of multi-dimensional time series data: When multi-dimensional time series data enters the system, it is classified by its attributes and plotted on different line graphs; (6.2) Fault knowledge graph: The fault knowledge base is visualized in the form of a knowledge graph. Different colors of criterion nodes represent different calculation types, and detailed information of criterion nodes can be viewed by hovering. (6.3) Fault diagnosis view: This is a flow view that allows users to view and analyze the abnormalities that occur at each moment in the device. The abnormality criteria that occur at each moment are represented by rectangles and stacked, and different colors are used to distinguish the calculation types. (6.4) Prompt word template building component: provides multiple input windows corresponding to different segmented parts. After the user completes all parts, a complete prompt word is formed and displayed; (6.5) Prompt word evaluation component: Provides prompt word performance feedback to users to help them optimize the prompt word. The component consists of text preview, bidirectional bar chart, reply projection view, reply detail display, and comparison view.

2. The method for diagnosing faults of power generation equipment based on a large language model according to claim 1, characterized in that: The complete fault graph includes one or more nodes, the node information includes the criterion itself and its corresponding calculation type, data encoding and fault information, the edge information includes the association relationship between the criteria, and a complete fault graph is represented by a single criterion or a plurality of criterion nodes and the association relationship represented by the edge; Each sub-criteria in the complete fault graph obtains the same attributes as the original criterion and is connected with edges. The criteria are classified according to the calculation type; according to the fault definition, multiple criteria corresponding to the fault are connected with edges.

3. The method for diagnosing faults of power generation equipment based on a large language model according to claim 2, characterized in that: Step (5) is specifically as follows: traverse the criterion node in the abnormal state set through each complete fault graph in the fault knowledge base; if the criterion node is a node connected by an edge, then the other node of the edge traverses the abnormal state set, otherwise the criterion node is deleted from the abnormal state set; if the other node of the edge is not any other criterion node in the abnormal state set, then the criterion node is deleted from the abnormal state set, otherwise the criterion node is retained; after each criterion node in the abnormal state set performs the above operations, an updated abnormal state set is formed, and the criterion node in the updated abnormal state set is the fault diagnosis result.

4. The method for diagnosing faults of power generation equipment based on a large language model according to claim 1, characterized in that: The text preview is a preview of the evaluation data set and the prompt words; The two-way bar chart shows the statistical results of consistency and accuracy of different versions of the prompt word templates in the large model answer on both sides; The response projection view embeds the response of the large model and forms a scatter plot after dimensionality reduction through the T-sne method, with different colors distinguishing different versions of the prompt words; The response details display shows the specific response content of the large model and distinguishes whether the answer is consistent or correct by background color; The comparison view provides an intuitive comparison of the performances of the prompt word templates through a radar chart.

Images