Power generation equipment fault diagnosis method based on large language model

Abstract

The application discloses a power generation equipment fault diagnosis method based on a large language model. The method first converts an expert knowledge base into a fault knowledge base composed of criterion nodes and edges, and completes fault retrieval through the knowledge base; uses input based on text function segmentation to complete natural language text synthesis, constructs a prompt word template based on the calculation types of different criteria, and promotes the large model to complete the calculation and reasoning tasks of the criteria; inputs multi-dimensional equipment state monitoring time series data, generates a prompt word set under the current time in combination with the prompt word template, inputs the prompt word set into the large language model for calculation and reasoning and generates the current state set of the equipment; uses the knowledge base to retrieve the state set to obtain a fault diagnosis result; and further designs a visual system to collect and visualize the output result of the large model and the fault diagnosis result, and provides an input and an interactive interface for the prompt word template construction. The fault diagnosis completed by the method has good practicability and expansibility.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of information technology, and particularly relates to a power generation equipment fault diagnosis method based on a large language model. BACKGROUND

[0002] With the development of information technology and the popularity of sensors, power plants can obtain and collect massive 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 diagnosis analysis, researchers have proposed traditional data-based methods, deep learning-based methods and model-based methods. However, the traditional data-based method requires domain expertise and a lot of effort to support the feature extraction process. The deep learning-based method greatly depends on the expertise of signal processing and deep learning, and also brings high cost of training models.

[0003] Therefore, in actual production applications, most intelligent power plants use model-based automated fault detection methods. This method fixes expert knowledge into computer executable software code and inputs multi-dimensional time series data to obtain a judgment result. Expert knowledge refers to a series of rules and knowledge summarized by experienced workers in the process of equipment maintenance and inspection. The criteria in the criterion knowledge formed by the expert knowledge are programmed into computer recognizable codes by artificial programming and embedded into the fault diagnosis system. Although this method can efficiently execute the preset specific diagnosis task through the code, there are still some defects. First, the model-based method highly depends on professional knowledge (W.J. Li, H.L, S. Gu, T. Chen, Process fault diagnosis with model- and knowledge-based approaches: Advances and opportunities, Control Engineering Practice. 105(11) (2020) 104637), and it takes a lot of manpower to accurately correspond and fix the expert knowledge into the code. Second, the updating of the knowledge base increases the difficulty of code maintenance. When adjusting or adding a certain point of the expert criterion, such as the change of the threshold judgment range, the change of the coefficient and the relationship combination mode of the fault, the corresponding code needs to be found and modified. This not only consumes a lot of manpower, but also causes the updating or even shutdown of the fault diagnosis system, which may cause a large amount of loss. In addition, even the same equipment in different power plants, the expert knowledge base will be different under different production conditions, which cannot realize code migration, and the above-mentioned code conversion and maintenance problems will also occur. These problems limit the expansibility and practicality of power equipment fault diagnosis. SUMMARY

[0004] The application aims to provide a power plant fault diagnosis method based on a large language model.

[0005] The method comprises the following steps:

[0006] Step (1) converts an existing power plant expert knowledge base into a fault knowledge base, and the fault knowledge base comprises a plurality of fault complete graphs, each fault complete graph representing a fault, and the fault complete graph is visualized, and different colors are used to distinguish the calculation types of each criterion node;

[0007] Step (2) sets a corresponding prompt word template according to different calculation types, which is used to automatically generate a prompt word; the prompt word template comprises four parts:

[0008] A head part is used to explain the requirements of the task and specify the output of the large model;

[0009] An external knowledge part contains the preliminary knowledge required to perform the task;

[0010] An example part contains a plurality of examples for performing the task;

[0011] A synthesis part is used to connect the other parts and the input data in natural language;

[0012] Step (3) sets a label matching module, the criterion state of the input data comprises normal false, abnormal true and error error, and the label matching module synthesizes the data with an abnormal criterion state true into an abnormal state set;

[0013] Step (4) inputs the prompt word template into the front end of the large language model, and inputs the label matching module into the rear end of the large language model, wherein the prompt word template, the large language model and the label matching module constitute a device state reasoning module; multi-dimensional device state monitoring time series data is input into the device state reasoning module, the prompt word template generates a prompt word set at the current time, the large language model performs calculation and reasoning on the prompt word set, and generates the state of the power plant generating equipment at the current time, which is represented in the form of a label language combined with a criterion state; the label matching module synthesizes the abnormal state set at the current time according to the criterion state, which is represented in the form of a plurality of criterion node sets;

[0014] Step (5) inputs the abnormal state set at the current time into the fault knowledge base to perform fault retrieval, and obtains a fault diagnosis result;

[0015] Step (6) constructs a power plant equipment fault diagnosis visual analysis system based on the device state reasoning module and the fault knowledge base; the system interface comprises the following parts:

[0016] (6.1) Multi-dimensional time series data line chart: when multi-dimensional time series data enters the system, it is classified by its attributes and drawn on different line charts;

[0017] (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 the criterion nodes can be viewed by hovering;

[0018] (6.3) Fault diagnosis view: a flowing view for users to view and analyze the abnormalities of the device at each moment, the abnormal criterion at each moment is represented by a rectangle and stacked, and different colors are used to distinguish the calculation types;

[0019] (6.4) Prompt word template construction component: provides multiple input windows corresponding to different segmented parts, and after the user completes all parts, a complete prompt word is formed and displayed;

[0020] (6.5) Prompt word evaluation component: feedback the performance of the prompt word to the user to help optimize the prompt word, the component consists of text preview, bidirectional bar chart, reply projection view, reply detail display, and comparison view.

[0021] Compared with the prior art, the beneficial effects of the present application are: the present application integrates large language models, knowledge bases, visualization, and human-computer interaction technologies, manages and maintains expert knowledge in the form of a knowledge base, which is conducive to the updating and expansion of fault models; the end-to-end large language model undertakes the calculation and reasoning tasks of the criterion, reducing the need for signal processing or machine learning professional knowledge and data sets; the use of prompt word templates ensures the efficiency and accuracy of large language model output, adapts to most fault detection tasks, and has strong portability between power plants with structured or semi-structured expert knowledge; the original time series data, knowledge base, and collected large model replies are visualized and provided with human-computer interaction, which is convenient for viewing fault details and making error-tolerant responses; the use of interactive input templates facilitates the construction of prompt word templates; the use of a richly designed prompt word evaluation template facilitates the optimization of the prompt word template. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a workflow schematic diagram of the present application;

[0023] Figure 2 is a visualization system schematic diagram of the present application;

[0024] Figure 3 is a quantitative evaluation result schematic diagram of the present application;

[0025] Figure 4 is an application example schematic diagram of the present application. DETAILED DESCRIPTION

[0026] The power plant fault diagnosis method based on a large language model will be further described below with reference to the accompanying drawings.

[0027] As shown in the specific process includes the following steps: Figure 1

[0028] Step (1) converts the existing power plant expert knowledge base into a fault knowledge base, and the fault knowledge base includes multiple fault complete graphs, each fault complete graph represents a fault, and the fault complete graph is visualized, and different colors distinguish the calculation types of each criterion node.

[0029] The fault complete graph includes one or more nodes, and the node information includes the criterion itself and its corresponding calculation type, data encoding and fault information, and the edge information includes the association relationship between criteria. A fault complete graph is represented by a single criterion or the association relationship represented by multiple criterion nodes and edges.

[0030] Each sub-criterion in the fault complete graph obtains the same attributes as the original criterion, and is connected by an edge. The criteria are classified according to the calculation type. According to the fault definition, multiple criteria corresponding to the fault are connected by an edge.

[0031] Step (2) sets the corresponding prompt word template according to different calculation types, which is used to automatically generate the prompt word.

[0032] The automatic generation of the prompt word is constructed by the natural language synthesis problem. The entire prompt word template is segmented based on the text content function, and the natural language is synthesized by splicing different contents. The prompt word template includes four parts:

[0033] The head part is used to explain the task requirements and specify the output of the large model;

[0034] The external knowledge part knowledge contains the preliminary knowledge required to perform the task. Although the large model may already contain such knowledge, by pre-inputting the large model through the prompt word, it can guide it to complete the task faster and reduce the response time;

[0035] The example part contains multiple examples of performing the task;

[0036] The synthesis part format connects the other parts and the input data in natural language.

[0037] ​Each of the above parts and input data is represented as a "tag name", such as "head", and the like, which facilitates the program to search and replace the content by string. Then, the format part connects these tags in natural language. Finally, the program automatically fills in the content according to the tags in the format, to complete the automatic generation of the prompt word.

[0038] Step (3) sets a label matching module, the criterion state of the input data includes normal false, abnormal true, and error error, and the label matching module synthesizes the data with the criterion state of abnormal true into an abnormal state set.

[0039] Step (4) connects the prompt word template to the front end of the large language model, and connects the label matching module to the rear end of the large language model, and the prompt word template, the large language model, and the label matching module constitute a device state reasoning module. The multi-dimensional device state monitoring time series data is input into the device state reasoning module, the prompt word template generates a prompt word set at the current time, the large language model calculates and reasons the prompt word set to generate the state of the power plant generating equipment at the current time, which is expressed in the form of a label language combined with a criterion state. The label matching module synthesizes the abnormal state set at the current time according to the criterion state, which is expressed as a set of multiple criterion nodes.

[0040] Step (5) inputs the abnormal state set at the current time into the fault knowledge base to perform fault retrieval, and obtains a fault diagnosis result.

[0041] The criterion nodes in the abnormal state set are traversed in each fault complete graph in the fault knowledge base, if the criterion node is one node connected by an edge, then the other node of the edge is traversed in the abnormal state set, otherwise the criterion node is deleted in the abnormal state set; if the other node of the edge is not any criterion node in the abnormal state set, then the criterion node is deleted in the abnormal state set, otherwise the criterion node is retained. After each criterion node in the abnormal state set performs the above operation, an updated abnormal state set is formed, and the criterion nodes in the updated abnormal state set are the fault diagnosis result.

[0042] Step (6) constructs a power plant equipment fault diagnosis visual analysis system based on the device state reasoning module and the fault knowledge base.

[0043] The system supports users to analyze original multi-dimensional time series data and criterion knowledge, uses a large model to perform fault diagnosis, and performs interactive prompt word construction and optimization. The system interface includes the following parts:

[0044] (6.1) Multi-dimensional time series data line chart: to avoid the visual clutter problem caused by drawing a large number of lines in one chart, to detect targets and data units, and to draw multiple line charts, such as Figure 2Part A. When multi-dimensional time series data enters the system, it is classified by its attributes and then plotted on different line charts, making it easy to observe the changing trend of multi-dimensional time series data, locate and compare data with the same unit.

[0045] (6.2) Fault knowledge graph: visualize the fault knowledge base in the form of a knowledge graph, different colors of criterion nodes represent different types of calculations, and detailed information of criterion nodes can be viewed by hovering over them, such as Figure 2 Part B. Users can clearly analyze the criterion composition of a single fault through the knowledge graph.

[0046] (6.3) Fault diagnosis view: a dynamic view for users to view and analyze the abnormalities that occur at each moment of the device, such as Figure 2 Part C. The abnormal criterion at each moment is represented by a rectangle and stacked, and different colors are used to distinguish the calculation types.

[0047] When a fault occurs, the background of the rectangle turns red to prompt the user. Users can view fault information by hovering over it, making it easy for them to respond. In the upper right corner of the chart, there is also a timeline to help users view historical data. When the timeline status is "Following", the chart will be updated in real time to display the latest results of fault diagnosis. Users can change the status of the timeline to "Stop" by pressing the button or dragging it, 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 position in the time series and the corresponding node in the knowledge graph. By hovering the mouse over the rectangle, users can check the detailed text content of the large model reply. In addition, due to the inevitable illusion problem of large models, an interactive method is provided to support users to revise the diagnosis results, and the diagnosis results can be modified by double-clicking the rectangle.

[0048] (6.4) Prompt word template construction component: provides multiple input windows corresponding to different segmentation parts, and after the user completes all parts, a complete prompt word is formed and displayed, such as Figure 2 Part D. Provide a prompt word template input component based on text content function segmentation to meet the task of automatic generation of prompt words and design requirements.

[0049] (6.5) Prompt word evaluation component: feedback the performance of the prompt word to the user 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.

[0050] Text preview, which is a preview of the evaluation dataset and the prompt words. Data sample preview allows users to input the number of samples and view the data. Prompt word preview generates complete prompt words through data and prompt word templates for users to view, which facilitates them to find errors in prompt words from natural language text, such as Figure 2 Part E in FIG. 1.

[0051] Bidirectional bar chart, two sides respectively show the statistical results of the consistency and accuracy of different versions of prompt word templates on the large model answers. The gray bar represents the number of inaccurate or inconsistent, and is stacked on the corresponding bar. Users filter the results by clicking the gray bar, such as Figure 2 Part F in FIG. 1.

[0052] Reply projection view, embedding the reply of the large model, and forming a scatter plot after dimension reduction by T-sne method, such as Figure 2 Part G in FIG. 1. Different colors distinguish different versions of prompt words. Users analyze the consistency of the large model's reply under a prompt word template through projection, and whether the points of the same color form a tight cluster in the projection. Users find outliers through projection to inspire them to optimize the prompt word. By clicking the point on the projection view, the reply detail display filters out other replies.

[0053] Reply detail display, showing the specific reply content of the large model, and distinguishing whether the answer is consistent or correct through background color, such as Figure 2 Part H in FIG. 1.

[0054] Comparison view, intuitively comparing the performance of prompt word templates through radar chart, such as Figure 2 Part I in FIG. 1. First, in order to help users find the best prompt word template under the same type, version comparison takes accuracy, consistency and running time as the three axes to compare different versions of prompt words under the same type. Template comparison evaluates on the complete dataset and uses different calculation logic types as the axis to compare the performance of different types of prompt word templates.

[0055] The effectiveness evaluation of the method is as follows:

[0056] The accuracy, consistency and running time are used to evaluate the different abilities of prompt word templates for fault detection.

[0057] Accuracy indicates the precision of the prompt word template for the fault detection task, and the accuracy , represents the number of correct results, represents the total number of data.

[0058] Consistency indicates the stability of the prompt word template, and the consistency , The number of results indicating output consistency, The total number of data.

[0059] The running time indicates the response speed performance using the prompt word template, the running time , The running end time is indicated. The running start time is indicated.

[0060] The constructed prompt word template is evaluated using GLM-4, Baichuan2-Turbo and Qwen-Turbo, and the evaluation results are shown in Figure 3 The fault diagnosis method provided by the present application has high accuracy, and the accuracy score is 86% to 100%.

[0061] The practicability of the method is actually operated, and the change of the fault diagnosis result is viewed through the updating of the knowledge base. As shown in Figure 4 When the data at time "00:17" enters the system, a fault occurs, and the fault information is "foreign matter alarm of coal mill". When "A mill current is greater than 40A" is modified to "A mill current is greater than 50A", the fault at "00:17" no longer occurs. When the criterion "A mill current is greater than 50A" is deleted, the fault at "00:17" occurs again. When a new criterion "steam temperature above the outlet drops by more than 5℃ for 1 minute" and its information are added. At time "00:33", a new fault "A coal feeder abnormal alarm" is reported.

Claims

1. A power plant fault diagnosis method based on a large language model, characterized by: Step (1) converting an existing power plant expert knowledge base into a fault knowledge base, the fault knowledge base including multiple fault complete graphs, each fault complete graph representing a fault, and the fault complete graphs being visualized in the form of fault complete graphs, with different colors distinguishing the calculation types of each criterion node; Step (2) setting corresponding prompt word templates according to different calculation types for automatically generating prompt words; The prompt word template includes four parts: The head part is used to explain the task requirements and specify the output of the large model; The external knowledge part contains the preliminary knowledge required to perform the task; The example part contains multiple examples of performing the task; The synthesis part connects the other parts and the input data in natural language; Step (3) setting a label matching module, the criterion state of the input data including normal false, abnormal true, and error error, and the label matching module synthesizing the abnormal state set of the data with the criterion state of abnormal true; Step (4) connecting the prompt word template to the front end of the large language model, and connecting the label matching module to the back end of the large language model, the prompt word template, the large language model, and the label matching module forming a device state reasoning module; Inputting multi-dimensional device state monitoring time series data into the device state reasoning module, the prompt word template generating a prompt word set at the current time, the large language model performing calculation reasoning on the prompt word set to generate the state of the power plant generating equipment at the current time, represented in the form of a label language combined with the criterion state; the label matching module synthesizes the abnormal state set at the current time according to the criterion state, represented as a set of multiple criterion nodes; Step (5) inputting the abnormal state set at the current time into the fault knowledge base to perform fault retrieval and obtain the fault diagnosis result; Step (6) constructing a power plant device fault diagnosis visual analysis system based on the device state reasoning module and the fault knowledge base; the system interface includes the following parts: (6.1) Multi-dimensional time series data line chart: when multi-dimensional time series data enters the system, it is classified by its attributes and plotted on different line charts; (6.2) Fault knowledge graph: the fault knowledge base is visualized using a knowledge graph, with different colors of criterion nodes representing different calculation types, and the detailed information of the criterion nodes can be viewed by hovering; (6.3) Fault diagnosis view: a flowing view for users to view and analyze the abnormalities of the device at each time, with each time's abnormal criterion represented by a rectangle and stacked, and different colors used to distinguish the calculation types; (6.4) Prompt word template construction component: provides multiple input windows corresponding to different segmentation parts, and the user forms a complete prompt word after completing all parts and displays it; (6.5) Prompt word evaluation component: feedback the performance of the prompt word to the user 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 power plant fault diagnosis method based on a large language model according to claim 1, wherein: The fault complete graph includes one or more nodes, node information contains criterion itself and its corresponding calculation type, data coding and fault information, edge information contains the association relationship between criteria, a fault complete graph is represented by a single criterion or the association relationship represented by multiple criterion nodes and edges; Each sub-criterion in the fault complete graph obtains the same attribute as the original criterion and is connected by an edge. The criteria are classified according to the calculation type. According to the fault definition, the multiple criteria corresponding to the fault are connected by an edge.

3. The power plant fault diagnosis method based on a large language model according to claim 2, characterized by, Step (5) is specifically: traversing each fault complete graph in the fault knowledge base for the criterion nodes in the abnormal state set, if the criterion node is one node connected by an edge, traversing the abnormal state set for the other node of the edge, otherwise deleting the criterion node in the abnormal state set; if the other node of the edge is not any other criterion node in the abnormal state set, deleting the criterion node in the abnormal state set, otherwise retaining the criterion node; after each criterion node in the abnormal state set performs the above operation, an updated abnormal state set is formed, and the criterion nodes in the updated abnormal state set are the fault diagnosis results.

4. The power plant fault diagnosis method based on a large language model according to claim 1, wherein: The text preview is a preview of the evaluation data set and the prompt word; The bidirectional bar chart displays statistical results of the consistency and accuracy of different versions of the prompt word template on the large model on the two sides; The reply projection view embeds the reply of the large model, performs dimension reduction through the T-sne method, and forms a scatter plot, different colors distinguish different versions of the prompt word; The reply detail display displays the specific reply content of the large model, and distinguishes whether the answer is consistent or correct through the background color; The comparison view intuitively compares the performance of the prompt word template through the radar chart.

Images