A fault recording intelligent analysis method and system based on a large language model

Abstract

The application discloses a fault recording wave intelligent analysis method and system based on a large language model, and relates to the technical field of power system automation and artificial intelligence. The method comprises the following steps: obtaining a fault recording wave file identifier to be analyzed and an analysis intention description; performing data retrieval according to the analysis intention description to obtain fault case features; based on the fault case features, constructing a dynamic execution rule corresponding to the fault recording wave file identifier through a large language model; encapsulating a preset analysis calculation function through an MCP protocol to obtain an analysis tool set suitable for the large language model; arranging the analysis tool set through the dynamic execution rule to obtain a tool execution chain; and based on the fault recording wave file identifier, performing analysis and calculation on a fault recording wave file through the tool execution chain to obtain fault feature parameters and an analysis result of a visualized chart. The application improves the flexibility of a fault recording wave analysis software and reduces the secondary development cost.

Description

Technical Field

[0001] This invention belongs to the field of power system automation and artificial intelligence technology, specifically relating to a fault recording intelligent analysis method and system based on a large language model. Background Technology

[0002] Fault waveform analysis is a crucial method for assessing equipment behavior and identifying potential hazards after a power system fault. Currently, this field mainly relies on specialized analysis software or manual interpretation of waveforms.

[0003] Existing technologies suffer from the following significant drawbacks: Rigid interaction: Existing software typically provides fixed menus and buttons, unable to understand the natural language commands of maintenance personnel. Users must manually click multiple menus to complete specific analyses. High secondary development costs: For different types of protection, the software's analysis logic is usually hard-coded. Adding new analysis functions or adapting to new protection devices requires professional programmers to modify and recompile the code, resulting in long development cycles. Poor adaptability: Files generated by fault recorders from different manufacturers and models have vastly different channel naming rules. Traditional software often relies on fixed channel index configurations, lacking automatic identification and mapping capabilities, leading to extensive manual intervention for channel configuration before analysis. Lack of experience reuse mechanisms: Existing systems are mostly isolated analysis tools, making it difficult to automatically link historical fault handling experience to the current analysis process. In recent years, although machine learning has been used for specific fault classification tasks, these methods are usually limited to single criteria, failing to cover the full picture of fault recording analysis, and still requiring programmed logic combination. Summary of the Invention

[0004] The purpose of this invention is to provide a fault recording intelligent analysis method and system based on a large language model. By performing channel fuzzy matching through the large language model, intelligent matching of the mapping relationship of fault filtering files is realized.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0006] On the one hand, this invention provides an intelligent analysis method for fault recording based on a large language model, comprising:

[0007] Obtain analysis instructions described in natural language by the user, the analysis instructions including the identifier of the fault waveform file to be analyzed and a description of the analysis intent;

[0008] A vector database is constructed using the acquired historical fault data, and fault case features are obtained by analyzing intent descriptions based on the vector database.

[0009] Based on the characteristics of the aforementioned fault cases, dynamic execution rules corresponding to the fault recording file identifiers are constructed using a large language model.

[0010] The pre-defined parsing and calculation functions are encapsulated using the MCP protocol to obtain an analysis toolkit adapted to large language models; the pre-defined parsing and calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions.

[0011] The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain;

[0012] Based on the fault recording file identifier, the fault recording file is parsed and calculated using the tool execution chain to obtain the analysis results of fault characteristic parameters and visualization charts.

[0013] Optionally, the analysis instructions may be in the form of unstructured text or voice commands.

[0014] Optionally, the step of constructing a vector database using acquired historical fault data and retrieving fault case features based on the described analytical intent includes:

[0015] Based on the acquired historical fault data, the data is vectorized using a pre-trained BERT model to obtain fault vector data.

[0016] A vector database is constructed based on the aforementioned fault vector data;

[0017] Based on the described analytical intent, the vector database is retrieved using an approximate nearest neighbor search algorithm to obtain fault case features.

[0018] Optionally, the analysis toolset includes multiple function wrapper tools;

[0019] The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain including:

[0020] Based on the dynamic execution rules, a list of computation step objects is obtained through rule parsing;

[0021] Based on the list of computational step objects, the analysis toolset is arranged using a topological sorting algorithm to obtain the execution order of each function encapsulation tool within the analysis toolset;

[0022] Based on the execution order, parameter binding is performed on each function encapsulation tool to obtain the tool execution chain.

[0023] Optionally, the function encapsulation tool includes input parameters and output parameters;

[0024] Based on the execution order, parameter binding is performed on each function encapsulation tool to obtain the tool execution chain, including:

[0025] Based on the execution order, the output parameters of the function wrapper tool are bound to the input parameters of the next function wrapper tool.

[0026] Optionally, based on the fault recording file identifier, the fault recording file is parsed and calculated using the tool execution chain, including:

[0027] The channel description information of the fault recording file is confirmed based on the fault recording file identifier; the channel description information includes the channel name, unit, and conversion factor;

[0028] Based on the channel description information of the fault recording file, the mapping relationship of the fault recording file is obtained by performing fuzzy matching of the channels through a large language model.

[0029] The tool execution chain calls the fault recording file for parsing and calculation based on the mapping relationship of the fault filtering file, and obtains the analysis results of fault characteristic parameters and visualization charts.

[0030] Optionally, based on the channel description information of the fault recording file, fuzzy matching of the channels is performed using a large language model to obtain the mapping relationship of the fault recording file, including:

[0031] The channel description information is semantically vectorized using a large language model to obtain the original description vector.

[0032] Similarity is calculated based on the original description vector and the preset standard term vector to obtain the similarity value of the original description vector;

[0033] If the similarity value of the original description vector is greater than the similarity threshold, a mapping relationship is constructed between the fault recording files corresponding to the channel description information.

[0034] Secondly, the present invention provides a fault recording intelligent analysis system based on a large language model, comprising:

[0035] The front-end user interaction module is used to obtain the analysis instructions described by the user in natural language. The analysis instructions include the identifier of the fault recording file to be analyzed and the description of the analysis intent.

[0036] The historical fault recording and storage module is used to construct a vector database using the acquired historical fault data and to retrieve fault case features based on the described analysis intent.

[0037] The backend intelligent analysis module is used to construct dynamic execution rules corresponding to the fault waveform file identifiers based on the fault case characteristics using a large language model; it encapsulates the preset parsing and calculation functions using the MCP protocol to obtain an analysis toolset adapted to the large language model; the preset parsing and calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions; it arranges the analysis toolset using the dynamic execution rules to obtain a tool execution chain; based on the fault waveform file identifiers, it uses the tool execution chain to parse and calculate the fault waveform files to obtain analysis results of fault feature parameters and visualization charts.

[0038] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0039] This invention constructs an analysis toolset using the MCP protocol, facilitating modification of the analysis logic and reducing secondary development costs. It utilizes a vector database to store historical fault data, enabling the reuse of historical fault handling experience and improving the efficiency of intelligent fault waveform analysis. Furthermore, it employs a large language model to construct dynamic execution rules corresponding to fault waveform file identifiers, enabling rapid adaptation to different types of protection systems. This invention effectively alleviates the technical problems of traditional fault waveform analysis software, such as poor flexibility, high secondary development costs, and inability to understand unstructured instructions.

[0040] The present invention uses unstructured text or voice commands for analysis, which improves the flexibility of user interaction; it uses a large language model for fuzzy matching of channels, which improves the flexibility of channel indexing and reduces the need for manual intervention in channel configuration for analysis. Attached Figure Description

[0041] Figure 1 The diagram shown is a schematic representation of the interactive process of this invention. Detailed Implementation

[0042] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0043] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0044] Example 1

[0045] This embodiment introduces an intelligent fault recording analysis method based on a large language model, specifically including:

[0046] The system obtains analysis instructions described in natural language by the user. The analysis instructions include the identifier of the fault waveform file to be analyzed and a description of the analysis intent. In this embodiment, the analysis instructions use unstructured text or voice commands, but can also be replaced with structured text as the analysis instructions input by the user.

[0047] A vector database is constructed using acquired historical fault data. Based on this vector database, fault case features are obtained through data retrieval by analyzing intent descriptions. The retrieval can employ a hybrid retrieval strategy combining metadata filtering and vector similarity retrieval. Metadata includes fault type and voltage level. This hybrid retrieval strategy improves retrieval accuracy.

[0048] Based on the characteristics of the aforementioned fault cases, dynamic execution rules corresponding to the fault waveform file identifiers are constructed using a large language model. This embodiment can dynamically generate dynamic execution rules corresponding to the fault waveform file identifiers using a large language model, and these execution rules can also be preset. In this embodiment, the large language model is preferably a pre-trained language model based on the Transformer architecture.

[0049] The pre-defined parsing and computation functions are encapsulated using the Model Context Protocol (MCP) to obtain an analysis toolset adapted to large language models. Specifically, based on the MCP protocol, an adapter layer is built to encapsulate Python functions into atomic tools that can be called by the large language model. In this embodiment, the pre-defined parsing and computation functions mainly include data parsing functions, mathematical operation functions, and logical judgment functions. Specifically, middleware for tool registration, parameter validation, and response serialization is implemented by parsing the MCP protocol specification. Exemplarily, typical function encapsulation tools include, but are not limited to, the following categories:

[0050] The data parsing tool parse_comtrade loads the COMTRADE parsing dynamic link library written in C by calling the ctypes library, achieving efficient binary data parsing and returning a structured dictionary containing sampling rate, channel configuration, and sampled data.

[0051] The mathematical calculation tool `calculate_impedance` uses the Discrete Fourier Transform algorithm to calculate voltage and current phasors. The impedance calculation formula is: Z = U / I, where U is the voltage phasor, I is the current phasor, and Z is the measured impedance.

[0052] The logic criterion tool check_distance_zone implements the distance protection impedance circle criterion, and its mathematical expression is: |Z - Zset| ≤ Rset, where Z is the measured impedance, Zset is the set impedance, and Rset is the radius of the impedance circle.

[0053] The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain; the tool execution chain is used to determine the execution order of function encapsulation tools within the toolset. The analysis toolset can be orchestrated using preset execution rules.

[0054] Based on the fault waveform file identifier, the tool execution chain is used to parse and calculate the fault waveform file, obtaining analysis results of fault characteristic parameters and visualization charts. The tool execution chain can construct the required physical channels for input data for parsing and calculation based on the fault waveform file identifier, and can automatically identify and intelligently map them, reducing the manual requirements for channel matching and ensuring the accuracy of input data when the tool execution chain performs parsing and calculation.

[0055] Example 2

[0056] Based on the same inventive concept as Embodiment 1, this embodiment introduces an intelligent fault recording analysis method based on a large language model, including:

[0057] The system obtains analysis instructions described in the user's natural language. The analysis instructions include the identifier of the fault recording file to be analyzed and a description of the analysis intent. The analysis instructions can be in the form of unstructured text or voice commands.

[0058] A vector database is constructed using the acquired historical fault data, and fault case features are obtained by analyzing intent descriptions based on the vector database. Based on the acquired historical fault data, a pre-trained BERT model is used to perform data vectorization operations to obtain fault vector data. The BERT model is a publicly available model in the prior art and can be used after pre-training with historical fault data.

[0059] A vector database is constructed based on the aforementioned fault vector data;

[0060] Based on the described analytical intent, the vector database is retrieved using an approximate nearest neighbor search algorithm to obtain fault case features. The approximate nearest neighbor search algorithm can achieve millisecond-level retrieval in datasets of millions of records.

[0061] Based on the fault case characteristics, dynamic execution rules corresponding to the fault recording file identifiers are constructed through a large language model; the fault case characteristics are input into the large language model as contextual information, which enhances the analysis and decision-making capabilities, and the large language model dynamically generates dynamic execution rules through a rule engine.

[0062] The preset parsing calculation functions are encapsulated using the MCP protocol to obtain an analysis toolkit adapted to large language models; the preset parsing calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions; the analysis toolkit includes multiple function encapsulation tools;

[0063] The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain including:

[0064] Based on the dynamic execution rules, a list of computation step objects is obtained through rule parsing;

[0065] Based on the list of computational step objects, the analysis toolset is arranged using a topological sorting algorithm to obtain the execution order of each function encapsulation tool within the analysis toolset;

[0066] Based on the execution order, parameter binding is performed on each function encapsulation tool to obtain the tool execution chain. The function encapsulation tool includes input parameters and output parameters. The specific method for parameter binding is to bind the output of the preceding function encapsulation tool as the input parameter of the subsequent step according to the execution order.

[0067] The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain;

[0068] The dynamic execution rules generated by the rule engine also include exception handling rules, and are equipped with timeout mechanisms and error retry strategies for the tool execution chain to ensure system stability.

[0069] Based on the fault waveform file identifier, the tool execution chain is used to parse and calculate the fault waveform file to obtain the analysis results of fault characteristic parameters and visualization charts. Specifically, the specific input data required for the tool execution chain to parse and calculate cannot be directly obtained from the fault waveform file; corresponding guidance is needed. In most existing technologies, a mapping channel for the fault waveform file is manually established to allow the tool execution chain to obtain the corresponding input data. However, in this embodiment, a large language model is used to automatically establish a mapping channel for the fault waveform file based on the needs of the tool execution chain and the fault waveform file identifier. Specifically, this includes: confirming the channel description information of the fault waveform file based on the fault waveform file identifier; the channel description information includes the channel name, unit, and conversion coefficient.

[0070] Based on the channel description information of the fault recording file, the mapping relationship of the fault recording file is obtained by performing fuzzy matching of the channels through a large language model.

[0071] The tool execution chain calls the fault recording file for parsing and calculation based on the mapping relationship of the fault filtering file, and obtains the analysis results of fault characteristic parameters and visualization charts.

[0072] The channel fuzzy matching through a large language model specifically includes: using a large language model to perform semantic vectorization processing on the channel description information to obtain the original description vector;

[0073] The similarity is calculated based on the original description vector and the preset standard term vector. The algorithm used is the cosine similarity algorithm, and the similarity value of the original description vector is obtained as similarity = (A·B) / (|A||B|), where A is the original description vector and B is the standard term vector.

[0074] If the similarity value of the original description vector is greater than the similarity threshold, in this embodiment the similarity threshold is set to 0.85, and the mapping relationship of the fault recording file corresponding to the channel description information is constructed.

[0075] Example 3

[0076] Based on the same inventive concept as other embodiments, this embodiment introduces a fault recording intelligent analysis system based on a large language model, such as... Figure 1 As shown, it specifically includes:

[0077] The front-end user interaction module is used to obtain the analysis instructions described by the user in natural language. The analysis instructions include the identifier of the fault recording file to be analyzed and the description of the analysis intent.

[0078] The historical fault waveform storage module is used to construct a vector database using the acquired historical fault data and to retrieve fault case features based on the described analysis intent. The original historical fault data is a Comtrade waveform file, which needs to be vectorized before it can be stored in the domain vector database.

[0079] The backend intelligent analysis module is used to construct dynamic execution rules corresponding to the fault waveform file identifiers based on the fault case characteristics using a large language model; it encapsulates the preset parsing and calculation functions using the MCP protocol to obtain an analysis toolset adapted to the large language model; the preset parsing and calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions; it arranges the analysis toolset using the dynamic execution rules to obtain a tool execution chain; based on the fault waveform file identifiers, it uses the tool execution chain to parse and calculate the fault waveform files to obtain analysis results of fault feature parameters and visualization charts.

[0080] The front-end user interaction module generates an analysis report based on the analysis results of fault characteristic parameters and visualization charts. The analysis report is presented on the user interaction interface in a mixed text and graphics format, and the mixed text and graphics format is semantically related to the analysis intent description. Channel mapping confirmation information needs to be generated and displayed, which includes the identified standard logical channel name and its corresponding original channel description text.

[0081] The front-end user interaction module provides users with a natural language input interface through a web interface or mobile application, while the back-end intelligent analysis module is deployed on a cloud server or edge computing device.

[0082] When a user enters a natural language command in the front-end user interaction module: Please analyze the EXAMPLE_FAULT_001.dat file to see if the three distance protection stages have malfunctioned, and find the A-phase current path.

[0083] Intent recognition: The large language model recognizes the entity "distance protection three segments", the action "analysis erroneous action", and the file "EXAMPLE_FAULT_001.dat", and generates formula call instructions in the corresponding format.

[0084] Channel matching: The system scans the file, calculates the similarity based on the channel description and semantic matching, and automatically maps it to the standard logical channel "Protection Current Phase A". A mapping confirmation prompt is generated in the front-end user interaction module.

[0085] Rule generation: The execution chain of the large language model dynamic generation tool is: data parsing tool → mathematical operation tool → logical judgment tool. The tool execution chain is used to parse and calculate to obtain the analysis results of fault characteristic parameters and visualization charts.

[0086] Output Results: The analysis results of fault characteristic parameters and visualization charts are used to generate an analysis report containing an impedance trajectory diagram through the front-end user interaction module. The diagram marks the Z3 circle characteristic and highlights the "channel mapping confirmation process". The analysis report also includes auxiliary decision-making suggestion text generated based on the characteristics of the historical fault cases.

[0087] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A fault recording intelligent analysis method based on a large language model, characterized in that, include: Obtain analysis instructions described in natural language by the user, the analysis instructions including the identifier of the fault waveform file to be analyzed and a description of the analysis intent; A vector database is constructed using the acquired historical fault data, and fault case features are obtained by analyzing intent descriptions based on the vector database. Based on the characteristics of the aforementioned fault cases, dynamic execution rules corresponding to the fault recording file identifiers are constructed using a large language model. The preset parsing calculation functions are encapsulated using the MCP protocol to obtain an analysis toolset adapted to large language models; The preset analytical calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions; The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain; Based on the fault recording file identifier, the fault recording file is parsed and calculated using the tool execution chain to obtain the analysis results of fault characteristic parameters and visualization charts.

2. The intelligent fault recording analysis method based on a large language model according to claim 1, characterized in that, The analysis commands are provided in the form of unstructured text or voice commands.

3. The intelligent fault recording analysis method based on a large language model according to claim 1, characterized in that, The process of constructing a vector database using acquired historical fault data and retrieving fault case features based on the described analytical intent includes: Based on the acquired historical fault data, the data is vectorized using a pre-trained BERT model to obtain fault vector data. A vector database is constructed based on the aforementioned fault vector data; Based on the described analytical intent, the vector database is retrieved using an approximate nearest neighbor search algorithm to obtain fault case features.

4. The intelligent fault recording analysis method based on a large language model according to claim 1, characterized in that, The analysis toolset includes multiple function wrapper tools; The analysis toolset is orchestrated using dynamic execution rules to obtain a tool execution chain including: Based on the dynamic execution rules, a list of computation step objects is obtained through rule parsing; Based on the list of computational step objects, the analysis toolset is arranged using a topological sorting algorithm to obtain the execution order of each function encapsulation tool within the analysis toolset; Based on the execution order, parameter binding is performed on each function encapsulation tool to obtain the tool execution chain.

5. The intelligent fault recording analysis method based on a large language model according to claim 4, characterized in that, The function encapsulation tool includes input parameters and output parameters; Based on the execution order, parameter binding is performed on each function encapsulation tool to obtain the tool execution chain, including: Based on the execution order, the output parameters of the function wrapper tool are bound to the input parameters of the next function wrapper tool.

6. The intelligent fault recording analysis method based on a large language model according to claim 1, characterized in that, Based on the fault recording file identifier, the fault recording file is parsed and calculated using the tool execution chain, including: The channel description information of the fault recording file is confirmed based on the fault recording file identifier; the channel description information includes the channel name, unit, and conversion factor; Based on the channel description information of the fault recording file, the mapping relationship of the fault recording file is obtained by performing fuzzy matching of the channels through a large language model. The tool execution chain calls the fault recording file for parsing and calculation based on the mapping relationship of the fault filtering file, and obtains the analysis results of fault characteristic parameters and visualization charts.

7. The intelligent fault recording analysis method based on a large language model according to claim 6, characterized in that, Based on the channel description information of the fault recording files, fuzzy matching of the channels is performed using a large language model to obtain the mapping relationship of the fault recording files, including: The channel description information is semantically vectorized using a large language model to obtain the original description vector. Similarity is calculated based on the original description vector and the preset standard term vector to obtain the similarity value of the original description vector; If the similarity value of the original description vector is greater than the similarity threshold, a mapping relationship is constructed between the fault recording files corresponding to the channel description information.

8. A fault recording intelligent analysis system based on a large language model, characterized in that, include: The front-end user interaction module is used to obtain the analysis instructions described by the user in natural language. The analysis instructions include the identifier of the fault recording file to be analyzed and the description of the analysis intent. The historical fault recording and storage module is used to construct a vector database using the acquired historical fault data and to retrieve fault case features based on the described analysis intent. The backend intelligent analysis module is used to construct dynamic execution rules corresponding to the fault recording file identifiers based on the characteristics of the fault cases through a large language model; to encapsulate the preset parsing calculation functions through the MCP protocol to obtain an analysis toolset adapted to the large language model; wherein the preset parsing calculation functions include data parsing functions, mathematical operation functions, and logical criterion functions; and to arrange the analysis toolset using dynamic execution rules to obtain a tool execution chain. Based on the fault recording file identifier, the fault recording file is parsed and calculated using the tool execution chain to obtain the analysis results of fault characteristic parameters and visualization charts.

Images