A stereo garage fault analysis system and method based on a large language model

Abstract

The application discloses a three-dimensional garage fault analysis system and method based on a large language model (LLM), wherein a private large language model is generated through low-rank adaptation (LoRA) of the large language model, and a device maintenance knowledge base is constructed through knowledge graph generation software; during system operation, device internet-of-things signals are processed through semantic processing according to attributes and features, and are input into local and global question sub-modules for processing, so that the whole process from device operation state description to automatic generation of fault root cause analysis results and fault solutions is realized. The application improves the efficiency and accuracy of fault diagnosis, reduces maintenance costs, and provides comprehensive fault root cause analysis results and fault solutions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault diagnosis, in particular to a stereo garage fault analysis system and method based on a large language model. BACKGROUND

[0002] In modern industry, complex equipment systems such as stereo garages are prone to faults during operation. These fault points are numerous and complex in combination, making it difficult to quickly locate the root cause of the fault, which in turn affects the timely repair of the fault. Traditional fault diagnosis models mainly have the following defects:

[0003] (1) The model based on reasoning or logical analysis is complex and costly to build, and its application in complex systems is limited;

[0004] (2) Traditional machine learning models require a large amount of labeled data and have poor interpretability;

[0005] (3) The existing LLM directly applied has the problem of "hallucination" caused by insufficient domain knowledge;

[0006] (4) Complex faults require comprehensive analysis of multi-source heterogeneous data. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application proposes a stereo garage fault analysis system and method based on a large language model.

[0008] The specific technical solutions are as follows:

[0009] A stereo garage fault analysis system based on a large language model, comprising: a private LLM module, a device maintenance knowledge base module, an Internet of Things and edge computing module, a retrieval and reasoning module;

[0010] The private LLM module is used to fine-tune the pre-trained large language model to generate a private large language model suitable for intelligent stereo garage fault analysis, which is used to output fault root cause analysis results and fault solutions;

[0011] The device maintenance knowledge base module is used to construct a knowledge graph-based device maintenance knowledge base through a knowledge graph generation software, which contains knowledge in the field of intelligent stereo garage;

[0012] The Internet of Things and edge computing module is used to collect daily operation data and fault information of the intelligent stereo garage, and upload them to the fault analysis cloud platform, and convert them into semantic information, i.e. problems;

[0013] The retrieval and reasoning module comprises a question classification submodule, a local questioning submodule and a global questioning submodule; the question classification submodule is used for classifying questions according to their attributes and characteristics, inputting questions that can give conclusions through retrieval summary into the local questioning submodule for processing, and inputting questions that need to evaluate influence and analyze root causes into the global questioning submodule for processing;

[0014] The local questioning submodule is used for performing near neighbor search on text segmentation, entities, relationships and graph communities in the equipment maintenance knowledge base according to the question, and screening out the content with the highest similarity and relevance scores, and then performing language organization and output through a private large language model;

[0015] The global questioning submodule performs retrieval based on a Map-Reduce architecture, first performs near neighbor search on graph communities in the equipment maintenance knowledge base according to the question, and takes the graph communities with a similarity reaching a set threshold as relevant communities; inputs the community summaries of the relevant communities into a private large language model to generate intermediate results; and then groups the intermediate results according to the similarity and relevance scores of the questions, and integrates the fault root cause analysis and fault solution scheme after summarizing through the private large language model.

[0016] Further, the fine-tuning of the pre-trained large language model in the private LLM module comprises the following steps:

[0017] (1.1) Data preprocessing: converting the maintenance manual and fault handling method of the intelligent stereo garage into input-output question and answer pairs, and performing data cleaning, removing redundant information and standardization processing;

[0018] (1.2) Using the data after data preprocessing, fine-tuning the pre-trained large language model through the LoRA method to obtain a private large language model.

[0019] Further, the knowledge graph generation software selects GraphRAG, and the construction of the equipment maintenance knowledge base comprises the following steps:

[0020] (2.1) Text segmentation: after unstructured processing of the structured historical maintenance records, the records are segmented into multiple text blocks according to a pre-defined block size;

[0021] (2.2) Using a private large language model to identify entities from each text block and extract relationships between entities;

[0022] (2.3) Community clustering: regarding an entity and its corresponding attribute information as a node, clustering the nodes into graph communities based on the relationships between the entities obtained in (2.2), and using a private large language model to generate community summaries;

[0023] (2.4) Constructing a knowledge graph including nodes, relationships, and graph communities, and storing the knowledge graph as an equipment maintenance knowledge base; the equipment maintenance knowledge base is plugged into the private large language model to obtain a fault analysis model.

[0024] Furthermore, in the step (2.2), entities are identified from the text block by performing named entity recognition, and the relationship between the entities is extracted by using a relationship extraction algorithm; and in the step (2.3), node clustering is performed using the Leiden algorithm.

[0025] Furthermore, the PLC controller of the intelligent stereo garage is the target device, the IoT box is connected to the serial port or Ethernet port of the target device, and the IoT box supports the industrial communication protocol to obtain the operating data and fault information of the target device;

[0026] Perform edge computing on the IoT box: For non-critical operating data, set up an abnormal data trigger mechanism. Only when a specific error code or abnormal situation is detected, the IoT box will publish non-critical operating data within a fixed period before and after the fault to the fault analysis cloud platform. For critical operating data, the IoT box will publish it to the fault analysis cloud platform in real time.

[0027] In the fault analysis cloud platform, data cleaning and format conversion are performed to convert the data into semantic information that can be processed by a large language model.

[0028] Furthermore, the IoT box is a data terminal unit IoT box, which is connected to the RS485 interface of the target device and supports Modbus or DLT protocol.

[0029] A method for analyzing a fault in a stereoscopic parking garage based on a large language model is implemented based on the aforementioned stereoscopic parking garage fault analysis system based on a large language model, and includes the following steps:

[0030] S1: In the private LLM module, based on the maintenance manual and standard fault handling methods of the intelligent parking garage, the pre-trained large language model is fine-tuned using the LoRa method to obtain a private large language model suitable for intelligent parking garage fault analysis;

[0031] S2: In the equipment maintenance knowledge base module, the GraphRAG method is used to build an equipment maintenance knowledge base based on the knowledge graph. The domain knowledge of the intelligent stereo garage is used as a database and plugged into the private large language model to form a fault analysis model.

[0032] S3: Through the Internet of Things and edge computing modules, IoT data collection and processing of daily operation and fault information of the intelligent parking garage are carried out, and semantic information is output to the retrieval and reasoning modules;

[0033] S4: The semantic information output by the IoT and edge computing modules is taken as questions, and the questions are classified into simple retrieval questions or complex analysis questions according to their attributes through the question classification submodule; the simple retrieval questions are questions that can give conclusions through retrieval and summary, and are processed by the local questioning submodule; the complex analysis questions require impact assessment and root cause analysis to give conclusions, and are processed by the global questioning submodule.

[0034] Furthermore, the S3 specifically includes the following steps:

[0035] The IoT box is connected to the serial port or Ethernet port of the target device and obtains the operating data and fault information of the target device through the industrial communication protocol;

[0036] The IoT box transmits the collected data to the fault analysis cloud platform, where it records and processes the data, including data cleaning and format conversion; and converts the processed data into semantic information that can be processed by a large language model.

[0037] During the above transmission process, edge computing is performed on the IoT box side: for non-critical operating data, an abnormal data trigger mechanism is set. Only when a specific error code or abnormal situation is detected, the IoT box will publish the non-critical operating data within a fixed time before and after the fault to the fault analysis cloud platform; for critical operating data, the IoT box will publish it to the fault analysis cloud platform in real time.

[0038] Furthermore, the characteristics of the simple search questions include: ① querying a single, clear fault code handling method; ② querying standard maintenance procedures or parts replacement methods; ③ querying equipment parameters or configuration information; ④ the problem description is simple and direct, without the need to associate multiple factors;

[0039] The characteristics of complex analysis problems include: ① containing multiple related fault codes; ② involving mutual influence or cascading failures between devices; ③ repeated failures; ④ the need to analyze the impact of environmental factors on the failure; ⑤ the problem description contains words that require in-depth analysis, and the words that require in-depth analysis include: non-standard, root cause, and correlation.

[0040] The beneficial effects of the present invention are:

[0041] The stereoscopic garage fault analysis system constructed in the application is based on large language model and graph retrieval enhancement generation, through combining a private large language model, a device maintenance knowledge base and Internet of Things data processing technology, realizes the whole process from device running state description to automatic generation of fault identification, root cause analysis result and fault solution, thereby improving the efficiency and accuracy of fault diagnosis, especially for complex faults that are difficult to solve through traditional rules or logical reasoning; and solves the "illusion" problem caused by insufficient knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the stereoscopic garage fault analysis system architecture based on large language model and graph retrieval enhancement generation in the application.

[0043] Figure 2 is the Internet of Things and edge computing architecture and model connection schematic diagram in the application.

[0044] Figure 3 is the local question flow chart of retrieval and reasoning of the fault analysis model in the application.

[0045] Figure 4 is the global question flow chart of retrieval and reasoning of the fault analysis model in the application. DETAILED DESCRIPTION

[0046] The application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings and preferred embodiments, and the application will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0047] As shown in Figure 1 and Figure 2 , a stereoscopic garage fault analysis system based on a large language model includes a private LLM module, a device maintenance knowledge base module, an Internet of Things and edge computing module, and a retrieval and reasoning module.

[0048] The private LLM module is used to fine-tune the pre-trained large language model to enhance the model performance, and generate a private large language model (hereinafter referred to as private LLM) suitable for intelligent stereoscopic garage fault analysis, so that it can output indefinite length descriptive fault root cause analysis and fault solution; the pre-trained large language model selects an existing open source large language model, including Qwen2.5-32B-Instruct, Gemma2-27B-instruct, Llama3-70B-Instruct, etc. The specific fine-tuning process includes:

[0049] (1.1) Data preprocessing: Convert the maintenance manual and fault handling methods of the intelligent stereo garage into input-output question and answer pairs, which include "input" and "output" question and answer pairs in JSON format. First, clean the question and answer pairs, then remove irrelevant characters, punctuation marks, stop words, and other redundant information, then standardize the text, and convert the data into JSON format suitable for inputting into the model.

[0050] As shown in Table 1 below, some question and answer pair examples are given, with the input being the question and the output being the indefinite-length descriptive fault root cause analysis (i.e., the possible cause of the problem) and the fault solution (including the location that needs to be checked).

[0051] Table 1 Question and answer pair examples

[0052]

[0053] (1.2) Based on the pre-processed data described above, the LoRA method is used to fine-tune the pre-trained large language model to obtain a private LLM. LoRA is a technique for adjusting the parameters of a pre-trained large language model to adapt to new tasks, which introduces an additional low-rank matrix to adapt to the intelligent stereo garage fault analysis, rather than directly modifying the original weight matrix; this method can reduce the demand for computing resources while maintaining the performance of the model. In this embodiment, the training parameters are set as: number of epochs epoch = 5, batch size batch_size = 8, learning rate learning_rate = 0.0003.

[0054] Device maintenance knowledge base module: used to construct a knowledge graph-based device maintenance knowledge base through a knowledge graph generation software (GraphRAG is used in this embodiment), so that the knowledge in the field of intelligent stereo garage is attached to the private LLM as an external database, which can effectively reduce the large model illusion and cover as much information knowledge as possible related to fault handling in the daily operation of the device. The device maintenance knowledge base and the private LLM form a fault analysis model.

[0055] The specific process of constructing a knowledge graph-based device maintenance knowledge base through GraphRAG includes:

[0056] (2.1) Text chunking: non-structured processing of structured historical maintenance records, and then blocking the non-structured maintenance records according to a pre-defined block size to obtain multiple text blocks; in this embodiment, the pre-defined block size is 500 characters.

[0057] (2.2) Entity Relationship Extraction: Use the private LLM to identify entities from each text block and extract the relationships between each entity. In this embodiment, entity recognition is performed using Named Entity Recognition (NER), and a relationship extraction algorithm is used to extract the relationships between each entity.

[0058] (2.3) Community Clustering: Record an entity and its corresponding attribute information as a node, based on the extracted entity relationships (i.e., the relationship between nodes), use Leiden algorithm to cluster nodes into graph communities, and generate community summaries by private LLM. To generate community summaries, first extract graph community features, including topological features (i.e., graph community size, internal connection density, central node, inter-graph community connectivity, etc.), semantic features (i.e., extract high-frequency entities within the graph community, and entity type distribution, to quickly determine the theme tendency of the community), relationship features (i.e., the relationship between entities within the community); then perform feature structural representation, converting graph community features into text templates or JSON format; then generate community summaries, combining feature information with natural language, and filtering redundant information to summarize the content of each community.

[0059] (2.4) Knowledge Storage: Build a knowledge graph containing 120,000 nodes and 350,000 relationships, stored in LanceDB vector database as a private LLM device maintenance knowledge base.

[0060] IoT and Edge Computing Module: Used to collect daily operation data and fault information of the intelligent stereo garage, and convert these operation data into semantic information that can be answered by a large language model. Specifically, it includes the following main contents:

[0061] (3.1) Hardware Connection: Use a Data Transfer Unit (DTU) or Remote Terminal Unit (RTU) IoT box to connect the serial port or Ethernet port of the target device (i.e., the PLC controller of the intelligent stereo garage), and transmit the data collected by the target device to the fault analysis cloud platform. In this embodiment, the IoT box is a DTU IoT box, which is connected to the RS485 interface of the target device.

[0062] (3.2) Protocol Conversion: The IoT box supports multiple industrial communication protocol parsing, allowing real-time operation data and fault information of the target device to be obtained based on physical connection. This embodiment supports Modbus-TCP or DLT645-2007 protocol.

[0063] (3.3) Edge computing strategy: edge computing is performed at the end of the Internet of Things box to save data traffic from the target device to the server end. Specifically, first, according to the preset running parameters, it is judged whether the running data is critical. For critical running data, the Internet of Things box publishes to the MQTT server (MQTT Broker) in the fault analysis cloud platform through the interface; and for non-critical running data, an abnormal data triggering mechanism is set, and only when a specific error code or abnormal situation is detected (in this embodiment, the error codes defined by a certain stereo garage equipment production company are used as the standard, and when the E80 "broken rope detection action" and E99 "lift platform switch error" fault codes are detected), the Internet of Things box uploads 30 seconds of data before and after the fault to the fault analysis cloud platform through the interface.

[0064] (3.4) In the fault analysis cloud platform, the data from the MQTT Broker is cleaned and converted in format, and the data is converted into semantic information that can be processed by a large language model, that is, a semantic fault description.

[0065] Retrieval and reasoning module: the semantic fault description output by the Internet of Things and edge computing module is taken as a question, classified according to the question attributes, and intelligently selected for local query (Local query) or global query (Global query) reasoning mode, combined with the fault analysis model for retrieval and reasoning. The retrieval and reasoning module is divided into question classification submodule, local query submodule and global query submodule according to function. The functions realized by each module are as follows.

[0066] Question classification submodule: since the local query and global query have a large difference in resource consumption of the system, the question classification submodule is designed in this embodiment, which uses a private LLM to judge the question attributes and features, and classifies the questions into different submodules (i.e. local query submodule or global query submodule) for processing to shorten the system response time. Specifically, the private LLM classifies the questions that can give conclusions through retrieval summary as simple retrieval class questions (retrieval_query) according to the question attributes and features (obtained by querying the private LLM once through the prompt word template), and classifies them into the local query submodule for processing; while classifying the questions that need to evaluate the impact and analyze the root cause to give conclusions as complex analysis class questions (complex_query), and classifying them into the global query submodule for processing. The following is an example of a prompt word template for a private LLM to judge question classification:

[0067] "System prompt: you are a question classification expert of the intelligent stereo garage fault analysis system, responsible for judging the complexity of the fault question to decide which query method to use.

[0068] Task: Please determine whether the following fault information is a simple retrieval problem that can be solved by simple retrieval or a complex analysis problem that requires in-depth analysis of the root cause.

[0069] Criteria for judgment: The characteristics of simple retrieval problems include: ① Query for a single and clear method of handling fault codes; ② Query for standard maintenance procedures or component replacement methods; ③ Query for device parameters or configuration information; ④ Problem description is simple and direct, without the need to associate multiple factors. The characteristics of complex analysis problems include: ① Contains multiple associated fault codes; ② Involves mutual influence or cascading failures between devices; ③ Faults repeatedly occur; ④ Need to analyze the impact of environmental factors (such as temperature, humidity) on the fault; ⑤ Problem description contains words such as 'non-standard', 'root cause', 'association' that require in-depth analysis.

[0070] Problem classification: {Problem text}. Please only answer'retrieval_query' or 'complex_query', without any explanation or other content."

[0071] Local question sub-module: As shown in Figure 3 , after receiving the input semantic fault description information, the nearest neighbor search is performed on the original text cut (i.e. the text cut generated in the first step of the device maintenance knowledge base generation process), entity, relationship and graph community in the device maintenance knowledge base, and the results of the four groups of nearest neighbor search are ranked (Ranking) according to similarity and relevance scores. According to the context window size of the private LLM, the top-ranked original text, entity, relationship and community are selected (Filtering) from top to bottom, and finally the filtered content is provided to the private LLM for language organization and output.

[0072] Global question sub-module: As shown in Figure 4 , after receiving the input semantic fault description information, the Map-Reduce architecture (Map-Shuffle-Reduce) is used for retrieval, which includes three stages of mapping (Map), intermediate process (Shuffle) and reduction (Reduce), as follows.

[0073] ① Map stage: According to the input semantic fault description information, the nearest neighbor search is performed on the graph community in the device maintenance knowledge base, i.e. the similarity between the semantic fault description information and the community summary of each graph community is calculated. If the similarity is greater than or equal to the set threshold, the graph community is selected as the relevant community, and a prompt word is constructed for each relevant community. With the community summary of the relevant community as input, the private LLM is used to generate intermediate results, which are generally partial answers related to the query or further refinements of the content of the relevant community.

[0074] This stage generates a prompt word template similar to the following:

[0075] "System prompt: You are a stereo garage fault analysis expert, please answer the questions related to the fault based on the provided community information.

[0076] Task: Analyze the content related to '{user question summary}' in the following community information, extract key information and generate partial answers.

[0077] Community information: {community summary}, {community key entities and relationships}.

[0078] Requirements: ① Only answer based on the provided community information; ② Extract the causes and solutions related to the fault '{fault code or phenomenon}'; ③ If the community information contains multiple related cases, summarize the commonalities and differences; ④ The output format should include 'possible causes' and'suggested measures'; ⑤ The answer length should be controlled within 200 words.

[0079] User question: {original question}."

[0080] ②Shuffle stage: Sort and group the intermediate results of each related community according to their similarity and relevance scores with the question. The similarity and relevance scores of the intermediate results in the same group are similar. This process can make answers from different graph communities with similar themes converge, preparing for the next step of reduction.

[0081] ③Reduce stage: The grouped intermediate results obtained in the Shuffle stage are summarized and executed by a private LLM, integrating the partial answers of each related community into a coherent and comprehensive final answer, i.e., fault root cause analysis and fault solution. The final answer is obtained by synthesizing all related community answers, which ensures that even complex global problems can be answered comprehensively and contextually.

[0082] Local questioning mainly uses semantic token neighborhood search to retrieve partial device maintenance knowledge base, which is fast and provides private LLM content related to the literal meaning of the question. Global questioning traverses all communities and calls private LLM multiple times to induce intermediate results, and finally integrates the most relevant content. Although the processing speed is slower, it can provide more comprehensive answers, suitable for integrating indirect information to analyze complex faults.

[0083] Based on the above stereo garage fault analysis system based on large language models, the embodiment also proposes a stereo garage fault analysis method based on large language models, including the following steps:

[0084] S1: In the private LLM module, based on the maintenance manual and standard fault handling method of the intelligent stereo garage, the LoRA method is used to fine-tune the pre-trained large language model to obtain a private LLM suitable for fault analysis of the intelligent stereo garage.

[0085] S2: In the equipment maintenance knowledge base module, the GraphRAG method is used to construct an equipment maintenance knowledge base based on the knowledge graph. The domain knowledge of the intelligent stereo garage is used as a database plug-in outside the private LLM to form a fault analysis model, which effectively reduces the large model illusion.

[0086] The equipment maintenance knowledge base is constructed using the GraphRAG method, which is implemented through the following sub-steps:

[0087] S2.1: Structured historical maintenance records are unstructured. These unstructured maintenance records are divided into predefined chunks, each containing a certain number of text units, such as sentences or paragraphs. This chunking method breaks down large amounts of text data into smaller, more manageable and processable parts.

[0088] S2.2: For each text block, named entity recognition is performed using a private LLM to identify specific entities in the text, such as names of people, places, organizations, etc. Subsequently, the semantic relationships between these entities are analyzed using a relation extraction algorithm to extract the relationship information between the entities.

[0089] S2.3: Use the Leiden algorithm to identify and cluster entities and relationships extracted from the text blocks, forming multiple graph communities. Then, use the private LLM to summarize the content of each community and generate a community summary for fast retrieval. These entities, relationships, and graph communities form an equipment maintenance knowledge base, which is plugged into the private LLM to form a fault analysis model.

[0090] S3: Through the IoT and edge computing modules, IoT data collection and processing of daily operation and fault information of the intelligent parking garage are carried out, and semantic fault descriptions are output to the retrieval and reasoning modules. This step specifically includes:

[0091] IoT Box Connection: Use a DTU or RTU IoT Box to connect to the serial or Ethernet port of the target device to achieve a physical connection between the target device and the IoT Box, providing a foundation for data collection. In this embodiment, the IoT Box uses a DTU type FBOX.

[0092] Protocol Data Acquisition: Based on physical connections, industrial communication protocols such as Modbus and DLT are used to obtain real-time operation and error signals from PLCs, inverters, HMIs, and other devices within the intelligent parking garage. These protocols were selected based on their widespread application and reliability in industrial automation.

[0093] Data transmission: Collected data is transmitted to the fault analysis cloud platform via the IoT gateway and ultimately reaches the MQTT broker. The MQTT protocol, due to its lightweight and low bandwidth consumption, is well-suited for IoT environments, especially where data transmission is limited.

[0094] Data Recording and Processing: The fault analysis cloud platform's server backend receives data from the MQTT broker and performs secondary processing. This secondary processing includes data cleaning and format conversion. The processed data is converted into semantic information that can be understood and answered by the large model.

[0095] Edge Computing: To effectively manage data traffic and reduce the burden on the server, edge computing is implemented on the IoT Box (implemented during the "Publish" step). Edge computing allows data processing to occur close to the source, reducing the amount of data transmitted to the fault analysis cloud platform. Specifically, for non-critical data, a conditionally triggered collection strategy is employed. Only when an error code or anomaly is detected will the IoT Box collect and send the current and previous data to the fault analysis cloud platform.

[0096] S4: The semantic fault descriptions output by the IoT and edge computing modules are treated as questions. The question classification submodule categorizes these questions into simple retrieval questions or complex analysis questions based on their attributes. Simple retrieval questions can be concluded simply through retrieval summarization and are handled by the local questioning submodule. Complex analysis questions require impact assessment and root cause analysis to reach a conclusion and are handled by the global questioning submodule. Finally, a final answer is output, including the root cause analysis and the solution.

[0097] Specifically, the problem classification submodule receives semantic fault descriptions output by the IoT and edge computing modules and performs classification judgments by calling a private LLM. A specialized prompt template is designed, including the instruction "Please determine whether the following fault information requires impact assessment and root cause analysis," and additional judgment criteria such as "includes multiple fault codes," "involves mutual impact between devices," or "recurring faults," among other complex situation characteristics. After receiving this prompt and fault information, the LLM outputs a classification result: "Simple retrieval question (retrieval_query)" or "Complex analysis question (complex_query)." Based on the classification results, simple retrieval questions are directed to the local questioning submodule, while complex analysis questions are directed to the global questioning submodule. This ensures the rational allocation of resources, rapid response to simple questions, and in-depth analysis of complex questions.

[0098] In this embodiment, 3825 fault records of intelligent stereo garage from 2018 to 2023 are collected as a sample data set, and the sample data set is analyzed by using the application. The average response time of the local question-based fault analysis is 1.2 seconds, and the accuracy rate is 89.7%. The average response time of the global question-based fault analysis is 8.5 seconds, and the accuracy rate is 93.4%. Compared with the existing fault analysis system based on the expert system, the efficiency of the fault root cause analysis of the application is improved by 47%, and the generation speed of the corresponding solution is improved by 62%. In summary, the application can realize fast and accurate fault analysis and solution generation, improve the maintenance efficiency of the intelligent stereo garage, and reduce the maintenance cost.

[0099] Those skilled in the art can understand that the above description is only preferred examples of the application and is not used to limit the application. Although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or replace some technical features with equivalent ones. Any modification, equivalent replacement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A large language model-based three-dimensional parking garage fault analysis system, characterized in that: include: Private LLM module, equipment maintenance knowledge base module, IoT and edge computing module, retrieval and reasoning module; The private LLM module is used to fine-tune the pre-trained large language model to generate a private large language model suitable for intelligent stereo garage fault analysis, and is used to output fault root cause analysis results and fault solutions; The equipment maintenance knowledge base module is used to construct an equipment maintenance knowledge base based on a knowledge graph through knowledge graph generation software, which contains knowledge in the field of intelligent stereo garages; The IoT and edge computing module is used to collect daily operation data and fault information of the intelligent stereo garage, upload it to the fault analysis cloud platform, and convert it into semantic information, i.e., problems; The retrieval and reasoning module includes a question classification submodule, a local questioning submodule, and a global questioning submodule; the question classification submodule is used to classify questions according to their attributes and characteristics using a private LLM, inputting questions that can be concluded through retrieval and summary into the local questioning submodule for processing, and inputting questions that require impact assessment and root cause analysis to give conclusions into the global questioning submodule for processing; The local question submodule is used to perform neighbor searches on text segments, entities, relationships, and graph communities in the equipment maintenance knowledge base based on the questions, and select the content with the highest similarity and relevance scores, which are then organized and output through a private large language model; The global question submodule performs retrieval based on the Map-Reduce architecture. First, it performs a neighbor search for graph communities in the equipment maintenance knowledge base based on the question, and selects graph communities whose similarity reaches a set threshold as relevant communities. The community summaries of the relevant communities are input into the private large language model to generate intermediate results. Intermediate results are then grouped based on their similarity and relevance to the problem, aggregated using a private large language model, and integrated into fault root cause analysis and solutions. The knowledge graph generation software is GraphRAG, and building the equipment maintenance knowledge base includes the following steps: (2.1) Text Segmentation: After unstructuring the structured historical maintenance records, they are segmented into blocks according to pre-defined block sizes to obtain multiple text blocks; (2.2) Use a private large language model to identify entities from each text block and extract the relationships between entities; (2.3) Community clustering: An entity and its corresponding attribute information is recorded as a node. Based on the relationship between entities obtained in (2.2), the nodes are clustered into graph communities, and a private large language model is used to generate community summaries. (2.4) Constructing a knowledge graph including nodes, relationships, and graph communities, and storing the knowledge graph as an equipment maintenance knowledge base; the equipment maintenance knowledge base is plugged into the private large language model to obtain a fault analysis model.

2. The large language model-based three-dimensional parking garage fault analysis system according to claim 1 is characterized in that: In the private LLM module, fine-tuning the pre-trained large language model includes the following steps: (1.1) Data preprocessing: The maintenance manual and troubleshooting methods for the intelligent parking garage are converted into input-output question-answer pairs, and data cleaning, redundant information removal, and standardization are performed. (1.2) Using the pre-processed data, the pre-trained large language model is fine-tuned using the LoRA method to obtain a private large language model.

3. The large language model-based three-dimensional parking garage fault analysis system according to claim 1 is characterized in that: In the step (2.2), entities are identified from the text block by performing named entity recognition, and the relationship between the entities is extracted by using a relationship extraction algorithm; in the step (2.3), node clustering is performed using the Leiden algorithm.

4. The large language model-based three-dimensional parking garage fault analysis system according to claim 1 is characterized in that: The PLC controller of the intelligent stereo garage is the target device, and the IoT box is connected to the serial port or Ethernet port of the target device. The IoT box supports industrial communication protocols to obtain the operating data and fault information of the target device; Perform edge computing on the IoT box: For non-critical operating data, set up an abnormal data trigger mechanism. Only when a specific error code or abnormal situation is detected, the IoT box will publish non-critical operating data within a fixed time before and after the fault to the fault analysis cloud platform; For critical operating data, the IoT box publishes it to the fault analysis cloud platform in real time; In the fault analysis cloud platform, data cleaning and format conversion are performed to convert the data into semantic information that can be processed by a large language model.

5. The large language model-based three-dimensional parking garage fault analysis system according to claim 4 is characterized in that: The IoT box uses a data terminal unit IoT box, which is connected to the RS485 interface of the target device and supports Modbus or DLT protocol.

6. A method for analyzing a fault in a stereoscopic parking garage based on a large language model, implemented based on the system for analyzing a fault in a stereoscopic parking garage based on a large language model according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1: In the private LLM module, based on the maintenance manual and standard fault handling methods of the intelligent parking garage, the pre-trained large language model is fine-tuned using the LoRa method to obtain a private large language model suitable for intelligent parking garage fault analysis; S2: In the equipment maintenance knowledge base module, the GraphRAG method is used to build an equipment maintenance knowledge base based on the knowledge graph. The domain knowledge of the intelligent stereo garage is used as a database and plugged into the private large language model to form a fault analysis model. S3: Through the Internet of Things and edge computing modules, IoT data collection and processing of daily operation and fault information of the intelligent parking garage are carried out, and semantic information is output to the retrieval and reasoning modules; S4: The semantic information output by the IoT and edge computing modules is taken as questions, and the questions are classified into simple retrieval questions or complex analysis questions according to their attributes through the question classification submodule; the simple retrieval questions are questions that can give conclusions through retrieval and summary, and are processed by the local questioning submodule; the complex analysis questions require impact assessment and root cause analysis to give conclusions, and are processed by the global questioning submodule.

7. The method for analyzing faults in a stereoscopic parking garage based on a large language model according to claim 6, characterized in that: The S3 specifically includes the following steps: The IoT box is connected to the serial port or Ethernet port of the target device and obtains the operating data and fault information of the target device through the industrial communication protocol; The IoT box transmits the collected data to the fault analysis cloud platform, where it records and processes the data, including data cleaning and format conversion; and converts the processed data into semantic information that can be processed by a large language model. During the above transmission process, edge computing is performed on the IoT box side: for non-critical operating data, an abnormal data trigger mechanism is set. Only when a specific error code or abnormal situation is detected, the IoT box will publish the non-critical operating data within a fixed time before and after the fault to the fault analysis cloud platform; for critical operating data, the IoT box will publish it to the fault analysis cloud platform in real time.

8. The method for analyzing faults in a stereoscopic parking garage based on a large language model according to claim 6 is characterized in that: The characteristics of simple search questions include: ① querying a single, clear fault code handling method; ② querying standard maintenance procedures or parts replacement methods; ③ querying equipment parameters or configuration information; ④ simple and direct problem descriptions without the need to correlate multiple factors; The characteristics of complex analysis problems include: ① containing multiple related fault codes; ② involving mutual influence or cascading failures between devices; ③ repeated failures; ④ the need to analyze the impact of environmental factors on the failure; ⑤ the problem description contains words that require in-depth analysis, and the words that require in-depth analysis include: non-standard, root cause, and correlation.

Images