Intelligent management method for equipment problems based on large model

Abstract

The application discloses an intelligent management method for equipment problems based on a large model and relates to the technical field of computer processing.The method comprises the following steps: S01, obtaining original equipment problem descriptions from different business systems, and then performing multi-stage intelligent classification on the original equipment problem descriptions according to a preset instruction and a list containing multi-level classification options based on a large language model after instruction fine-tuning, and outputting corresponding equipment problem categories; S02, inputting the corresponding equipment problem categories into a pre-trained named entity recognition model to recognize and extract entity information having a unique identification function; and S03, extracting abnormal event information or key component state feature information contained in the entity information.The application realizes intelligent classification of equipment problems, accurate positioning of fine-grained repeatability and common problems, automatic cause analysis closed-loop management, and significantly reduces the labor cost and the equipment repeat failure rate.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to an intelligent management method for equipment problems based on a large model. Background Technology

[0002] In daily business operations, information on various equipment problems is scattered across multiple independent business systems. This dispersion stems from the fact that different business processes (such as production monitoring, work order reporting, preventative maintenance, and quality inspection) each have their own dedicated information management platforms. For each specific equipment problem, operators typically need to follow established business process specifications to conduct detailed inspections, records, and assessments. Simultaneously, to gain a deeper understanding of the root cause of the problem and develop solutions, staff must rely on their professional knowledge and experience to manually interpret and analyze the textual descriptions of the problems. While this approach can achieve a relatively high accuracy rate based on accumulated experience, its significant drawbacks include cumbersome processes, enormous time consumption, and high labor costs, severely limiting processing efficiency and the speed of response to massive numbers of problems.

[0003] The shortcomings of the aforementioned prior art are: Currently, there is a lack of cross-system data aggregation capabilities, and problems are scattered across different systems, making it difficult to automatically categorize them based on different device issues.

[0004] The inability to provide quantitative analysis results for recurring problems within the same department, or common problems occurring repeatedly throughout the company or across departments, prevents managers from making more targeted decisions, thus hindering the timely handling of these recurring issues.

[0005] Given a problem description, it is not possible to automatically extract the corresponding inspection clauses and solutions from the relevant management rules and provide the results of the cause analysis. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention provides a method for intelligent management of equipment problems based on a large model, the method comprising the following steps: S01. Obtain original device problem descriptions from different business systems, and then, based on the large language model after fine-tuning the instructions, perform multi-stage intelligent classification on the original device problem descriptions according to the preset instructions and a list containing multi-level classification options, and output the corresponding device problem category. S02. Input the corresponding equipment problem category into the pre-trained named entity recognition model to identify and extract the entity information with unique identification function; S03. Extract abnormal event information or key component status feature information contained in the entity information to construct an equipment knowledge graph, wherein the knowledge graph includes equipment entity nodes, accident mode nodes and their relationships. S04. The equipment knowledge graph, combined with graph representation learning technology, analyzes recurring accident pattern nodes associated with the same equipment entity node, as well as common accident pattern nodes associated across different equipment entity nodes or departments, and generates analysis results for recurring and common problems. S05. The analysis results of the repetitive and common problems are refined to generate a concise problem description. Then, the retrieval enhancement generation technology is used to retrieve relevant text blocks from a predefined institutional document library based on the concise problem description. S06. Combine the retrieved relevant text blocks with the concise problem description and input them into the large language model to generate the cause analysis results of the device problem.

[0007] Preferably, step S1, which involves constructing the large language model after instruction fine-tuning and the preset instructions and list containing multi-level classification options, includes the following steps: S11. Based on historical labeled device problem samples and prompt word engineering, construct an instruction fine-tuning sample dataset containing instructions, device problem description inputs, multi-level classification option list inputs, and expected category outputs; S12. Use the instruction fine-tuning sample dataset to perform instruction fine-tuning training on the large language model to obtain the fine-tuned large language model. S13. Input the prompt words containing preset instructions, descriptions of the original equipment problems to be classified, and a list of classification options for the current stage into the fine-tuned large language model, and let it output the corresponding classification results; S14. Based on business rules, the corresponding classification results are classified in multiple stages to obtain multiple preset instructions and a list of multi-level classification options corresponding to each preset instruction.

[0008] Preferably, the entity information with unique identification function in step S02 includes at least device ID and apparatus ID; The named entity recognition model is trained on a dataset annotated using the BIO annotation method.

[0009] Preferably, the construction of the device knowledge graph in step S3 specifically includes: S31. Entities are identified using a pre-trained named entity recognition model, and relation triples between entities are extracted from the device problem description using a pre-trained relation extraction model. The relation triples include a head entity, a relation, and a tail entity. S32. Parse or extract abnormal fluctuation event information or key component status feature information from the original equipment problem description; S33. Add the entity, the relation triplet, the abnormal fluctuation event information or key component status feature information and visual attribute nodes to the knowledge graph network. If the entity or accident mode node already exists, update its association information or count; if it does not exist, add a new node and establish an association relationship with the device entity node.

[0010] Preferably, the graph representation learning technique in step S04 includes: S41. Calculate the cosine similarity of the embedding vectors of the original device entity nodes to perform similarity detection; S42. Perform graph clustering on the entity nodes of the original device to identify similar devices that exhibit the same accident mode; S43. Use the embedded vectors of the aforementioned similar devices to perform link prediction and discover potential common cause associations; S44. Identify abnormal device segments that significantly deviate from the group after the device embedding vector.

[0011] Preferably, the cause analysis using retrieval enhancement generation technology in step S5 specifically includes: S51. Rewrite the concise problem description to generate multiple semantically similar queries; S52. In the system document library, smaller text blocks and their corresponding larger text blocks are used for retrieval to obtain mixed retrieval results; S53. The mixed retrieval results are merged and sorted, and the candidate text blocks are re-sorted using a retrieval sorting model to select the most relevant text blocks. S54. Combine the retrieved relevant text blocks with a concise problem description to form prompt words, and input them into the large language model to generate the cause analysis results.

[0012] Preferably, the hybrid retrieval result obtained in step S52 is a combination of retrieval results based on dense semantic vectors and keyword retrieval results based on sparse retrieval algorithms.

[0013] Preferably, the sparse retrieval algorithm is BM25, and the fusion sorting in step 53 adopts the RRF algorithm.

[0014] The present invention has at least the following beneficial effects: 1. By constructing a fine-tuning dataset, a single large language model is fine-tuned, enabling it to automatically perform multi-stage, fine-grained device problem classification based on a dynamically input list of category options. This avoids the high GPU resource consumption and model management complexity of traditional methods that require training and deploying independent models for each problem type, significantly reducing the reliance on manual labor and labor costs for classification, and achieving efficient and flexible classification automation.

[0015] 2. Named entity recognition is used to accurately extract unique identifiers such as equipment IDs and device IDs, and a knowledge graph of fused accident pattern nodes is built and updated in real time based on these identifiers. Combined with graph representation learning technology, recurring problems with the same equipment / device can be accurately identified within departments, and common problems with shared accident patterns can be identified across departments. This entity- and relationship-based analysis provides fine-grained quantitative results far exceeding traditional text similarity or clustering, enabling managers to accurately pinpoint management weaknesses and allocate resources for targeted rectification.

[0016] 3. This invention first extracts the core description of the problem during execution, and then uses advanced methods such as query rewriting, multi-granularity retrieval, fusion retrieval, and result reordering to efficiently and accurately retrieve the most relevant clauses from massive amounts of policy documents. Finally, the large model generates causal analysis based on the accurately retrieved context. This significantly reduces the time cost of manually reviewing and analyzing policy documents, improves the accuracy and efficiency of analysis, and allows business personnel to focus on higher-value work. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 Here is a flowchart of an intelligent management method for equipment problems based on a large model, provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of intelligent classification of equipment problems provided in Embodiment 1 of the present invention; Figure 3 This is a flowchart for analyzing repetitive and common problems provided in Embodiment 1 of the present invention; Figure 4 The flowchart for cause analysis provided in Embodiment 1 of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] Example 1

[0022] This embodiment provides an intelligent management method for equipment problems based on a large model. The method includes the following steps: Figure 1 As shown: S01. Obtain original device problem descriptions from different business systems, and then, based on the large language model after fine-tuning the instructions, perform multi-stage intelligent classification of the original device problem descriptions according to the preset instructions and a list containing multi-level classification options, and output the corresponding device problem categories. The above implementation of building a large language model after fine-tuning the instructions, as well as the preset instructions and the list containing multi-level classification options, includes the following steps: S11. Based on historical labeled device problem samples and prompt word engineering, construct an instruction fine-tuning sample dataset containing instructions, device problem description inputs, multi-level classification option list inputs, and expected category outputs; S12. Use the instruction fine-tuning sample dataset to train the large language model with instruction fine-tuning to obtain the fine-tuned large language model. S13. Input the prompt words containing preset instructions, descriptions of the original equipment problems to be classified, and a list of classification options for the current stage into the fine-tuned large language model, and let it output the corresponding classification results. S14. Based on business rules, perform multi-stage progressive classification on the corresponding classification results to obtain multiple preset instructions, and a list of multi-level classification options for each preset instruction.

[0023] Specifically, in combination Figure 2 As shown, using a finely tuned Large Language Model (LLM) based on pre-defined business rules (a multi-level classification option list) and instructions, the system automatically performs multi-stage classification of equipment problems. The input is raw equipment problem description text from different business systems. The output is the final problem category determined after multi-level classification (e.g., "Static Equipment - Defect Problem"). The pre-defined instructions tell the model what to do (e.g., "Based on the following problem description and option list, determine which category of equipment problem it belongs to?"). The multi-level classification option list reflects the structured options of the business classification process (e.g., first-stage options: "Equipment Problem", "Other Problems"); second-stage options: "Management Problems", "Defect Problems"; third-stage options: "Comprehensive", "Dynamic Equipment", "Static Equipment", "Electrical Equipment", "Instrumentation Equipment"). "Equipment Problems" can be things like common electronic component failure leading to inability to start, or incorrect wiring (a common problem with incorrect wiring), while "Other Problems" can be things like unstable voltage causing equipment malfunction, or lack of lubrication.

[0024] For original equipment issues collected from different systems, the business process is divided into three stages based on equipment issue classification. The first stage categorizes original equipment issues into equipment issues and other issues. The second stage categorizes the issues identified as equipment issues in the first stage into management issues and defect issues. The third stage categorizes the issues identified as management issues in the second stage into comprehensive, dynamic equipment, static equipment, electrical equipment, and instrumentation equipment based on chemical engineering specialties.

[0025] Then, a general large language model (such as LLaMA, ChatGLM, GPT, etc.) is selected as the base model. This base model is then supervised fine-tuned (SFT) using the instruction fine-tuning dataset constructed in step S11. The training process teaches the model to understand the intent of various classification instructions and to generate the desired category label as output based on the input problem description and option list. Essentially, it teaches the model to follow instructions to complete a specific text classification task. Multi-stage classification reasoning refers to constructing corresponding prompts: combining the pre-set instructions for that stage, the problem description text to be classified, and the classification option list for that stage. This prompt is input into the large model fine-tuned in step S12. The model reasons based on the prompt, generating a text-based classification result (i.e., outputting the desired category option). Based on the model's output and business rules, it decides whether to output the final category (e.g., if stage 2 outputs "defect problem," the process ends) or proceed to the next stage (e.g., if stage 1 outputs "equipment problem," it proceeds to stage 2; if stage 2 outputs "management problem," it proceeds to stage 3).

[0026] Traditional methods require training and deploying a separate model (such as BERT) for each classification task (stage 1 classification, stage 2 classification, stage 3 classification). The steps described above, however, only require training and deploying a single, finely tuned large model that can handle all stages of the classification task. This significantly reduces the number of models that need to be maintained and saves GPU resources required for training and inference.

[0027] S02. Input the corresponding equipment problem category into the pre-trained named entity recognition model to identify and extract the entity information with unique identification function; The entity information with unique identification function in the above embodiments includes at least device ID and apparatus ID. The named entity recognition model is trained on a dataset annotated using the BIO annotation method.

[0028] Specifically, this involves extracting key entity information from equipment problem description text that uniquely identifies a specific device or unit. To achieve this, a Named Entity Recognition (NER) model specifically designed for equipment management is first required. The model's training relies on a meticulously annotated dataset using the "BIO" (Begin, Inside, Outside) annotation specification. In this dataset, annotators explicitly mark keywords representing specific equipment identities, such as equipment IDs (e.g., "P-102A reactor") and unit IDs (e.g., "hydrocracking unit"), along with their boundaries. Using this high-quality annotated dataset, a pre-trained language model (e.g., BERT or its variants) is fine-tuned to learn and master the ability to accurately identify and extract target entities (equipment IDs, unit IDs, etc.) from the natural language text of equipment problem descriptions. The trained NER model is typically deployed and invoked via a REST API. When new equipment problem description text is input, the model automatically scans the text content, accurately identifies and extracts key entity information such as equipment IDs and unit IDs, providing structured identifier data for subsequent analysis steps. This lays a precise entity foundation for subsequent refined analysis. Traditional methods based on overall text similarity (such as TF-IDF or sentence vectors) remain at a vague "text similarity" level when analyzing repetitive or common problems, failing to accurately correlate them to specific physical devices or apparatuses. This step, however, successfully extracts uniquely identifying entities (device IDs, apparatus IDs) from unstructured text descriptions using named entity recognition technology, essentially giving each device problem a clear "identity label." This precise entity recognition capability reduces the dimensionality of subsequent analysis of device problems (especially repetitive and common problems) from the coarse textual semantic level to the specific device entity level. It enables the system to clearly know "which specific device (P-102A) or apparatus (hydrocracking unit) has a problem," rather than simply "a similarly described problem has occurred again." This ability to pinpoint specific entities is a key prerequisite and foundation for subsequent analysis of repetitive and common problems using knowledge graphs (step S04), significantly improving the accuracy of problem identification and the pertinence of management measures.

[0029] S03. Extract abnormal event information or key component status feature information contained in entity information to construct an equipment knowledge graph. The knowledge graph includes equipment entity nodes, accident mode nodes and their relationships. The construction of the device knowledge graph in the above embodiments specifically includes: S31. Use a pre-trained named entity recognition model to identify entities, and use a pre-trained relation extraction model to extract relation triples between entities from the device problem description. The relation triples include head entity, relation, and tail entity. S32. Parse or extract abnormal fluctuation event information or key component status characteristic information from the original equipment problem description; S33. Add entities, relation triples, abnormal fluctuation event information or key component status feature information and visual attribute nodes to the knowledge graph network. If an entity or accident mode node already exists, update its association information or count; if it does not exist, add a new node and establish an association relationship with the device entity node.

[0030] Specifically, a pre-trained Named Entity Recognition (NER) model is used to accurately identify and extract key entity information with unique identifiers, such as specific equipment IDs and device IDs, from the original equipment problem description or classified problem descriptions. Simultaneously, another pre-trained relation extraction model is used to extract the semantic relationships between entities from these descriptive texts. These relationships are represented in structured "triples" (head entity-relation-tail entity), such as "(pump P-101)-(belongs to)-(device A)" or "(valve V-202)-(occurs)-(leakage fault)". This step is crucial for transforming unstructured text information into structured knowledge.

[0031] In addition to identifying static device entities and their relationships, it is also necessary to extract dynamic information that characterizes specific fault conditions from the problem description. This mainly includes two types of key information: Abnormal fluctuation events: By analyzing the sensor time-series data (such as temperature, pressure, vibration, etc.) associated with the problem description, abnormal fluctuation patterns or events are identified using time-series analysis models (such as LSTM, Informer, etc.).

[0032] Key component status characteristics: For maintenance work orders or recorded equipment damage photos that may be included in the problem description, a Visual Modeling (VLM) is used to understand the images and generate descriptive text, from which key component status characteristics (such as "impeller wear," "seale aging and breakage," etc.) are extracted. These extracted abnormal events and status characteristics represent specific accident patterns.

[0033] All the information extracted in the above steps—identified device / device entities, relation triples, abnormal event nodes, and key component status feature nodes (collectively referred to as "accident pattern nodes")—is integrated into the device knowledge graph network. For device entities or accident pattern nodes identified in new problems, if they do not exist in the existing knowledge graph, these nodes are added to the graph, and associations between device entity nodes and these new accident pattern nodes are established (e.g., "(Device X) - (Occurrence) - (Abnormal Event Y)"). If device entities or accident pattern nodes identified in new problems (especially those with the same abnormal event or the same component damage characteristics) already exist in the knowledge graph, the association information of the accident pattern node (such as the list of associated devices) is updated, and the occurrence count of the accident pattern node is increased. This "counting" mechanism is the core quantitative basis for subsequent analysis of repetitive and common problems. The final knowledge graph is a dynamically evolving semantic network. It uses specific equipment and devices as core nodes, connects organizational information nodes such as departments and locations through relational edges, and connects various nodes representing specific accident modes (abnormal events, component damage) through relational edges such as "occurred" and "has characteristics". Each accident mode node accumulates and records the number of times it has occurred (count) and the specific equipment entity it is associated with.

[0034] S04. Using equipment knowledge graphs combined with graph representation learning technology, we analyze recurring accident pattern nodes associated with the same equipment entity node, as well as common accident pattern nodes associated across different equipment entity nodes or departments, and generate analysis results for recurring and common problems. The graph representation learning techniques in the above embodiments include: S41. Calculate the cosine similarity of the embedding vectors of the original device entity nodes to perform similarity detection; S42. Perform graph clustering on the entity nodes of the original equipment to identify similar equipment that exhibit the same accident mode. S43. Use embedded vectors from similar devices to predict links and discover potential common-cause associations; S44. Identify abnormal device segments that significantly deviate from the group after the device embedding vector.

[0035] Specifically, in combination Figure 3As shown, the device knowledge graph constructed and updated in step S03, along with graph representation learning technology, is utilized. The knowledge graph contains nodes representing specific devices (device entity nodes) and nodes representing specific fault phenomena (such as specific abnormal events, critical component damage states, etc., i.e., accident mode nodes), connected by edges (associations), recording the number of times the fault occurs (count information) and information such as the department to which it belongs. When a new device problem enters the system, after extracting its unique identifier entity (such as device ID, unit ID), the system accurately locates the corresponding device entity node in the knowledge graph. Simultaneously, the specific fault details contained in the problem (through extracted abnormal event information or critical component status feature information) are associated with specific accident mode nodes in the graph (such as "bearing temperature too high," "seal leakage," etc.). If the accident mode node already exists, its association count with the device node is updated (count increases); if it does not exist, a new accident mode node is created and associated with the device node.

[0036] The system employs graph representation learning techniques (such as GNN graph neural networks) to perform in-depth analysis of knowledge graphs. This includes: Calculate the "distance" or similarity between different device nodes in the graph space. Highly similar device nodes may indicate that they are in similar environments or have similar designs, making them more prone to the same problems. Group (cluster) device nodes based on their connectivity, attribute similarity, etc. Different devices appearing in the same cluster are associated with the same incident mode nodes, indicating a common problem across devices. Analyze the graph structure to predict which device nodes may have undocumented but actually related potential connections (e.g., two seemingly unrelated devices frequently experiencing the same failure may be because they share a defective component or maintenance process). Identify device nodes that are clearly "isolated" or "offset" from their group (such as similar devices or devices in the same department) in the graph embedding space. This may suggest that the device has unique design flaws, installation problems, or operational vulnerabilities.

[0037] S05. Extract the analysis results of repetitive and common problems to generate concise problem descriptions, and then use retrieval enhancement generation technology to retrieve relevant text blocks from a predefined institutional document library based on the concise problem descriptions; The above embodiments utilize search-enhanced generation technology for cause analysis, specifically including: S51. Rewrite the concise problem description to generate multiple semantically similar queries; S52. In the institutional document database, smaller text blocks and their corresponding larger text blocks are used for retrieval to obtain mixed retrieval results; S53. The mixed search results are merged and ranked (RRF algorithm), and the candidate text blocks are re-ranked using the search ranking model to select the most relevant text blocks; S54. Combine the retrieved relevant text blocks with a concise problem description to form prompt words, and input them into the large language model to generate the cause analysis results.

[0038] S06. Combine the retrieved relevant text blocks with a concise problem description and input them into the large language model to generate the cause analysis results of the device problem.

[0039] Furthermore, the hybrid retrieval results obtained in step S52 are obtained by combining the retrieval results based on dense semantic vectors and the keyword retrieval results based on the sparse retrieval algorithm (BM25).

[0040] Specifically, in combination Figure 4 As shown, firstly, a concise description of the equipment problem is extracted. Then, the Retrieval Enhanced Generation (RAG) technique is used to retrieve relevant text blocks from a predefined institutional document library. Finally, these text blocks are combined with the concise description and input into a Large Language Model (LLM) to generate causal analysis results.

[0041] Prompt engineering is used to refine the original description of the device problem after intelligent classification, generating a concise and core problem description (the concise problem description). To overcome the low recall rate that may result from a single query, the system automatically rewrites the semantics of this concise description, generating multiple queries that express the same core meaning but may use different wording. This simulates the different ways humans might search for information, improving the coverage of subsequent searches.

[0042] The retrieval in step S52 not only searches entire document paragraphs (large text blocks) but also smaller text fragments (small text blocks). Small text blocks can more accurately match key points in the query (improving accuracy), and after finding a matching small text block, the system will also include its larger contextual text blocks (such as entire chapters or paragraphs) as candidates, ensuring that the LLM has sufficient background information for understanding (improving analysis quality). The system simultaneously uses two retrieval methods: one is retrieval based on dense semantic vectors, which can understand the deep semantic similarity between the query and the document; the other is sparse retrieval based on keyword matching (such as the BM25 algorithm). These two methods each have their advantages (semantics vs. keywords). The system merges the results obtained from the two methods. The merged hybrid retrieval results are sorted using the RRF (Reverse Ranking Fusion) algorithm. RRF can comprehensively consider the ranking positions of the results returned by different retrieval methods, intelligently merging them into a better comprehensive ranking list, ensuring that the most relevant text blocks are ranked first.

[0043] S06. Combine the retrieved relevant text blocks with a concise problem description and input them into the large language model to generate the cause analysis results of the device problem.

[0044] Specifically, the candidate text blocks initially ranked by RRF are then scored and re-ranked by an advanced retrieval ranking model (such as the open-source model from the Beijing Academy of Artificial Intelligence). This model can more precisely evaluate the relevance of each candidate text block to the original query, further filtering out the most relevant and accurate text blocks as the final context provided to the LLM. Finally, the system combines the original concise problem description with the most relevant text blocks obtained through the complex retrieval, fusion, and ranking process to form a structured prompt, which is then input into the Large Language Model (LLM). LLM utilizes its powerful language understanding and reasoning capabilities to automatically generate a cause analysis of the device problem based on the retrieved precise and relevant context. This process is completely automated, requiring no manual intervention in retrieval and analysis.

[0045] This first embodiment fine-tunes a single large language model by constructing an instruction fine-tuning dataset, enabling it to automatically complete multi-stage, fine-grained equipment problem classification based on a dynamically input list of category options. This avoids the high GPU resource consumption and model management complexity of traditional methods that require training and deploying independent models for each problem type, significantly reducing the reliance on manual labor and labor costs, and achieving efficient and flexible classification automation. Secondly, named entity recognition is used to accurately extract unique identifiers such as device IDs and apparatus IDs, and a knowledge graph of fused accident pattern nodes is constructed and updated in real time based on this. Combined with graph representation learning technology, recurring problems of the same equipment / apparatus can be accurately identified within departments, and common problems with shared accident patterns can be identified across departments. This entity- and relationship-based analysis provides fine-grained quantitative results far exceeding traditional text similarity or clustering, enabling managers to accurately locate management weaknesses and allocate resources for targeted rectification. Furthermore, during execution, the core description of the problem is first extracted, and then advanced methods such as query rewriting, multi-granularity retrieval, fusion retrieval, and result reordering are used to efficiently and accurately retrieve the most relevant clauses from massive amounts of policy documents. Finally, the large model generates causal analysis based on the precisely retrieved context. This significantly reduces the time cost of manually reviewing and analyzing policy documents, improves the accuracy and efficiency of the analysis, and allows business personnel to focus on higher-value work.

[0046] Example 2

[0047] This invention provides a non-transitory computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the following steps: The system acquires original device problem descriptions from different business systems, and then performs multi-stage intelligent classification of the original device problem descriptions based on the large language model after fine-tuning the instructions, according to the preset instructions and a list containing multi-level classification options, and outputs the corresponding device problem categories. The corresponding equipment problem category is input into the pre-trained named entity recognition model to identify and extract entity information that has a unique identifier. Extract abnormal event information or key component status feature information contained in entity information to construct an equipment knowledge graph. The knowledge graph includes equipment entity nodes, accident mode nodes and their relationships. The equipment knowledge graph, combined with graph representation learning technology, analyzes recurring accident pattern nodes associated with the same equipment entity node, as well as common accident pattern nodes associated across different equipment entity nodes or departments, generating analysis results for recurring and common problems. The analysis results of repetitive and common problems are refined to generate concise problem descriptions. Then, retrieval enhancement generation technology is used to retrieve relevant text blocks from a predefined institutional document library based on the concise problem descriptions. The relevant text blocks are combined with a concise problem description and input into a large language model to generate the cause analysis results of the device problem.

[0048] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0050] Example 3

[0051] This invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one instruction or at least one program segment, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the following steps: Create target user permissions and obtain several target project records. The target project records are the project records corresponding to the target project. The target project is the project in which the target user edits and updates several data of the BIM model. The project records include the estimated materials, actual production materials, and materials generated into the warehouse.

[0052] The system acquires original device problem descriptions from different business systems, and then performs multi-stage intelligent classification of the original device problem descriptions based on the large language model after fine-tuning the instructions, according to the preset instructions and a list containing multi-level classification options, and outputs the corresponding device problem categories. The corresponding equipment problem category is input into the pre-trained named entity recognition model to identify and extract entity information that has a unique identifier. Extract abnormal event information or key component status feature information contained in entity information to construct an equipment knowledge graph. The knowledge graph includes equipment entity nodes, accident mode nodes and their relationships. The equipment knowledge graph, combined with graph representation learning technology, analyzes recurring accident pattern nodes associated with the same equipment entity node, as well as common accident pattern nodes associated across different equipment entity nodes or departments, generating analysis results for recurring and common problems. The analysis results of repetitive and common problems are refined to generate concise problem descriptions. Then, retrieval enhancement generation technology is used to retrieve relevant text blocks from a predefined institutional document library based on the concise problem descriptions. The relevant text blocks are combined with a concise problem description and input into a large language model to generate the cause analysis results of the device problem.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent management of equipment problems based on a large model, characterized in that, The method includes the following steps: S01. Obtain original device problem descriptions from different business systems, and then, based on the large language model after fine-tuning the instructions, perform multi-stage intelligent classification on the original device problem descriptions according to the preset instructions and a list containing multi-level classification options, and output the corresponding device problem category. S02. Input the corresponding equipment problem category into the pre-trained named entity recognition model to identify and extract the entity information with unique identification function; S03. Extract abnormal event information or key component status feature information contained in the entity information to construct an equipment knowledge graph, wherein the knowledge graph includes equipment entity nodes, accident mode nodes and their relationships. S04. The equipment knowledge graph, combined with graph representation learning technology, analyzes recurring accident pattern nodes associated with the same equipment entity node, as well as common accident pattern nodes associated across different equipment entity nodes or departments, and generates analysis results for recurring and common problems. S05. The analysis results of the repetitive and common problems are refined to generate a concise problem description. Then, the retrieval enhancement generation technology is used to retrieve relevant text blocks from a predefined institutional document library based on the concise problem description. S06. Combine the retrieved relevant text blocks with the concise problem description and input them into the large language model to generate the cause analysis results of the device problem.

2. The intelligent management method for equipment problems based on a large model according to claim 1, characterized in that, The step S1 of constructing the large language model after fine-tuning the instructions, as well as the preset instructions and the list containing multi-level classification options, includes the following steps: S11. Based on historical labeled device problem samples and prompt word engineering, construct an instruction fine-tuning sample dataset containing instructions, device problem description inputs, multi-level classification option list inputs, and expected category outputs; S12. Use the instruction fine-tuning sample dataset to perform instruction fine-tuning training on the large language model to obtain the fine-tuned large language model. S13. Input the prompt words containing preset instructions, descriptions of the original equipment problems to be classified, and a list of classification options for the current stage into the fine-tuned large language model, and let it output the corresponding classification results; S14. Based on business rules, the corresponding classification results are classified in multiple stages to obtain multiple preset instructions and a list of multi-level classification options corresponding to each preset instruction.

3. The intelligent management method for equipment problems based on a large model according to claim 1, characterized in that, The entity information with unique identification function mentioned in step S02 includes at least device ID and device ID; The named entity recognition model is trained on a dataset annotated using the BIO annotation method.

4. The intelligent management method for equipment problems based on a large model according to claim 1, characterized in that, The construction of the device knowledge graph in step S3 specifically includes: S31. Entities are identified using a pre-trained named entity recognition model, and relation triples between entities are extracted from the device problem description using a pre-trained relation extraction model. The relation triples include a head entity, a relation, and a tail entity. S32. Parse or extract abnormal fluctuation event information or key component status feature information from the original equipment problem description; S33. Add the entity, the relation triplet, the abnormal fluctuation event information or key component status feature information and visual attribute nodes to the knowledge graph network. If the entity or accident mode node already exists, update its association information or count; if it does not exist, add a new node and establish an association relationship with the device entity node.

5. The intelligent management method for equipment problems based on a large model according to claim 1, characterized in that, The graph representation learning technique mentioned in step S04 includes: S41. Calculate the cosine similarity of the embedding vectors of the original device entity nodes to perform similarity detection; S42. Perform graph clustering on the entity nodes of the original device to identify similar devices that exhibit the same accident mode; S43. Use the embedded vectors of the aforementioned similar devices to perform link prediction and discover potential common cause associations; S44. Identify abnormal device segments that significantly deviate from the group after the device embedding vector.

6. The intelligent management method for equipment problems based on a large model according to claim 1, characterized in that, The cause analysis using retrieval enhancement generation technology in step S5 specifically includes: S51. Rewrite the concise problem description to generate multiple semantically similar queries; S52. In the system document library, smaller text blocks and their corresponding larger text blocks are used for retrieval to obtain mixed retrieval results; S53. The mixed retrieval results are merged and sorted, and the candidate text blocks are re-sorted using a retrieval sorting model to select the most relevant text blocks. S54. Combine the retrieved relevant text blocks with a concise problem description to form prompt words, and input them into the large language model to generate the cause analysis results.

7. The intelligent management method for equipment problems based on a large model according to claim 6, characterized in that, The hybrid retrieval result obtained in step S52 is a combination of retrieval results based on dense semantic vectors and keyword retrieval results based on sparse retrieval algorithms.

8. The intelligent management method for equipment problems based on a large model according to claim 7, characterized in that, The sparse retrieval algorithm is BM25, and the fusion sorting in step 53 adopts the RRF algorithm.

9. A non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the intelligent device problem management method based on a large model as described in any one of claims 1-8.

10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the intelligent device problem management method based on a large model as described in any one of claims 1-8.

Images