Equipment exception attribution method and system based on knowledge graph

Abstract

The invention provides an equipment exception attribution method and system based on a knowledge graph, and the method comprises the steps: firstly, recognizing a user intention according to a problem inputted by a user; next, the user intentions are classified, and under the condition that the classification result is equipment anomaly reason analysis, the equipment anomaly agent is called to process the problem input by the user, specifically, according to the problem, a pre-constructed knowledge graph is retrieved, and the retrieval result comprises historical anomaly reasons of equipment anomaly in the problem; if the retrieval result comprises a plurality of possible historical abnormal reasons, continuing to retrieve an inspection action corresponding to each possible historical abnormal reason, and asking a question to a user based on the inspection actions to obtain user feedback; and according to the user feedback, determining a real reason of the equipment abnormity and feeding back the real reason to the user.

Description

Technical Field

[0001] The present invention relates to the field of intelligent diagnosis of industrial equipment faults, and in particular to a method and system for attributing equipment anomalies based on a knowledge graph. Background Art

[0002] With the improvement of industrial automation, the complexity of equipment continues to increase, which puts higher requirements on the rapid diagnosis of equipment anomalies. Nowadays, the analysis and tracking process after anomalies occur in factory equipment is heavily dependent on the experience of professional engineers, which often leads to problems such as low efficiency, low accuracy, lack of migration flexibility on new equipment, and being restricted by information islands and knowledge transfer barriers.

[0003] In recent years, the development of large language models (LLMs) has provided a new solution for device anomaly attribution analysis. However, existing large language models still have many problems in understanding complex queries in professional fields and providing accurate causes of device anomalies, including: 1. Limitations of query understanding: Natural language processing technology of large language models may not be sufficient to accurately understand the professional terms or abbreviations of a specific industry.

[0004] 2. Insufficient information: When faced with an equipment anomaly of unknown cause, the user may not be able to accurately provide all the information related to the equipment anomaly at the beginning. In this case, the large model cannot ask questions based on the missing information, resulting in incorrect answers.

[0005] 3. Difficulty in reasoning about equipment anomaly analysis: Equipment anomalies are often the result of multiple factors. The manifestations of equipment anomalies caused by different reasons may also be similar. As a result, equipment anomaly analysis often requires a lot of professional knowledge, comprehensive judgment of multiple aspects of information, and complex reasoning to obtain accurate results. Summary of the invention

[0006] To solve the above problems, the present invention provides a method and system for attributing device anomalies based on knowledge graph, including: According to a first aspect, the present invention provides a method for analyzing the cause of an abnormality of a device, comprising the following steps: Identify user intent based on the questions they enter; Classify the user intention, and when the classification result is an analysis of the cause of the device abnormality, assign the problem to a device abnormality agent for processing; Calling the device abnormality agent to process the question input by the user specifically includes: searching a pre-built knowledge graph according to the question, the search results include the historical abnormality causes of the device abnormality in the question, and determining the next action according to the search results: If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is fed back to the user as the real cause of the device abnormality; If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the true cause of the equipment abnormality and feed it back to the user.

[0007] In some embodiments, determining the real cause of the device abnormality according to the user feedback specifically includes: Determining whether the user feedback exceeds the capability range of the abnormal intelligent agent of the device; If the user feedback does not exceed the capability scope of the device abnormality intelligent agent, it is determined based on the user feedback whether the historical abnormality cause corresponding to the inspection action is the real cause of the device abnormality; if it is determined that the historical abnormality cause is not the real cause of the device abnormality, other inspection actions are reselected to ask questions to the user until the real cause of the device abnormality is determined.

[0008] In some embodiments, after determining whether the user feedback exceeds the capability range of the abnormal device agent, the method further includes: If it is determined that the user feedback exceeds the capability range of the abnormal intelligent agent of the device, the user intention is re-identified based on the user feedback.

[0009] In some embodiments, the pre-construction of the knowledge graph includes the following steps: The domain semantic network is used to design the ontology of engineers' equipment operation and maintenance experience, and the semantic representation of the network ontology language (OWL) is obtained; Based on the prompt words, the knowledge graph instance is extracted from the equipment analysis report and the device manual using the domain big model; the prompt words are extracted based on the OWL semantic representation; The OWL semantic representation and the instance are stored on a knowledge graph platform, which includes an elastic search engine (Elasticsearch, ES) and a distributed database.

[0010] In some embodiments, identifying the user's intention according to the question input by the user specifically includes: Use the domain model to identify private words in user input questions; Based on the private vocabulary recognition results, the user input question is expanded through retrieval augmentation generation (RAG), and the user intention is identified based on the expanded question.

[0011] In some embodiments, classifying the user intent specifically includes: Combine user intent with hierarchical dynamic few-shot prompts to form a query, which is sent to different agents. The agents determine the query category and give a vote. The large model rewrites the query using different prompt words, repeats the voting multiple times, counts the average number of votes for each category, and determines the user intent classification based on the average number of votes.

[0012] In some embodiments, the checking action includes multiple steps, wherein asking the user questions based on the checking action specifically includes: Use the large model to select the target inspection action from multiple inspection actions; The user is asked a question regarding an inspection result of the target inspection action.

[0013] In some embodiments, after determining the true cause of the device abnormality, the method further includes: according to the true cause of the device abnormality, obtaining monitoring parameters associated with the device abnormality through path retrieval of the knowledge graph, and feeding them back to the user.

[0014] According to a second aspect, the present invention further provides a device abnormality attribution analysis system, comprising: The intention recognition module is used to recognize the user's intention based on the question input by the user; An intention routing module is used to classify the user intention, and when the classification result is an analysis of the cause of the device abnormality, assign the problem to the device abnormality agent for processing; A retrieval processing module is used to call the device abnormality agent to process the question input by the user, specifically including: according to the question, searching the pre-built knowledge graph, and the search results include the historical abnormality causes of the device abnormality in the question; If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is fed back to the user as the real cause of the device abnormality; If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the true cause of the equipment abnormality and feed it back to the user.

[0015] In some embodiments, the retrieval processing module is further used to: Determining whether the user feedback exceeds the capability range of the abnormal intelligent agent of the device; If it is determined that the user feedback does not exceed the capability range of the device abnormality agent, then judging whether the historical abnormality cause corresponding to the inspection action is the real cause of the device abnormality according to the user feedback; if it is determined that the historical abnormality cause is not the real cause of the device abnormality, then reselecting other inspection actions to ask questions to the user until the real cause of the device abnormality is determined; If it is determined that the user feedback exceeds the capability range of the abnormal intelligent agent of the device, the user intention is re-identified based on the user feedback.

[0016] In some embodiments, the system further includes a knowledge graph construction module for constructing a knowledge graph, and the knowledge graph construction module specifically includes: The ontology design submodule is used to design the ontology of engineers’ equipment operation and maintenance experience using the domain semantic network to obtain OWL semantic representation; An instance extraction submodule is used to extract knowledge graph instances from equipment analysis reports and device manuals based on prompt words and using a domain big model; the prompt words are extracted based on the OWL semantic representation; The storage submodule is used to store the OWL semantic representation and the instance on a knowledge graph platform, wherein the knowledge graph platform includes ES and a distributed database.

[0017] In some embodiments, the intention recognition module is specifically configured as follows: Use the domain model to identify private words in user input questions; Based on the private vocabulary recognition results, the user input question is expanded through RAG document search, and the user intention is identified based on the expanded question.

[0018] In some embodiments, the intent routing module is specifically configured as follows: Combine user intent and hierarchical dynamic few-shot prompts to form a query, which is sent to different agents. The agents judge the query category and give a vote. The large model rewrites the query using different prompt words, repeats the voting multiple times, counts the average number of votes for each category, and determines the user intent classification based on the average number of votes.

[0019] In some embodiments, asking the user questions based on the checking action specifically includes: Use the large model to select the target inspection action from multiple inspection actions; The user is asked a question regarding an inspection result of the target inspection action.

[0020] In some embodiments, the system further includes a monitoring parameter recommendation module for obtaining monitoring parameters associated with the device abnormality through path retrieval of the knowledge graph based on the actual cause of the device abnormality, and feeding it back to the user.

[0021] The device anomaly attribution analysis method and system provided by the present invention can quickly understand the user's intention through the user's questions, reduce the time for manual information entry and interpretation in traditional device anomaly analysis, and reduce the operating threshold of device anomaly analysis, so that device anomaly analysis tracking does not have to rely on experienced engineers, thereby improving the overall device anomaly analysis speed and equipment maintenance efficiency, and reducing the manpower cost of device anomaly analysis. In addition, the method and system can guide users to perform inspection actions and supplement missing information, thereby improving the accuracy of equipment anomaly cause analysis. The method and system perform equipment anomaly analysis inspections based on historical anomaly causes and inspection action experience recorded in the knowledge graph, simplifying the device anomaly attribution analysis process. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 The construction and application process of the knowledge graph in the present invention is shown; Figure 2 The process of the device abnormality attribution analysis method provided by the present invention is shown; Figure 3 The dialogue interaction process of the device abnormal intelligent agent provided by the present invention is shown; Figure 4 A specific implementation process of the device abnormality attribution analysis method provided by the present invention is shown; Figure 5 The structure of the device abnormality attribution system provided by the present invention is shown. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0025] In order to overcome the problems in the current analysis and tracking process after an abnormality occurs in a device, the present invention provides a device abnormality attribution analysis and monitoring parameter recommendation method based on knowledge graph and dialogue linkage. The following is a detailed description of the implementation of the method in conjunction with the accompanying drawings.

[0026] Before conducting equipment anomaly attribution analysis, it is necessary to pre-build an equipment anomaly knowledge graph as the basis for equipment anomaly attribution analysis. The knowledge graph used in the present invention collects and integrates engineers' equipment operation and maintenance experience, field-related knowledge, and relevant examples in factory analysis reports and device manuals as the basis for equipment anomaly attribution analysis. Figure 1 The construction and application process of the knowledge graph is shown in Figure 2. Figure 1 As shown, the construction of the knowledge graph includes the following steps: Firstly, the domain semantic network is used to design the engineers' equipment operation and maintenance experience into an ontology, and the OWL semantic representation of the ontology is obtained.

[0027] It should be understood that a domain semantic network is a knowledge representation model specific to a certain discipline or industry domain. It simulates the knowledge organization and association within a domain by organizing the concepts, entities, attributes, and relationships within the domain into a directed graph or network structure.

[0028] Ontology is a structured and formal description of knowledge in a certain field. It clarifies the concepts and relationships between them in a specific field and is the skeleton of the knowledge graph. The core of ontology design is to create a shared conceptual model, which includes elements such as defining concepts, relationships between concepts, functions, axioms, and instances.

[0029] After the ontology design is completed, the prompt words extracted based on the OWL semantic representation of the ontology are combined with the domain big model to extract knowledge graph instances from equipment analysis reports and device manuals.

[0030] Specifically, the domain big model first extracts prompt words according to the ontology design results, and informs the domain big model when extracting prompt words that the prompt words are used to extract knowledge graph instances from the factory analysis report and device manual. The domain big model then extracts knowledge graph instances from the factory analysis report and device manual in combination with the prompt words.

[0031] The domain big model is obtained by fine-tuning the pre-trained big language model with instructions, and the fine-tuning sample is constructed based on the domain information with proper nouns. The domain big model can understand the language patterns, proper nouns and concepts specific to the field. The domain big model is designed to solve complex problems in a specific field and show excellent performance in the field. The domain big model used in the present invention is a domain big model that has been pre-trained with equipment operation and maintenance field data and fine-tuned with instructions.

[0032] The knowledge graph instance includes nodes, relationships and attributes. Among them, nodes refer to specific objects with specific meanings or references in the text, which are entities in the knowledge graph, such as equipment, equipment type, maintenance methods, fault types, etc. Edges refer to semantic relationships between entities, such as "belongs to", "cause", "responsible for", etc. Attributes are supplementary descriptions of nodes or edges, such as the model of the equipment, the impact of the fault, etc.

[0033] The prompt words used to extract knowledge graph instances can be: (1) Extract entity prompts: "Identify the device name and related information in the text, such as model, manufacturer, and other attributes." "Find the fault type in the text and extract its priority and probability of occurrence." (2) Prompt words for extracting relationships: “Identify the relationship between a certain failure described in the text and its consequences.” “Extract the relationship between maintenance methods and corresponding failures from the following text.” (3) Prompt words for extracting attributes: "Extract the attribute values ​​of entities (such as the model of the equipment, the priority of the fault, etc.) from the following text." "Please extract the operating status (normal / abnormal) of the equipment in the text." "Find the description of the operating steps of each maintenance method in the text." After the knowledge graph instance is extracted, the OWL semantic representation and the instance are stored on a knowledge graph platform, which includes ES (Elasticsearch) and a distributed database.

[0034] The knowledge graph is stored in a dual database structure of ES + distributed database. ES is used for reasoning and retrieval of the knowledge graph, and the distributed database is used for storing triple data and related multimodal attachments (pictures or videos). The distributed database can be Accumulo, Apache Cassandra, Apache HBase, etc.

[0035] The above introduces the construction process of the knowledge graph. The following briefly introduces the application process of the knowledge graph. Figure 1 As shown in the figure, after the knowledge graph is built, the user asks the domain big model questions about the cause of equipment abnormality and the recommendation of monitoring parameters. After the big model recognizes the intent of the user's input questions, it obtains the user's intent and the key entities in the intent, associates and binds the entities with the word slots, and uses them as the input of the equipment abnormality agent. The equipment abnormality agent retrieves the knowledge graph based on the user's intent and the key entities in the intent, obtains the cause of the equipment abnormality, and based on this, filters and recommends the monitoring parameters associated with the equipment abnormality, reducing the workload of engineers in parameter selection during equipment monitoring and realizing fast and intelligent parameter selection.

[0036] The following describes the device abnormality cause analysis method provided by the present invention in conjunction with a specific embodiment. The flowchart of the method is as follows: Figure 2 As shown, the following steps are included: 1. Intent recognition: Identify user intent based on the questions entered by the user.

[0037] In some embodiments, this step specifically includes: using the domain macro model to identify private vocabulary in the question input by the user. Based on the private vocabulary identification result, the question input by the user is expanded by searching the document through RAG, and the user intention is identified based on the expanded question.

[0038] 2. Intent routing: Classify the user intent, and when the classification result is an analysis of the cause of the device abnormality, assign the problem to the device abnormality agent for processing.

[0039] In some embodiments, classifying the user intent specifically includes: combining the user intent and the hierarchical dynamic few-shot prompt words to form a query, sending it to different agents, and the agents respectively determine the category of the query and give votes. The large model rewrites the query using different prompt words, and after repeating the voting process multiple times, the large model counts the average number of votes for each category, and determines the user intent classification based on the average number of votes.

[0040] When classifying the user intent, the historical classification results of similar intents can also be referred to. Specifically, the large model retrieves similar historical intents based on the intent recognition results, and uses the classification process (i.e., agent voting records) and classification results of similar historical intents as the basis for classification decision-making to enhance the accuracy of classification.

[0041] The different agents are good at handling different types of problems, including device anomaly agents, time series database query agents, log analysis agents, etc. The large model can optimize the intent routing process through continuous learning. For example, it can learn which agents perform best when handling specific types of requests and adjust future routing decisions based on this information.

[0042] After classifying the user's intention, the big model routes the user's input questions to different agents for processing according to the classification results. In the case where the classification result is an analysis of the cause of an equipment abnormality, the question is assigned to the equipment abnormality agent for processing.

[0043] 3. Calling the device abnormality agent to process the problem input by the user, specifically including: According to the problem, a pre-built knowledge graph is retrieved, and the retrieval results include historical abnormal causes of the device abnormality in the problem.

[0044] If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is used as the real cause of the device abnormality.

[0045] If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the real cause of the equipment abnormality.

[0046] The following is a detailed introduction to the process of device abnormality intelligent agent handling problems.

[0047] First, according to the problem, a pre-built knowledge graph is retrieved, and the retrieval results include historical abnormal causes of the equipment abnormality in the problem.

[0048] In some embodiments, a tree search algorithm without backtracking is used to search the knowledge graph, which can reduce the search space and improve the search efficiency.

[0049] The agent retrieves the knowledge graph based on the description of the device anomaly in the question, and based on the historical analysis process of similar historical anomalies, it can obtain the historical anomaly cause of the device anomaly that is consistent with the description in the question. Next, the next action is performed based on the retrieval results.

[0050] If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is fed back to the user as the real cause of the device abnormality.

[0051] In this case, since the user's description of the equipment abnormality is clear and the input information is complete, the equipment abnormality agent can retrieve the real cause of the equipment abnormality based on this information by retrieving the knowledge graph and feedback it to the user.

[0052] If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the real cause of the equipment abnormality.

[0053] In this case, since the user's problem description does not accurately provide all the information related to the device abnormality, there are multiple historical abnormality causes that match the description. The user needs to provide further information to confirm the true cause of the device abnormality.

[0054] To this end, the device abnormality agent needs to retrieve and obtain the inspection action corresponding to each possible historical abnormality cause, ask questions to the user based on the inspection action, and obtain the inspection results fed back by the user.

[0055] In some embodiments, asking the user questions based on the checking action specifically includes: Use the large model to select the target inspection action from multiple inspection actions; The user is asked a question regarding an inspection result of the target inspection action.

[0056] For example, the device abnormality agent allows the big model to autonomously select the best target inspection action from three candidate inspection actions, and asks the user about the inspection result of the target inspection action. This controlled execution process can balance the cost-effectiveness and speed of task execution and improve execution efficiency.

[0057] After asking the user questions based on the inspection action, the device abnormality intelligent experience will make a decision on the next action based on the user feedback, including: Determine whether the user feedback exceeds the capability range of the abnormal intelligent agent of the device.

[0058] If the user feedback includes the inspection result of the inspection action, or includes questions related to equipment abnormality, the determination is no.

[0059] If the user feedback does not exceed the capability of the device abnormality agent, the user feedback includes the answer to the previous question asked by the device abnormality agent, that is, the inspection result corresponding to the inspection action. The device abnormality agent determines whether the historical abnormality cause corresponding to the inspection action is the real cause of the device abnormality based on the inspection result. If it is not, it reselects other inspection actions to ask questions to the user, and repeats multiple rounds of dialogue until the real cause of the device abnormality is determined.

[0060] If it is determined that the user feedback exceeds the capability range of the abnormal intelligent agent of the device, the user intention is re-identified based on the user feedback.

[0061] In this way, by checking in each round of dialogue whether the current user feedback exceeds the capabilities of the current agent, sudden topic changes can be avoided during the dialogue. Once a topic that exceeds the capabilities of the current agent appears, rerouting is triggered to ensure the user experience of multiple rounds of dialogue.

[0062] The above describes the process of the device abnormality agent processing the user input problem and obtaining the real cause of the device abnormality. In general, the dialogue interaction process of the device abnormality agent is as follows: Figure 3As shown. After the device abnormality agent receives the question input by the user, it searches the knowledge graph. If the real cause of the device abnormality can be obtained, it will be fed back to the user and the conversation is completed. After searching the knowledge graph, if the real cause of the device abnormality cannot be obtained due to unclear user description, the inspection action is obtained through dialogue interaction until the user description is clear or the input is complete, and the real cause of the device abnormality can be obtained. During the dialogue interaction, if the topic is switched, causing the topic to exceed the capability of the device abnormality agent, the intention recognition is performed again.

[0063] In some embodiments, after determining the real cause of the equipment abnormality, the method further includes: according to the real cause of the equipment abnormality, obtaining monitoring parameters associated with the equipment abnormality through the path retrieval of the knowledge graph, and feeding it back to the user. Recommending monitoring parameters based on the cause of equipment abnormality can significantly improve the accuracy of fault diagnosis and prediction, and can achieve preventive maintenance, reduce maintenance costs, extend equipment life, and ensure production safety.

[0064] The following is a description of how the method can perform equipment abnormality attribution analysis and monitoring parameter recommendation with reference to specific examples. A specific implementation process of the equipment abnormality attribution analysis method provided by the present invention is as follows: Figure 4 It should be noted that this example is only used to show the implementation process of the device abnormality attribution analysis method provided by the present invention, and the device abnormality and its specific cause mentioned therein are only used as examples.

[0065] Example User input question: The film thickness of PECVD01PM2 of PF-300T equipment is abnormal. Which parameters need to be monitored throughout the process? Intent recognition: User intent is used for device anomaly attribution analysis and monitoring parameter recommendation.

[0066] Intent routing: Based on the user’s intent, the problem is classified into device abnormality cause analysis and assigned to the device abnormality agent for processing.

[0067] Processing process: The question input by the user includes entities: PF-300T type equipment, PECVD01 PM2, abnormal film thickness, which are used as keywords to search the knowledge graph.

[0068] Three possible causes of equipment abnormality related to the above information were retrieved, namely metal burr protrusion, RFStrap aging, and poor surface contact.

[0069] For these three possible reasons, we continue to search the knowledge graph and obtain the inspection actions corresponding to these three possible reasons: check the HF discharge circuit, check the bottom of the cavity, and check the edge of the cavity wall.

[0070] The big model is used to select the best one from the above three candidate inspection actions as the target inspection action, and the big model asks the user about the inspection result of the target inspection action.

[0071] Question about the large model: Are there any signs of aging on the bottom of the cavity? User feedback: No.

[0072] The device abnormality agent determines that the user feedback includes the inspection result of the inspection action and does not exceed the capability range of the device abnormality agent.

[0073] The device abnormality agent receives the inspection result from the user: no signs of aging appear on the bottom of the cavity. Based on the inspection result, it is determined that the real cause of the device abnormality is not: RF Strap aging. The next step is to reselect other inspection actions to ask questions to the user.

[0074] The remaining candidate inspection actions include: inspecting the HF discharge loop, inspecting the edge of the cavity wall contact.

[0075] The large model selects the best one from the above two candidate inspection actions as the target inspection action and asks the user about the inspection result of the target inspection action.

[0076] Question about the large model: Are there any metal burrs or protrusions in the HF discharge circuit? User feedback: Yes.

[0077] The device abnormality agent determines that the user feedback includes the inspection result of the inspection action and does not exceed the capability range of the device abnormality agent.

[0078] The device abnormality agent receives the inspection result from the user: metal burrs appear in the HF discharge circuit. Based on the inspection result, the real cause of the device abnormality is determined to be: metal burrs and tip discharge.

[0079] According to the real cause of the equipment abnormality, the monitoring parameters associated with the equipment abnormality are obtained through the path retrieval of the knowledge graph, and are fed back to the user together with the real cause of the equipment abnormality.

[0080] The above is the method for analyzing the cause of equipment anomaly and recommending monitoring parameters provided by the present invention. It can be seen that the method has the following beneficial effects: The device abnormality attribution analysis method provided by the present invention can quickly understand the user's intention through the user's questions, reduce the time for manual information entry and interpretation in traditional device abnormality analysis, and reduce the operating threshold of device abnormality analysis, so that the device abnormality analysis tracking does not have to rely on experienced engineers, thereby improving the overall device abnormality analysis speed and equipment maintenance efficiency, and reducing the manpower cost of equipment abnormality analysis. In addition, the method can guide the user to perform inspection actions and supplement missing information to obtain accurate equipment abnormality causes. The method performs equipment abnormality analysis inspection based on the historical abnormality causes and inspection action experience recorded in the knowledge graph, simplifying the equipment abnormality attribution analysis process.

[0081] The equipment abnormality attribution analysis method provided by the present invention can also recommend monitoring parameters according to the cause of the equipment abnormality, which can significantly improve the accuracy of fault diagnosis and prediction, and can achieve preventive maintenance, reduce maintenance costs, extend equipment life, and ensure production safety. It provides technical support for intelligent manufacturing, promotes the improvement of industrial automation and intelligence levels, and meets the needs of transformation and upgrading of the modern manufacturing industry.

[0082] The equipment anomaly attribution analysis method provided by the present invention is based on a large language model and an intelligent agent, has deep learning and generalization capabilities, can learn historical equipment anomaly data and its corresponding anomaly causes and inspection actions, and continuously optimize the knowledge graph and large model, thereby continuously improving the equipment anomaly attribution analysis capability.

[0083] On the other hand, the present invention also provides a device abnormality attribution system, the schematic diagram of which is shown in Figure 5 As shown, including: The intention recognition module 520 is used to recognize the user intention based on the question input by the user.

[0084] In some embodiments, the intention recognition module is specifically configured as follows: Leverage domain-wide models to identify private vocabulary in user-entered questions.

[0085] Based on the private vocabulary recognition results, the user input question is expanded through RAG document search, and the user intention is identified based on the expanded question.

[0086] The intention routing module 530 is used to classify the user intention, and when the classification result is an analysis of the cause of the device abnormality, assign the problem to the device abnormality agent for processing.

[0087] In some embodiments, the intent routing module is specifically configured as follows: The user intention and hierarchical dynamic few-shot prompt words are combined to form a query, which is sent to different agents. The agents judge the category of the query and give a vote.

[0088] The large model rewrites the query using different prompt words, repeats the voting multiple times, counts the average number of votes for each category, and determines the user intent classification based on the average number of votes.

[0089] The retrieval processing module 540 is used to call the device abnormality agent to process the question input by the user, specifically including: according to the question, searching the pre-built knowledge graph, and the retrieval results include the historical abnormality causes of the device abnormality in the question; If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is fed back to the user as the real cause of the device abnormality; If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the true cause of the equipment abnormality and feed it back to the user.

[0090] In some embodiments, the retrieval processing module is further used to: Determine whether the user feedback exceeds the capability range of the abnormal intelligent agent of the device.

[0091] If it is determined that the user feedback does not exceed the capability scope of the device abnormality intelligent agent, then based on the user feedback, determine whether the historical abnormality cause corresponding to the inspection action is the real cause of the device abnormality; if it is determined that the historical abnormality cause is not the real cause of the device abnormality, then reselect other inspection actions to ask questions to the user until the real cause of the device abnormality is determined.

[0092] If it is determined that the user feedback exceeds the capability range of the abnormal intelligent agent of the device, the user intention is re-identified based on the user feedback.

[0093] In some embodiments, asking the user questions based on the checking action specifically includes: The large model is used to select the target inspection action from multiple inspection actions.

[0094] The user is asked a question regarding an inspection result of the target inspection action.

[0095] In some embodiments, the system further includes a knowledge graph construction module 510 for constructing a knowledge graph. The knowledge graph construction module specifically includes: The ontology design submodule 511 is used to design the ontology of the equipment operation and maintenance experience of engineers by using the domain semantic network to obtain OWL semantic representation.

[0096] The instance extraction submodule 512 is used to extract knowledge graph instances from equipment analysis reports and device manuals based on the prompt words extracted from the OWL semantic representation and the domain big model.

[0097] The storage submodule 513 is used to store the OWL semantic representation and the instance on a knowledge graph platform, and the knowledge graph platform includes ES and a distributed database.

[0098] In some embodiments, the system further includes a monitoring parameter recommendation module for obtaining monitoring parameters associated with the device abnormality through path retrieval of the knowledge graph based on the actual cause of the device abnormality, and feeding it back to the user.

[0099] It should be noted that the above system can execute the above-mentioned device abnormality attribution method. For the functions of each module, please refer to the above-mentioned introduction to the method and will not be repeated here.

[0100] In the description of the embodiments of the present application, words such as "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0101] In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more.

[0102] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprises", "has" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0103] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A device abnormality attribution method, characterized in that: include: Identify user intent based on the questions they enter; Classify the user intention, and when the classification result is the analysis of the cause of the device abnormality, call the device abnormality agent to process the question input by the user, specifically including: according to the question, retrieve the pre-built knowledge graph, and the retrieval result includes the historical abnormality cause of the device abnormality in the question; If the search result includes only a single historical abnormality cause, the single historical abnormality cause is fed back to the user as the real cause of the device abnormality; If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the true cause of the equipment abnormality and feed it back to the user.

2. The method according to claim 1, characterized in that According to the user feedback, the real cause of the device abnormality is determined, specifically including: Determining whether the user feedback exceeds the capability range of the abnormal intelligent agent of the device; If the user feedback does not exceed the capability of the device abnormality agent, it is determined based on the user feedback whether the historical abnormality cause corresponding to the inspection action is the real cause of the device abnormality; if it is determined to be not, other inspection actions are reselected to ask questions to the user until the real cause of the device abnormality is determined.

3. The method according to claim 2, characterized in that After determining whether the user feedback exceeds the capability range of the abnormal device agent, the method further includes: If it is determined that the user feedback exceeds the capability range of the abnormal intelligent agent of the device, the user intention is re-identified based on the user feedback.

4. The method according to claim 1, characterized in that The pre-construction of the knowledge graph includes the following steps: The domain semantic network is used to design the ontology of engineers' equipment operation and maintenance experience, and the semantic representation of the network ontology language OWL is obtained; Based on the prompt words, the knowledge graph instance is extracted from the equipment analysis report and the device manual using the domain big model; the prompt words are extracted based on the OWL semantic representation; The OWL semantic representation and the instance are stored on a knowledge graph platform, which includes an elastic search engine and a distributed database.

5. The method according to claim 1, characterized in that The identifying the user intention according to the question input by the user specifically includes: Use the domain model to identify private words in user input questions; Based on the private vocabulary recognition results, the user input question is expanded through retrieval enhancement generation, and the user intention is identified based on the expanded question.

6. The method according to claim 1, characterized in that Classify the user intent, specifically including: Combine user intent and hierarchical dynamic small sample prompt words to form a query, which is sent to different agents. The agents judge the category of the query and give a vote. The large model rewrites the query using different prompt words, repeats the voting multiple times, counts the average number of votes for each category, and determines the user intent classification based on the average number of votes.

7. The method according to claim 1, characterized in that The checking action includes multiple steps, wherein the questioning of the user based on the checking action specifically includes: Use the large model to select the target inspection action from multiple inspection actions; The user is asked a question regarding an inspection result of the target inspection action.

8. The method according to claim 1, characterized in that: After determining the real cause of the device abnormality, the method further includes: according to the real cause of the device abnormality, obtaining monitoring parameters associated with the device abnormality through path retrieval of the knowledge graph, and feeding them back to the user.

9. A device abnormality attribution system, characterized in that: include: An intention recognition module is used to recognize the user's intention based on the question input by the user; An intention routing module is used to classify the user intention, and when the classification result is an analysis of the cause of the device abnormality, assign the problem to a device abnormality agent for processing; A retrieval processing module is used to call the device abnormality agent to process the question input by the user, specifically including: according to the question, searching the pre-built knowledge graph, and the search results include the historical abnormality causes of the device abnormality in the question; If the search result includes only a uniquely determined historical abnormality cause, the uniquely determined historical abnormality cause is fed back to the user as the real cause of the device abnormality; If the search result includes multiple possible historical abnormal causes, continue to search for the inspection action corresponding to each possible historical abnormal cause, ask the user questions based on the inspection action, and obtain user feedback; based on the user feedback, determine the true cause of the equipment abnormality and feed it back to the user.

10. The system according to claim 9, characterized in that The system also includes a knowledge graph construction module for constructing a knowledge graph, and the knowledge graph construction module specifically includes: The ontology design submodule is used to design the ontology of engineers' equipment operation and maintenance experience using the domain semantic network to obtain the semantic representation of the network ontology language OWL; An instance extraction submodule is used to extract knowledge graph instances from equipment analysis reports and device manuals based on prompt words combined with a domain big model; the prompt words are extracted based on the OWL semantic representation; The storage submodule is used to store the OWL semantic representation and the instance on a knowledge graph platform, wherein the knowledge graph platform includes an elastic search engine and a distributed database.

Images