New energy equipment troubleshooting method based on knowledge graph and large model and storage medium

Abstract

The invention discloses a new energy equipment troubleshooting method based on a knowledge graph and a large model. The method is used for solving the problems of fault positioning and troubleshooting of photovoltaic and wind power new energy equipment and the like. The method mainly comprises the following steps: performing data cleaning and preprocessing on a new energy equipment troubleshooting professional field document, and constructing a structured knowledge graph of equipment types, fault types, fault information and solution nodes in a graph database through Cypher; the big language model analyzes the user fault description, extracts and maps node attributes in the knowledge graph, and generates a graph query statement; the question-tracing type multi-round interaction is realized, the missing information is dynamically complemented, and the basic fault is accurately positioned in a step-by-step reasoning mode; according to the method, multi-modal resources such as texts, schematic diagrams and videos stored in the knowledge graph are combined, a step-by-step reasoning path and a solution with illustrative features are generated through a preset template, the whole reasoning process is visualized in a branch thinking graph form, and the accuracy, interpretability and user experience of troubleshooting are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, in particular to a new energy equipment fault troubleshooting method based on a knowledge graph and a large model and a storage medium. BACKGROUND

[0002] In the field of new energy, especially in the operation and maintenance of wind power and photovoltaic power generation, traditional manual item-by-item inspection and fault troubleshooting based on experience rules not only consume time and effort, but also have a sharp increase in cost and risk in remote or harsh environments. Moreover, the diagnosis results are easily influenced by individual experience, often resulting in misjudgment or missed judgment, leading to extended downtime, decreased efficiency, and even secondary damage.

[0003] A knowledge graph organizes devices, faults, and treatment schemes in a structured triple form, enabling the conversion of scattered fault cases, device manuals, and expert experience into a visual multi-level network. Through path deduction, it automatically associates relevant knowledge, enabling efficient diagnosis across devices, manufacturers, and models. However, question and answer systems relying solely on static graphs lack dynamic interaction capabilities, making it difficult to complete key information in multiple rounds of dialogue and unable to show the reasoning process to users, making it difficult to gain trust.

[0004] With the rapid development of large model technology, generative pre-training models such as DeepSeek have strong language understanding and generation capabilities and are widely used in intelligent question and answer, knowledge retrieval, and other scenarios. Through fine-tuning or prompting engineering, they can better understand natural language and generate fluent responses. However, relying solely on these models cannot meet the needs of structured and multi-step reasoning in new energy equipment fault troubleshooting. Large models lack explicit mapping of structured knowledge and are unable to show their internal reasoning logic to users, lacking traceability and interpretability. In actual operation and maintenance, users often provide incomplete information such as "the device does not generate electricity" or "inverter failure". Traditional question and answer systems cannot actively ask for additional details or show their internal basis, making it difficult to provide accurate and reliable solutions. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a new energy equipment fault troubleshooting method based on a knowledge graph and a large model. The system constructed by this method can solve the problems of single interaction, untraceable reasoning path, low solution credibility, and limited multi-hop reasoning capability of the existing new energy equipment fault troubleshooting question and answer system.

[0006] The technical solution of the present application specifically includes the following content.

[0007] New energy equipment troubleshooting knowledge graph construction: clean and preprocess the relevant new energy equipment troubleshooting professional field documents provided by business experts, use Neo4j graph database and Cypher statement to construct structured graph. The graph contains device type, fault type, fault information and solution nodes, and stores device name, manufacturer, model, fault code and other attributes, and stores online identifiers of fault pictures and operation videos in Markdown form, supporting multi-modal information retrieval.

[0008] Semantic analysis and dynamic questioning: use a large model to analyze the user's fault description, automatically extract device type, manufacturer, model, fault code and other key information and map it to the graph nodes. If the information is incomplete, generate follow-up questions through the conversation memory mechanism combined with the current Session ID to complete the missing attributes and ensure the integrity of the reasoning information.

[0009] Buffer pool ambiguity resolution: when there are nodes with the same name in the graph and belong to different reasoning paths, the system will temporarily store such nodes in the buffer pool, generate a Cypher query to locate the nearest branch node across the branch path, and generate a follow-up prompt based on the large model. After the user responds, filter out irrelevant nodes to determine the unique reasoning path.

[0010] Progressive reasoning and multi-modal answer generation: the system generates device positioning statements based on the completed information and extracts related fault information from the graph. The large model performs multi-round reasoning according to the "device positioning → fault phenomenon confirmation → associated fault deduction → root cause determination" logic, gradually converging to the root cause. Combined with multi-modal resources such as text, pictures and videos in the graph, generate troubleshooting answers with pictures and text.

[0011] Visualization and case pushing: build a branch thinking graph based on the reasoning path, show the complete process from the device node to the solution node, and store the session key information in the fault case library. Match the latest and most similar Top-k cases in the case library through the Jaccard similarity matching algorithm, and push them to the user through the visualization interface to improve the reference value of the solution and the user's trust.

[0012] (1) In the preferred scheme, the construction of the new energy equipment troubleshooting knowledge graph can be further refined as follows:

[0013] Obtain the original text from wind power / photovoltaic equipment manuals, operation logs, expert reports and industry standard documents, and simultaneously collect related pictures and videos; upload the pictures and videos to an online storage platform, generate Markdown format URLs, and establish an index mapping relationship with the corresponding explanatory text.

[0014] Based on the domain modeling objectives, define extraction rules for node types such as "device", "fault type", "treatment scheme", etc. Combine dependency syntax analysis and regular expressions to identify key information such as device name, model, fault code, and treatment steps from the text.

[0015] Use the context sequence and keyword prompts of devices, faults, and solutions in the original text to build a multi-level entity relationship architecture of "device → fault type → fault phenomenon → solution", and develop corresponding relationship extraction rules.

[0016] Generate a unique ID for each extracted entity, generate a "node attribute table" and a "relationship linking table", and import them as CSV files.

[0017] Import into Neo4j through the LOAD CSV statement of Cypher, use Neo4j Bloom for visual inspection, and manually review and verify the correctness of node attributes and relationships, and real-time correct the graph.

[0018] (2) In the preferred scheme, the semantic analysis and dynamic questioning can be further refined as follows:

[0019] Parse the user's natural language input through a large model, extract key information such as device type, manufacturer, model, and fault code, and map them to the corresponding nodes and attributes in the graph, and build an initial query context.

[0020] If necessary attributes (such as manufacturer, model, etc.) are missing, based on the current Session ID, pass the missing fields into the preset prompt template, and the large model generates targeted guided questioning to guide the user to supplement the information.

[0021] After the user supplements the information, the new attributes are mapped back to the knowledge graph and the dialogue state is updated synchronously, and the questioning and mapping process is repeated until the key information is complete.

[0022] According to the completed device information, automatically generate a Cypher statement to retrieve the fault information nodes associated with the device.

[0023] The large model dynamically generates questions based on the query results, guides the user to confirm the fault phenomenon, and maps it back to the graph to update the memory.

[0024] Through multiple rounds of interaction, gradually complete the fault information until the root cause of the fault is located.

[0025] (3) In the preferred scheme, the buffer pool ambiguity resolution can be further refined as follows:

[0026] When Cypher query finds multiple different branches of the same name or attribute repeated fault nodes, it is automatically identified as an ambiguity scenario.

[0027] The candidate node and its path information are stored in the buffer pool, and a cross-branch query statement is generated to locate the nearest neighbor branch node.

[0028] The large model automatically generates a Cypher query for the nearest neighbor branch node based on the large model, retrieves the next level of related fault sub-nodes of the branch node, and generates a new round of follow-up questions based on the query results to guide the user to confirm the specific fault information.

[0029] According to the user feedback, the irrelevant buffer pool nodes are removed, and only the paths consistent with the user feedback are retained to complete the ambiguity resolution.

[0030] (4) In the preferred scheme, the progressive reasoning and multi-modal answer generation can be further refined as follows:

[0031] The large model combines the completed "manufacturer", "model", "device name" and "fault phenomenon" attributes, automatically assembles and executes the graph query statement, and obtains the target phenomenon node and its next level of associated node list.

[0032] Based on the query results, the large model generates multiple rounds of guided follow-up questions according to the "device positioning → fault phenomenon confirmation → associated fault deduction → root cause determination" logic, guides the user to confirm the fault phenomenon, and updates the graph query context to gradually refine the reasoning.

[0033] When the reasoning converges to a unique root cause node, the system extracts the text, schematic diagram and operation video of the solution, combines the dialogue memory to generate a final answer with pictures and text.

[0034] (5) In the preferred scheme, the visualization and case pushing can be further refined as follows:

[0035] According to the determined reasoning link, a branch thinking diagram is drawn in real time to show the complete process of device → fault → root cause → solution.

[0036] The session ID, device information, fault node and solution key information are extracted, and a record is appended to the fault case library.

[0037] The Jaccard similarity matching algorithm is used to calculate the similarity between the current dialogue and the records in the case library, and the latest and most similar Top-k cases are selected and pushed to the user through the visualization interface.

[0038] The application also provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the new energy equipment fault troubleshooting method based on the knowledge graph and the large model. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a construction flowchart of the new energy fault troubleshooting knowledge graph in the application;

[0040] Figure 2 is a flowchart of constructing a question and answer system by using the method of the application;

[0041] Figure 3 is a schematic flowchart of fault troubleshooting knowledge retrieval in the application. DETAILED DESCRIPTION

[0042] The technical solutions of the application will be further described in detail below with reference to the embodiments.

[0043] With the rapid development of deep learning technology, large models have become the core tool in the field of intelligent question and answer due to their powerful natural language understanding and generation capabilities. They can parse the user's free text input into structured semantic expressions and generate answers. However, due to the limitations of training data and the "black box" characteristics of the model, it often produces "hallucination" problems in professional fields, i.e., giving coherent but inaccurate answers, especially in fields such as medicine, finance, and new energy fault maintenance, where accuracy requirements are extremely high, making it difficult to ensure the reliability and traceability of the answers, and lacking support for multi-round follow-up questions and gradual reasoning.

[0044] Knowledge graphs organize professional knowledge in the form of "entity-relation-entity" triples and can quickly locate related nodes according to conditions. However, when facing complex reasoning paths that span multiple intermediate entities, traditional graph retrieval is easily affected by path redundancy and difficult to capture complete information. At the same time, knowledge graphs themselves do not have flexible natural language understanding and multi-round interaction capabilities, making it difficult to actively initiate supplementary questions when user input is incomplete or ambiguous.

[0045] To overcome the above-mentioned defects, the application provides a new energy equipment fault troubleshooting method that combines knowledge graphs and large models. To address the pain points of low efficiency, information silos, and high safety risks in traditional wind power and photovoltaic operation and maintenance, this method proposes to integrate multi-source operation and maintenance documents with multi-modal resources to construct a structured fault troubleshooting knowledge graph. By combining the language understanding of large models with the reasoning ability of graphs, a six-step integrated intelligent system is formed: "semantic analysis → dynamic follow-up questions → ambiguity resolution → gradual reasoning → multi-modal answers → reasoning visualization and case pushing". This system not only addresses the "hallucination" risk of large models but also addresses the limitations of knowledge graphs in multi-hop reasoning, actively completing incomplete information and generating traceable solutions in the form of text and images, thereby improving the response speed, diagnostic accuracy, and user trust in new energy equipment operation and maintenance.

[0046] The technical solutions of the embodiments of the application will be further described below with reference to the accompanying drawings in the embodiments of the application.

[0047] APPENDIXFigure 1 The flowchart for constructing the new energy troubleshooting knowledge graph is as follows, and specifically includes the following key processes:

[0048] Text, tables, pictures and video resources are obtained from new energy operation knowledge documents such as device manuals, operation logs and expert reports. The text part is de-duplicated, segmented, segmented and segmented, and the terminology is standardized; pictures and videos are uploaded to the picture bed, and the embedded resources in the original text are replaced with Markdown format online links.

[0049] Combining dependency syntax analysis and rule method, extract "device name", "fault code", "fault phenomenon", "treatment steps" and other entities and attributes, and check with the field term library to improve the extraction accuracy and coverage.

[0050] Synonyms, ambiguous entities and conflicting information are manually reviewed and merged, and redundant information is removed to generate a structured "node attribute table".

[0051] According to the pre-defined structure of "device → fault → solution", a "relationship linking table" is generated to realize the semantic association between entities. The semantic-URL index of the document is established to facilitate the subsequent question and answer to realize the linkage display of text and pictures.

[0052] By analyzing the descriptive text corresponding to the image or video in the document, a mapping between Markdown URL and image / video resource is established to ensure that subsequent questions and answers can automatically call pictures / video resources according to the semantics, and realize the display effect of pictures and text.

[0053] The "node attribute table" and "relationship linking table" are exported as CSV files, and the example import statement is as follows:

[0054] LOAD CSV FROM'file: / / / device_nodes.csv'AS line

[0055] WITH line SKIP 1

[0056] CREATE(d:Device{id:line[0],name:line[1],manufacturer:line[2],model:line[3]});

[0057] Import Neo4j in batches through Cypher LOAD CSV to build a "device-fault-solution" multi-level graph structure. Finally, use Neo4j Bloom to visualize the correctness of the structure and content, and iteratively optimize the extraction logic with manual review to ensure the completeness and accuracy of the graph.

[0058] Appendix Figure 2A flowchart for constructing a troubleshooting question and answer system according to the present application, specifically comprising the following key steps:

[0059] Step 1, receiving user fault description:

[0060] Obtain the natural language description input by the user through the man-machine interaction interface, and extract the device type, manufacturer, model, fault phenomenon or code, etc.

[0061] Based on the pre-defined knowledge graph node label and attribute name, the input content is compared with the dictionary to determine whether it involves the new energy troubleshooting field;

[0062] If the relevant nodes of the knowledge graph are matched, the semantic analysis process is entered; otherwise, a pre-trained large model is called to answer generally, and a guide prompt is attached.

[0063] Step 2, preliminary semantic analysis:

[0064] Call the large model to perform deep semantic understanding on the received free text;

[0065] Automatically identify and extract key entities and their attribute values such as "device type", "manufacturer", "model", "fault code", "fault phenomenon", etc.

[0066] Map the extracted entity attributes to the corresponding nodes and relationships in the new energy troubleshooting knowledge graph to construct a preliminary question and answer context.

[0067] Step 3, attribute completeness verification:

[0068] The system maps the entity attributes extracted in step 2 to the knowledge graph and initiates a quick positioning query;

[0069] If the key attributes such as "manufacturer", "model" and "device name" are missing, go to step 4;

[0070] Otherwise, continue to step 5 for graph reasoning.

[0071] Step 4, guided attribute completion follow-up:

[0072] Based on the current session, inject all missing attribute names into the large model prompt template;

[0073] Automatically generate targeted completion follow-up questions to guide the user to provide missing attributes, such as "Please provide the manufacturer information of the device";

[0074] After receiving the user's supplement, update the dialogue memory synchronously and return to step 2 to re-perform semantic analysis and mapping.

[0075] Step 5, gradual multi-level reasoning and query:

[0076] The large model combines the completed entity attributes and the current question and answer context, assembles a single-hop or multi-hop Cypher query statement according to the prompt template, generates and executes the query statement in the format of "MATCH(d:Device

[0077] {…})-[:LINK]->(f:Fault)-[:LINK*0..n]->(x)RETURN x".

[0078] Step 6: Target triple and path list construction:

[0079] The large model reasons according to the prompt template and the question and answer context, and follows the hierarchical logic of "device positioning → fault phenomenon confirmation → associated fault deduction → root fault positioning".

[0080] According to the normal follow-up reasoning process of the method, a single-hop Cypher query statement is generated and executed each time the question and answer is generated, and the query statement is generated in the format of "MATCH(d:Device{…})-[:LINK]->(n)RETURN n" to obtain the next level node.

[0081] When the user provides multiple associated clues at one time or needs to cross multiple levels, a multi-hop Cypher query statement is generated, and the query statement is generated in the format of "MATCH(d:Device{…})-[:LINK]->(f:Fault)-[:LINK*0..n]->(x)RETURN x" to parallelly traverse multiple levels of relationships.

[0082] The Cypher execution result is parsed, and the entity triples and complete traversal paths that meet the conditions are extracted and stored in the "path list" for subsequent reasoning.

[0083] Step 7: ambiguity resolution:

[0084] When the same name or repeated attribute fault nodes appear in the path list, the system identifies it as an ambiguity scenario.

[0085] Backtracking to the branch node closest to the ambiguity node, and passing the branch node information into the large model prompt template to generate targeted follow-up questions.

[0086] For example, in the fan troubleshooting, the "current normal" node appears in both the "power low" and "abnormal vibration" branches, and the system locates to the "no fault code" branch node and asks "Please tell me if the power is low or if the vibration is abnormal?", to guide the user to clearly identify which branch is faulty.

[0087] According to the user's supplementary answer, the system filters out the branches that do not match, and only keeps the paths consistent with the user's situation.

[0088] Step 8, step-by-step reasoning and answer generation:

[0089] According to the four-level logic of "device positioning → fault phenomenon confirmation → associated fault deduction → root fault determination", the large model combines the path list and the original question semantics to generate multiple rounds of guided follow-up questions or conclusive answers;

[0090] In each round of question and answer, the node judgment is performed on the triples obtained by knowledge retrieval to determine whether the triples contain a solution node;

[0091] If it contains, the solution prompt template is called to generate a conclusive answer by combining user semantic information and target triples;

[0092] If it does not contain, the large model follow-up question prompt template is called to generate the next round of guided follow-up questions;

[0093] When generating guided follow-up questions, the system not only outputs natural language prompts, but also generates "picture-guided follow-up questions" by combining the pictures associated with the fault nodes;

[0094] For example, when it is necessary to confirm whether the fan yawing system is abnormal, the system synchronously displays a typical abnormality diagram and prompts "Have you observed such a phenomenon?", helping users to intuitively judge;

[0095] The node set of the "path list" in this round and the multi-modal resource index are recorded in the dialogue memory as the basis for subsequent visual reasoning and case pushing.

[0096] Step 9, case extraction and pushing:

[0097] Determine whether a conclusive answer has been generated in this round of question and answer;

[0098] If a conclusive answer is generated, the SessionID, device information, fault node, and solution are extracted according to the case extraction template and stored in the fault case library;

[0099] If a follow-up question answer is generated, the current session key information is extracted and similarity matching calculation is performed with each historical case in the fault case library.

[0100] In order to balance the matching efficiency and implementation complexity, Jaccard similarity is used to quickly match the "path list" and "case node list" in the fault case pushing stage. The specific calculation is as follows:

[0101]

[0102] Where Q is the path list node set extracted in the current dialogue, C is the case node set corresponding to a historical fault case in the library, |Q∩C| and |Q∪C| represent the intersection and union sizes of the two sets, respectively;

[0103] Sort the results from high to low, and select the latest Top-k cases.

[0104] Step 10, visual reasoning path display:

[0105] When generating a conclusion answer, the visualization engine reads the "path list" node set saved in the dialogue memory and renders it as a tree-like branch structure diagram, showing the complete reasoning chain from the specified device to the related fault and finally to the solution.

[0106] Through intuitive graphical reasoning paths, users can trace back and review each step of the large model's reasoning logic in the knowledge graph, significantly enhancing the transparency and credibility of the question and answer results.

[0107] The above process not only supports multi-hop structured retrieval and accurate positioning of business experts, but also combines the natural language interaction capabilities of large models to achieve active completion and step-by-step reasoning of user input, ultimately outputting traceable, visual, and reliable troubleshooting answers.

[0108] Appendix Figure 3 A schematic flowchart for a segment of troubleshooting knowledge retrieval, which specifically includes the following content:

[0109] First, the user raises the issue of "the photovoltaic inverter equipment of the manufacturer Sunlight and model SG320HX has stopped working", and the system calls the large model to analyze the semantics and locate the "SG320HX" device node.

[0110] According to the device node and the graph structure, two types of fault nodes, "fault code" and "no fault code", are retrieved.

[0111] The large model generates initial follow-up information, asking the user: "Has the inverter equipment appeared relevant fault codes, such as 1276, 537, etc.?"

[0112] The user replies, "There is no fault code information, but the device cannot start."

[0113] The large model retrieves the relevant fault type nodes of "no fault code" under the device in the graph, including "output power constant" and "low power" fault information nodes.

[0114] The large model proposes a new round of questioning: "Since your inverter does not display fault codes, further investigation of other fault phenomena is needed. Please check if the inverter's power generation and generation hours are low. Observe the two refresh cycles to determine if the power is low or the output power is constant? (see the figure below) so that we can accurately locate the fault problem and provide a solution." Through the automatic indexing of descriptive text and the embedding of corresponding illustrations, users can intuitively identify the problem;

[0115] The user replies: "I observed that the device information table of the inverter device shows that the current is 0. What should I do now?" Based on the context and the node information of the knowledge graph, the fault node "actual current is 0" is located, and the troubleshooting path to this node is returned to the large model.

[0116] After multiple rounds of troubleshooting, the "MC4 plug contact failure" is finally locked as the root cause, and the large model gives a conclusion: try replacing the MC4 plug and push the relevant operation demonstration video.

[0117] Throughout the process, each round of questioning and retrieval uses the structured triple of the knowledge graph for accurate positioning, combined with the natural language interaction ability of the large model, to finally realize the closed-loop troubleshooting process from device shutdown description to root cause positioning to solution output.

Claims

1. A new energy equipment troubleshooting method based on knowledge graph and large model, comprising new energy equipment troubleshooting knowledge graph construction, semantic analysis and dynamic questioning, buffer pool ambiguity resolution, progressive reasoning and multi-modal answer generation, and visualization and case pushing; characterized in that: the buffer pool ambiguity resolution comprises: when the same name fault nodes appear in the graph and belong to different reasoning paths, the system temporarily stores such nodes in the buffer pool, generates a Cypher query to locate the nearest branch node across the branch path, combines the questioning prompts generated by the large model, filters irrelevant nodes after the user responds, and determines the unique reasoning path; the progressive reasoning and multi-modal answer generation comprises: the system generates equipment positioning statements according to the completion information and extracts associated fault information from the graph; the large model performs multi-round reasoning according to the "equipment positioning → fault phenomenon confirmation → associated fault deduction → root cause determination" logic, and gradually converges to the root cause; combined with the text, picture and video resources in the graph, a troubleshooting answer with text and pictures is generated; the visualization and case pushing comprises: a branch thinking graph is constructed according to the reasoning path, the complete process from the equipment node to the solution node is displayed, and the session key information is extracted and stored in the fault case library; the latest and most similar Top-k cases are selected by matching with the historical cases in the case library through the Jaccard similarity matching algorithm, and pushed to the user through the visualization interface.

2. The method of claim 1, wherein the construction of the new energy equipment troubleshooting knowledge graph comprises: (1) obtaining original text from wind power / photovoltaic equipment manuals, operation and maintenance logs, expert reports and industry standard documents, and synchronously collecting related pictures and videos; uploading the pictures and videos to an online storage platform, generating Markdown format URLs, and establishing index mapping relationship with the corresponding explanation text; (2) based on the domain modeling target, defining the extraction rules of equipment, fault type and processing scheme node types, combining dependency syntax analysis and regular expressions to identify equipment name, model, fault code and processing steps from the text; (3) using the context sequence and keyword prompts of equipment, fault and scheme in the original text, constructing a multi-level entity relationship architecture of equipment → fault type → fault phenomenon → solution, and formulating the corresponding relationship extraction rules; (4) generating a unique ID for each extracted entity, generating a node attribute table and a relationship link table, and importing as a CSV file; (5) importing into Neo4j through the LOAD CSV statement of Cypher, using Neo4j Bloom for visualization inspection, and manually auditing and checking the correctness of node attributes and relationships, and correcting the graph in real time.

3. The method of claim 1, wherein the semantic analysis and dynamic questioning comprises: analyzing the user's natural language input through the large model, extracting key information such as equipment type, manufacturer, model and fault code, and mapping it to the corresponding nodes and attributes in the graph to build an initial query context; ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ If the necessary attributes are missing, based on the current Session ID, the missing fields are passed into the preset prompt template, and the large model generates targeted guided follow-up questions to guide the user to supplement information; After the user supplements the information, the new attributes are mapped back to the knowledge graph and the dialogue state is updated synchronously, and the follow-up question and mapping process are executed in a loop until the key information is complete; According to the completed device information, a Cypher statement is automatically generated to retrieve the fault information nodes associated with the device; The large model dynamically generates follow-up questions based on the query results, guides the user to confirm the fault phenomenon, and maps it back to the graph to update the memory; Through multiple rounds of interaction, the fault information is gradually completed until the root cause of the fault is located.

4. The method of claim 1, wherein the buffer pool ambiguity resolution step comprises: When the Cypher query finds multiple different branches of the same name or attribute repeated fault nodes, it is automatically identified as an ambiguity scenario; Store the candidate nodes and their path information into the buffer pool, generate a cross-branch query statement to locate the nearest neighbor branch node; The large model automatically generates a Cypher query for the nearest neighbor branch node based on the nearest neighbor branch node, retrieves the next level of related fault sub-nodes of the branch node, and generates a new round of follow-up questions based on the query results, guiding the user to confirm the specific fault information; According to the user feedback, eliminate irrelevant buffer pool nodes, and only keep the path consistent with the user feedback to complete the ambiguity resolution.

5. The method of claim 1, wherein the progressive reasoning and multi-modal answer generation step comprises: The large model combines the completed manufacturer, model, device name and fault phenomenon attributes to automatically assemble and execute graph query statements to obtain the target phenomenon node and its next level of associated node list; Based on the query results, the large model generates multiple rounds of guided follow-up questions according to the logic of device positioning → fault phenomenon confirmation → associated fault deduction → root cause determination, guiding the user to confirm the fault phenomenon and updating the graph query context, gradually refining the reasoning; When the reasoning converges to a unique root fault node, the system extracts the text, schematic diagram and operation video of the solution, combines the dialogue memory to generate a final answer with text and graphics.

6. The method of claim 1, wherein the visualization and case pushing step comprises: According to the determined reasoning link, a branch thinking diagram is drawn in real time to show the complete process of device → fault → root cause → solution; Extract the session ID, device information, fault node and solution key information to append records to the fault case library; Use the Jaccard similarity matching algorithm to calculate the similarity between the current dialogue and the records in the case library, select the latest and most similar Top-k cases, and push them to the user through the visualization interface.

7. A non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the method of any one of claims 1-6.

Images