Intelligent hoisting machinery fault questioning and answering method based on knowledge graph and large model fusion and application of intelligent hoisting machinery fault questioning and answering method

Abstract

The invention discloses a hoisting machinery fault intelligent question-answering method based on knowledge graph and large model fusion and application thereof.The method comprises the steps that firstly, a hoisting machinery fault knowledge graph containing equipment types, parts, fault modes, symptoms, working conditions and other nodes is constructed, and instructions, standards, maintenance records and fault cases are written into the graph through text extraction; the method comprises the following steps: aiming at a natural language problem and operation data, generating a graph query instruction containing equipment, parts, symptoms and working condition constraints by a large language model subjected to field fine tuning, and executing query and mode matching on a knowledge graph to obtain a plurality of candidate fault interpretation sub-graphs; then inputting the candidate subgraph abstract, the question and the working condition abstract into a large language model together to generate a diagnosis answer, analyzing the answer into a structured triple, performing consistency verification with a knowledge graph, triggering rewriting under constraint for the conflicting answer, and finally obtaining the diagnosis result. According to the scheme, the accuracy, structural rationality and interpretability of crane mechanical fault questions and answers can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crane safe operation diagnosis and large language model application, in particular to a crane fault intelligent question and answer method based on knowledge graph and large model fusion and application thereof. BACKGROUND

[0002] With the wide application of tower cranes, gantry cranes and other hoisting machinery in construction engineering, port loading and unloading, industrial manufacturing and other scenes, the mechanical structure of the hoisting machinery is more and more complex, the working environment is variable, and the fault types are diversified and hidden. Once the mechanical fault cannot be identified and handled in time, it is easy to cause long-term operation with disease, unplanned downtime, and even cause major safety accidents. Therefore, how to provide convenient and reliable fault question and answer and diagnosis assistance tools for front-line operators and maintenance personnel is an important technical demand in the field of crane operation and maintenance.

[0003] In the prior art, crane mechanical fault diagnosis and question and answer mainly exist in the following several ways: (1) Traditional method based on artificial experience and document search In most application scenarios, operators rely on personal experience, on-site oral communication, telephone consultation with experts, and search of equipment manuals, maintenance procedures, accident case compilations and other documents to judge fault causes and treatment measures. This method has the problems of scattered knowledge, low search efficiency, dependence on personnel experience, unstable diagnosis results, and difficulty in timely comprehensive analysis of historical operation data and alarm events for faults.

[0004] (2) Expert system based on rule base or simple model Some systems pre-construct a "symptom-fault-treatment measure" rule base, combine limited logical reasoning or statistical models, and realize semi-automatic diagnosis of crane faults. However, this kind of system usually covers a limited number of symptoms, and is insufficient to describe complex working conditions and multiple fault coupling conditions; the rule base has high maintenance cost, and is difficult to dynamically expand with device model updates and on-site experience accumulation; at the same time, the rules usually lack explicit association with device structure, working condition, and event sequence, resulting in limited applicability and explainability of the diagnosis results.

[0005] (3) Mechanical fault question and answer technology based on knowledge graph With the development of knowledge graph technology, existing solutions attempt to extract equipment structure, failure mode, and maintenance measures in the mechanical field as entities and relationships to form a graph database, and provide question and answer services for users through graph query or pattern matching. However, existing knowledge graph question and answer services are mostly based on "keyword matching + fixed query template", which has certain limitations. On the one hand, the understanding ability of natural language questions is limited, and it is difficult to accurately identify the implied intent, constraint conditions and time sequence information in the question. On the other hand, the graph query logic is mostly artificially predefined, lacking deep integration with actual operation data and alarm events, making it difficult to provide personalized diagnosis for "this crane at the current time, this time period, and this set of working conditions".

[0006] (4) General question answering or retrieval enhanced question answering technology based on large language model In recent years, large language models have shown strong natural language understanding and generation capabilities in general question answering and retrieval enhanced generation. Some studies attempt to directly use large models for mechanical fault question and answer, or combine document retrieval to provide context. Although such methods improve the natural language interaction experience, they have the following shortcomings: (1) Large models focus on text semantic reasoning, and lack explicit modeling of structured knowledge such as crane equipment structure, failure mechanism, and working condition constraints, which can easily produce "illusion" answers that are inconsistent with the real structure and mechanism; (2) Retrieval enhancement mainly targets text snippets, and does not fully utilize the entity relationships and multi-hop paths contained in the knowledge graph, making it difficult to form a verifiable "explanation subgraph"; (3) There is a lack of mechanism to check the consistency of the diagnosis conclusion output by the large model with the knowledge graph, and it is difficult to effectively filter or correct answers that do not conform to the device structure constraints and contradict existing knowledge.

[0007] In summary, the existing crane mechanical fault question and answer technology has the following common problems: 1. Loose coupling between natural language question answering and knowledge graph, with large models and graphs often fighting separately, lacking a deep integration mechanism based on unified tasks; 2. Lack of automated bridging between question analysis and graph query, with graph query templates and parameters mostly relying on manual design, making it difficult to cover complex and variable field problems; 3. Diagnosis conclusions usually lack structured consistency checking based on knowledge graphs, and cannot constrain and correct possible errors in large model reasoning; 4. It is difficult to form a "graph reasoning path" and a "large model semantic path" for dual-path collaboration, resulting in limited accuracy, stability, and explainability of fault question and answer.

[0008] Therefore, there is an urgent need for a hoisting machinery fault intelligent question and answer method that deeply integrates a knowledge graph and a large language model, so that a natural language question can be automatically converted into a graph query task, candidate explanation subgraphs are reasoned using a graph, and a large model is used to comprehensively explain and check results through a knowledge graph, thereby improving the accuracy and reliability of hoisting machinery fault question and answer. SUMMARY

[0009] Therefore, the present application aims to provide a hoisting machinery fault intelligent question and answer method based on the integration of a knowledge graph and a large model, which is reliable to implement, flexible to apply, and accurate and reliable in question and answer.

[0010] In order to achieve the above technical purpose, the technical scheme adopted by the present application is as follows: A hoisting machinery fault intelligent question and answer method based on the integration of a knowledge graph and a large model, comprising: S01, constructing a hoisting machinery fault knowledge graph, which comprises a plurality of nodes and relationship edges connecting the nodes, and also extracting entities and relationships related to hoisting machinery information, and writing the extracted entities and relationships into the hoisting machinery fault knowledge graph; S02, receiving an operation personnel input question and answer request about hoisting machinery fault, converting the question and answer request into a natural language question, and obtaining the running data of the hoisting machinery related to the question and answer request, and converting it into a running data summary; S03, inputting the natural language question and running data summary into a large language model fine-tuned by hoisting machinery field corpus, and outputting a structured instruction including task type identification and graph query template from the large language model, wherein the graph query template contains a pattern description of event time sequence; S04, according to the graph query template, performing entity matching and multi-hop relationship search on the hoisting machinery fault knowledge graph, and combining the pattern description of event time sequence to perform pattern matching of events-faults, obtaining a plurality of candidate explanation subgraphs, and then scoring and sorting the candidate explanation subgraphs according to a preset condition; S05, converting the natural language question, running data summary and sorted candidate explanation subgraph into a fusion context in the form of a graph summary, and inputting the fusion context into the large language model, and generating a hoisting machinery fault diagnosis answer from the large language model under the constraint of the fusion context, wherein the diagnosis answer at least includes a main fault mode and optional alternative fault modes, corresponding reason explanation, and recommended inspection steps and repair scheme.

[0011] As a possible implementation, further, in the scheme S01, the plurality of nodes of the hoisting machinery fault knowledge graph comprises a plurality of nodes of device type, configuration and component, fault mode, symptom, working condition, inspection item, maintenance scheme, and event and state.

[0012] The relationship edge is formed by one or more above-mentioned corresponding relationships in the hoisting machinery fault knowledge graph, including "device-component", "component-fault mode", "fault mode-symptom", "working condition-fault mode", "fault mode-inspection item", "fault mode-maintenance scheme", and "event-fault mode".

[0013] In addition, the hoisting machinery related information is crane manual, standard specification, maintenance record, fault case, and expert experience.

[0014] As a preferred implementation option, preferably, in the scheme S01, the hoisting machinery fault knowledge graph adopts a hierarchical modeling manner, wherein the device type node is located at the device layer, the configuration and component node is located at the structure layer, the fault mode node is located at the fault layer, the symptom node is located at the symptom layer, the working condition node is located at the working condition layer, the inspection item node and the maintenance scheme node are located at the disposal layer, and the event and state node is located at the time event layer; the event and state node comprises an alarm event node, a shutdown event node, and a sensor out-of-limit event node, and is connected with the corresponding fault mode node through the event-fault mode relationship edge, for supporting event-fault mode matching based on time sequence.

[0015] In addition, in the scheme S02, the running data is obtained by information extraction on the question and answer request, or is obtained from the crane control system.

[0016] In addition to the above, in the scheme S02, the running data comprises load, amplitude, lifting height, operation mode, wind speed, alarm record, and statistical characteristics of part of key sensors of the hoisting machinery in the current or recent period of time; and the running data summary is in the form of natural language description of the running data.

[0017] As a preferred implementation option, preferably, in the scheme S03, the graph query template is used to define target device type, candidate component, candidate symptom, working condition constraint, and expected output fault mode and processing suggestion.

[0018] The graph query template includes multiple information slots to be filled, including at least a target equipment type slot, a target component slot, a target symptom slot, a working condition constraint slot, and an output item slot. The output item slot is used to indicate the query return fault mode, the corresponding evidence node, and the recommended inspection and maintenance nodes. The large language model fills the information in each slot according to the description in the natural language question and the running data summary, generating a graph query statement containing specific entities or limiting conditions.

[0019] As a preferred implementation option, preferably, in this scheme S04, each candidate interpretation subgraph includes at least one suspected fault mode node and its connected component nodes, symptom nodes, operating condition nodes, event and status nodes, inspection item nodes, and maintenance plan nodes.

[0020] As a preferred implementation option, in this scheme S04, the candidate interpretation subgraphs are scored and ranked according to the degree of matching between each candidate interpretation subgraph and the current operating data and observed symptoms, as well as the reliability of the knowledge source.

[0021] The scoring of each candidate explanatory subgraph should consider at least the following factors: (1) The degree of matching between the symptom nodes in the candidate explanatory subgraph and the symptom expressions mentioned in the natural language problem; (2) The consistency between the working condition nodes in the candidate interpretation subgraph and the working condition features in the current operating data; (3) The reliability level of the source of the knowledge entries involved in the candidate explanation subgraph; (4) The structural complexity and path length of the candidate interpretation subgraph; The comprehensive score of the candidate explanatory subgraph is obtained by weighted summation of one or more of the above factors.

[0022] As a preferred implementation option, this solution also includes: S06. Extract information from the diagnostic response generated by the large language model and parse it into a set of structured triples that include one or more of the following relationships: "equipment-component-fault mode", "fault mode-symptom", "fault mode-operating condition", "fault mode-inspection item", and "fault mode-maintenance plan". Compare the set of triples with the knowledge graph of crane machinery faults.

[0023] If all key relationships in the triplet set have connected paths in the knowledge graph of crane machinery faults and do not contradict the equipment structure or fault mechanism, then the diagnostic answer will be output as the final question and answer result. If a significant conflict is detected with the knowledge graph of crane machinery faults, the conflict information and the relevant graph node summary are input into the large language model to trigger a rewrite of the diagnostic response. If the conflict cannot be eliminated after a preset number of rewrites, a conservative diagnostic result with uncertainty prompts is output.

[0024] As a preferred implementation option, in S06 of this scheme, when extracting information from the diagnostic response, a large language model or sequence labeling model is used to parse the natural language response into a set of structured triples. Each triple in the set of triples includes at least a subject entity, a relation type, and an object entity. The subject entity and the object entity correspond to the equipment type node, configuration and component node, fault mode node, symptom node, operating condition node, inspection item node, or maintenance plan node in the knowledge graph of crane machinery faults, respectively. The relation type corresponds to the relation edge type in the knowledge graph of crane machinery faults.

[0025] As a preferred implementation option, in S06 of this scheme, when comparing the set of triples with the knowledge graph of crane machinery faults, the system queries the knowledge graph of crane machinery faults to see if there is a predefined type of connection path between the nodes corresponding to each triple in the set of triples. When a triple does not have any connection path in the knowledge graph of crane machinery faults that satisfies the equipment structure constraints and fault mechanism constraints, it is marked as a conflicting triple. If a conflicting triple exists, a conflict description containing the conflicting triple and related node information is generated and input into the large language model again along with the original natural language question and the runtime data summary to trigger the large language model to rewrite the diagnostic answer under the conflict constraints.

[0026] Based on the above, this solution also proposes an intelligent question-answering system for crane machinery faults based on the fusion of knowledge graphs and large models, which includes: The graph construction unit is used to construct a knowledge graph of crane machinery faults, which includes multiple nodes and relation edges connecting the nodes. It also extracts entities and relations related to crane machinery and writes the extracted entities and relations into the knowledge graph of crane machinery faults. The question collection unit is used to receive question and answer requests about crane malfunctions input by operators, convert the question and answer requests into natural language questions, and at the same time acquire the crane's operating data related to the question and answer requests and convert it into an operating data summary. The model loading unit loads a large language model that has been fine-tuned and trained on a crane domain corpus and is used to provide question-and-answer information processing. The information scheduling unit is used to input the natural language problem and the summary of operation data into a large language model fine-tuned by the crane domain corpus, and the large language model outputs structured instructions including task type identifiers and graph query templates, wherein the graph query templates contain pattern descriptions of the time sequence of events. The information processing unit is used to perform entity matching and multi-hop relationship search on the knowledge graph of crane machinery faults according to the graph query template, and to perform event-fault pattern matching in combination with the pattern description of event time sequence to obtain several candidate explanation subgraphs, and then to score and sort the candidate explanation subgraphs according to preset conditions. The answer generation module converts the natural language question, the runtime data summary, and the sorted candidate explanation subgraphs into a fusion context in graph summary form and inputs it into the large language model. The large language model generates a crane mechanical fault diagnosis answer under the constraints of the fusion context. The diagnosis answer includes at least one main fault mode and optional alternative fault modes, corresponding explanations, and suggested inspection steps and maintenance plans. The model optimization module is used to package the confirmed fault results from actual maintenance with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. Based on multiple question-and-answer examples, it updates the edge weights and adds new nodes and relationships in the crane machinery fault knowledge graph. At the same time, it extracts samples of "question-graph query-candidate explanation subgraph-diagnostic answer" from the question-and-answer examples to incrementally fine-tune the large language model or use them as dialogue context examples to improve its ability to generate graph query templates and diagnostic answers in the future.

[0027] Based on the above, this solution also proposes a method for querying crane machinery faults, which applies the aforementioned intelligent question-and-answer method for crane machinery faults based on the fusion of knowledge graphs and large models; it also includes: Based on the fault results confirmed in actual maintenance, these results are packaged with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. The edge weights, newly added nodes, and relationships in the crane machinery fault knowledge graph are updated based on multiple question-and-answer examples. At the same time, samples of "question-graph query-candidate explanation subgraph-diagnostic answer" are extracted from the question-and-answer examples to incrementally fine-tune the large language model or to serve as dialogue context examples, thereby improving its ability to generate graph query templates and diagnostic answers in the future.

[0028] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This solution bridges the gap between natural language questions and knowledge graph queries. It utilizes a large language model refined from crane scenario corpora and literature to convert operators' natural language questions and operational data summaries into structured graph query templates and instructions. These are then filled with information slots such as equipment type, related components, symptom descriptions, operating condition constraints, and output requirements, eliminating the need for manually writing numerous fixed query templates. This approach can adapt to diverse and complex natural language questioning scenarios, transforming the knowledge graph from a "passive query" to an "active collaboration with the large model for understanding," significantly improving the system's flexibility and scalability.

[0029] In addition, this solution constructs a candidate explanation subgraph based on knowledge graph reasoning to improve the structural rationality of the diagnosis. Through graph query templates, multi-hop relationship search and pattern matching are performed on the crane machinery fault knowledge graph to generate a candidate explanation subgraph centered on the fault mode and including nodes such as components, symptoms, operating conditions, events and states, inspection items, and maintenance plans. The candidate explanation subgraph is then comprehensively scored based on one of the following factors: symptoms, operating conditions, reliability of knowledge sources, and path length. Compared with existing solutions that rely solely on text retrieval or pure model reasoning, this approach can define "which fault explanations are well-founded in the graph" from a structured knowledge level, significantly enhancing the rationality of the diagnostic conclusions regarding equipment structure and fault mechanisms.

[0030] This solution employs a dual-path question-answering approach combining graph summarization and large-scale model fusion to improve interpretability and human-computer interaction quality. Candidate explanatory subgraphs are compressed into structured graph summaries containing fault modes, key symptoms, operating conditions, and recommended remedial measures. These summaries, along with natural language questions and operating condition summaries, are then input into a large language model, which generates natural language diagnostic answers under graph constraints. When answering, the large model can explicitly reference the relationship chain of "component-fault-symptom-operating condition," explaining the reasoning basis to the user. Compared to a simple "black box output" model, this approach offers better interpretability and professional credibility.

[0031] Building upon the above, this solution further introduces a knowledge graph-based consistency verification and rewriting mechanism to enhance the reliability and robustness of the results. After the large model generates diagnostic answers, information extraction is used to parse the answer content into structured triples such as "equipment-component-fault-symptom-operating condition-inspection-maintenance," and then performs connectivity and mechanistic constraint checks with the knowledge graph, marking relationships that do not conform to the graph structure or are obviously contradictory. When overall consistency is insufficient, conflict information is fed back to the large model in the form of a prompt, triggering it to rewrite the answer under the constraints of the knowledge graph. Compared to traditional solutions that rely solely on the output of the large model, this invention performs "secondary verification" of the answer through the knowledge graph, effectively reducing the risk of the large model generating erroneous reasoning or illusory answers, and improving the reliability and robustness of the system in safety-critical scenarios.

[0032] This solution also supports bidirectional updates based on question-and-answer examples, possessing continuous evolution capabilities. By forming complete question-and-answer examples from "natural language question - runtime data summary - graph query - candidate explanation subgraph - diagnostic answer - actual maintenance result," these examples are used to update entity and relation weights in the knowledge graph and serve as training or suggestion samples for "large model generating graph queries, understanding graph summaries, and outputting structured diagnoses." Through continuous accumulation of real-world cases, this invention achieves bidirectional enhancement of the knowledge graph and the large model, enabling the system to continuously optimize as equipment models evolve, operating conditions change, and field experience increases. It possesses significant engineering deployment value and long-term maintainability.

[0033] In summary, this solution constructs a fusion architecture of "knowledge graph reasoning path + large model semantic path + consistency verification loop". While maintaining a natural language-friendly interaction, it introduces explicit structured knowledge constraints and result verification mechanisms, which can significantly improve the accuracy, stability and interpretability of intelligent question answering for crane mechanical faults. It overcomes the shortcomings of existing technologies where knowledge graphs and large models are used in isolation and lack deep integration and mutual constraints. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0035] Figure 1 This is a schematic diagram of one implementation process of the question-and-answer method in this solution; Figure 2 This is a brief illustration of the layered modeling method used in the knowledge graph of crane machinery faults in this solution, showing the location of different nodes in each layer. Figure 3 This is a schematic diagram of another implementation process for the question-and-answer method in this solution; Figure 4 This is a schematic diagram showing the connection of the unit modules of the question-and-answer system in this solution. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Combination Figure 1 As shown in the figure, this embodiment presents an intelligent question-answering method for crane machinery faults based on the fusion of knowledge graphs and large models, which includes: S01. Construct a knowledge graph of crane machinery faults, which includes multiple nodes and relational edges connecting the nodes. Also, extract entities and relations related to crane machinery, and write the extracted entities and relations into the knowledge graph of crane machinery faults. S02. Receive the question and answer request about the crane malfunction input by the operator, convert the question and answer request into a natural language question, and at the same time obtain the crane operation data related to the question and answer request and convert it into an operation data summary. S03. Input the natural language question and the summary of the operation data into a large language model fine-tuned by the crane domain corpus. The large language model outputs a structured instruction including a task type identifier and a graph query template. The graph query template contains a pattern description of the time sequence of events. S04. Based on the graph query template, perform entity matching and multi-hop relationship search on the knowledge graph of crane machinery faults, and combine the event-fault pattern matching with the pattern description of the event time sequence to obtain several candidate explanation subgraphs. Then, score and sort the candidate explanation subgraphs according to preset conditions. S05. The natural language question, the running data summary, and the sorted candidate explanation subgraph are converted into a fusion context in graph summary form and input into the large language model. The large language model generates a crane mechanical fault diagnosis answer under the constraints of the fusion context. The diagnosis answer includes at least one main fault mode and optional alternative fault modes, corresponding explanations, and suggested inspection steps and maintenance plans.

[0038] This solution bridges the gap between natural language questions and knowledge graph queries. Utilizing a large language model refined from crane scenario corpora and literature, it transforms operators' natural language questions and operational data summaries into structured graph query templates and instructions. These are then populated with information slots for equipment type, related components, symptom descriptions, operating constraints, and output requirements, eliminating the need for manually writing numerous fixed query templates. This approach adapts to diverse and complex natural language questioning scenarios, transforming the knowledge graph from a "passive query" to an "active collaboration with the large model for understanding," significantly improving the system's flexibility and scalability.

[0039] In terms of the specific details of the knowledge graph of crane machinery faults, as a possible implementation, in this solution S01, the multiple nodes of the knowledge graph of crane machinery faults include multiple nodes of equipment type, configuration and component, fault mode, symptom, operating condition, inspection item, maintenance plan, and event and status.

[0040] The relationship edges are formed by connecting nodes in the crane machinery fault knowledge graph through one or more of the following corresponding relationships: "equipment-component", "component-fault mode", "fault mode-symptom", "operating condition-fault mode", "fault mode-inspection item", "fault mode-maintenance plan", and "event-fault mode".

[0041] As an example, the function definition of the hoisting machinery fault knowledge graph is as follows: ; in, It combines entity nodes, such as: equipment type nodes, configuration and component nodes, fault mode nodes, symptom nodes, operating condition nodes, inspection item nodes, maintenance plan nodes, and one or more of event and status nodes; This refers to a set of relation types, such as one or more of the following correspondences: "Equipment-Component", "Component-Failure Mode", "Failure Mode-Symptom", "Operating Condition-Failure Mode", "Failure Mode-Inspection Item", "Failure Mode-Maintenance Plan", and "Event-Failure Mode". It is a set of directed edges in the form of ternary components.

[0042] In addition, regarding the extraction of entity and relational information related to lifting machinery, the lifting machinery-related information mentioned in this solution includes crane manuals, standards and specifications, maintenance records, failure cases, and expert experience.

[0043] In terms of information extraction, as an example, consider a piece of information text related to lifting machinery. Entities and relationships can be obtained through an information extraction model, the function of which is defined as follows:

[0044] in, For entity recognition models, For information text The entity set extracted from the middle, For relation extraction models, For information text The extracted triple set is used. The entity recognition model and relation extraction model for the text can directly use existing technical strategies, which will not be elaborated further here. Furthermore, as a further implementation example, the extracted triple set can be further evaluated for confidence level, and then the entities and relations pointed to by the triple sets that meet the confidence requirements are written into the crane machinery fault knowledge graph, thereby improving the reliability of the crane machinery fault knowledge graph.

[0045] Regarding the data structure of the knowledge graph for crane machinery faults, combined with Figure 2 As shown, as a preferred implementation option, in this scheme S01, the crane machinery fault knowledge graph adopts a hierarchical modeling approach. Equipment type nodes are located in the equipment layer, configuration and component nodes are located in the structure layer, fault mode nodes are located in the fault layer, symptom nodes are located in the symptom layer, operating condition nodes are located in the operating condition layer, inspection item nodes and maintenance plan nodes are located in the handling layer, and event and status nodes are located in the time event layer. The event and status nodes include alarm event nodes, shutdown event nodes, and sensor over-limit event nodes, and are connected to the corresponding fault mode nodes through event-fault mode relationship edges to support time-sequence-based event-fault mode matching.

[0046] In addition, in this solution S02, the operating data is obtained by extracting information from the question and answer request, or by obtaining it from the crane control system.

[0047] In addition to the above, for the types or items of operational data collection, in this scheme S02, the operational data includes the load, amplitude, lifting height, operating mode, wind speed, alarm records, and statistical characteristics of some key sensors of the lifting machinery in the current or recent period; the operational data summary is the operational data described in natural language.

[0048] As an example of implementation, in S02 of this solution, the question-and-answer request about crane malfunction input by the operator can be a natural language question or information input according to a preset syntax. During the conversion process, it can be transformed into a corresponding question semantic vector by a text encoder. This is for subsequent model recognition processing, and its function definition is as follows: ; in, For natural language questions about question-answering requests, This is a text encoding model, which can be the encoding part of BERT or LLM; The dimension of the problem vector.

[0049] To transform operational data into operational data summaries, this can be achieved by creating feature vectors related to operational conditions. This is done by concatenating different operational data items into a feature vector r. Furthermore, to facilitate understanding by larger models, the model or function can be further described using natural language. The method of converting the feature vector r of the running data into a text summary, i.e., running data summary. Its function is defined as follows: .

[0050] As a preferred implementation option, preferably, in this scheme, step S02 can address the natural language problem. Feature vector r of the running data, and summary of the running data. Problem semantic vector As an output item, it can be reused later.

[0051] In step S03, as an example, the model input sequence can be formed by concatenating the natural language question with the running data summary. The input is defined as follows: ; in, The system prompt text for guiding the generation of task types and graph query templates by the large language model can be pre-built and directly retrieved, so it will not be elaborated here.

[0052] When generating task types, this approach can treat the large language model as a conditional generation model, or it can abstract the probability distribution of task types and then define the task types based on the probability. This is existing technology and will not be elaborated here.

[0053] As a preferred implementation option, in this scheme S03, the graph query template is used to limit the target equipment type, candidate components, candidate symptoms, operating condition constraints, and the expected output fault modes and handling suggestions.

[0054] The graph query template includes multiple information slots to be filled, including at least a target equipment type slot, a target component slot, a target symptom slot, a working condition constraint slot, and an output item slot. The output item slot is used to indicate the query return fault mode, the corresponding evidence node, and the recommended inspection and maintenance nodes. The large language model fills the information in each slot according to the description in the natural language question and the running data summary, generating a graph query statement containing specific entities or limiting conditions.

[0055] As an example of the graph query template, it can be predefined and then associated with the system prompt text as an input item, as shown in the following definition: ; in, These correspond to the target equipment type slot, target component slot, target symptom slot, operating condition constraint slot, and output item slot, respectively.

[0056] In this approach, the large language model can generate the content for each slot based on a predefined template. The function definition is as follows: ; in, The specific results after querying the slots within the template for filling the large language model's graph. This defines a function for generating content within a graph query template for a large language model.

[0057] The large language model output includes structured instructions that include task type identifiers and graph query templates. and task type identifier Convert into an executable query statement for a graph database (knowledge graph of crane machinery faults). It can be defined as follows: .

[0058] In solution S04, a graph database is invoked via a query statement. Entity matching and multi-hop relationship search are performed on the aforementioned knowledge graph of crane machinery faults, and the function is defined as follows: ;

[0059] in, For knowledge graphs of crane machinery faults The function that executes the query. Let k be the number of candidate explanatory subgraphs, and k be the number of the candidate explanatory subgraph. Let the k-th candidate explanation subgraph be defined as follows: ; in, These are the sets of nodes and edges of the candidate interpretation subgraph, respectively.

[0060] By combining the pattern description of the event time sequence, event-fault pattern matching is performed to obtain several candidate explanatory subgraphs, and then the candidate explanatory subgraphs are scored and sorted according to preset conditions.

[0061] Based on the above scheme, as a preferred implementation option, in S04 of this scheme, each candidate interpretation subgraph includes at least one suspected fault mode node and its connected component nodes, symptom nodes, operating condition nodes, event and status nodes, inspection item nodes and maintenance scheme nodes.

[0062] In addition, in S04 of this scheme, the candidate interpretation subgraphs are scored and ranked according to the degree of matching between each candidate interpretation subgraph and the current operating data and observed symptoms, as well as the reliability of the knowledge source; wherein, the scoring of each candidate interpretation subgraph considers at least the following factors: (1) The degree of matching between the symptom nodes in the candidate explanatory subgraph and the symptom expressions mentioned in the natural language problem, i.e., the symptom matching score. For the symptom matching score, symptoms can be parsed from natural language questions and operational data about working conditions, and the aggregated symptom set is denoted as... The symptom information can be obtained through front-end NER or matching for a given candidate explanatory subgraph. The set of symptom nodes is The symptom matching degree is defined as follows: ; in, This refers to the set of observed symptoms parsed from natural language questions input by operators and operational data regarding operating conditions. Candidate Explanation Subgraph The set of symptom nodes, For the set of symptom nodes The standardized text label corresponding to node v in the middle; For text similarity functions; Candidate Explanation Subgraph The score for the degree of match of the symptoms; v represents a single observed symptom; v is a candidate explanatory subgraph. A single symptom node in the data.

[0063] (2) The consistency between the working condition nodes in the candidate interpretation subgraph and the working condition characteristics in the current operating data, i.e., the candidate interpretation subgraph Matching score with working conditions For candidate explanation subgraph The matching score with the operating condition can be obtained from the feature vector r of the operating data to obtain a summary of the operating condition features, and an interpretation subgraph can be selected. The working condition nodes are Each working condition node can have a feature vector. (e.g., standard operating condition description embedding), candidate interpretation subgraph The matching degree score with the working condition is defined as follows: ; in, Candidate Explanation Subgraph Matching score with operating conditions For a single working condition node in the candidate subgraph, This is the similarity function between working condition vectors.

[0064] (3) The reliability level of the knowledge items involved in the candidate explanation subgraph, i.e., the reliability score. For the source reliability level of knowledge items (information related to lifting machinery includes crane manuals, standards and specifications, maintenance records, failure cases, and expert experience), assuming that all edges involved in reasoning in the candidate explanation subgraph have pre-set source weights, the credibility function of the candidate explanation subgraph is defined as follows: ; in, Candidate Explanation Subgraph The credibility score of the knowledge source. Candidate Explanation Subgraph The corresponding set of directed edges in the form of ternary elements, The credibility weights of the edges involved in the reasoning. The index number is the edge number.

[0065] (4) The structural complexity and path length of the candidate explanatory subgraph, i.e., the simplicity score. By analyzing the candidate explanation subgraph The complexity can be constrained, for example, by using average path length or number of nodes as functions to control overly complex interpretations, as defined below: ; in, The conciseness score is used to evaluate the candidate explanatory subgraphs. As a length penalty factor, This represents the average length of the edges involved in the reasoning.

[0066] The comprehensive score of the candidate explanatory subgraph is obtained by weighted summation of one or more of the above factors. As an example, its function can be defined as follows: ; in, , , , These are related to the symptom matching score and the selected subgraph. Weighting coefficients for the matching score with the working conditions, the credibility score, and the simplicity score.

[0067] Based on the comprehensive score of candidate explanatory subgraphs The sorting process allows us to obtain the top K candidate interpretation subgraphs and aggregate them to form a candidate interpretation subgraph set, which is then used as input for step S05.

[0068] Combination Figure 3 As shown, in addition to the solutions mentioned above, as a preferred implementation option, this solution also includes: S06. Extract information from the diagnostic response generated by the large language model and parse it into a set of structured triples that includes one or more of the following relationships: "equipment-component-failure mode", "failure mode-symptom", "failure mode-operating condition", "failure mode-inspection item", and "failure mode-maintenance plan". Compare the set of triples with the knowledge graph of crane machinery faults. If all key relationships in the triplet set have connected paths in the knowledge graph of crane machinery faults and do not contradict the equipment structure or fault mechanism, then the diagnostic answer will be output as the final question and answer result. If a significant conflict is detected with the knowledge graph of crane machinery faults, the conflict information and the relevant graph node summary are input into the large language model to trigger a rewrite of the diagnostic response. If the conflict cannot be eliminated after a preset number of rewrites, a conservative diagnostic result with uncertainty prompts is output.

[0069] As a preferred implementation option, in S06 of this scheme, when extracting information from the diagnostic response, a large language model or sequence labeling model is used to parse the natural language response into a set of structured triples. Each triple in the set of triples includes at least a subject entity, a relation type, and an object entity. The subject entity and the object entity correspond to the equipment type node, configuration and component node, fault mode node, symptom node, operating condition node, inspection item node, or maintenance plan node in the knowledge graph of crane machinery faults, respectively. The relation type corresponds to the relation edge type in the knowledge graph of crane machinery faults.

[0070] As a preferred implementation option, in S06 of this scheme, when comparing the set of triples with the knowledge graph of crane machinery faults, the system queries the knowledge graph of crane machinery faults to see if there is a predefined type of connection path between the nodes corresponding to each triple in the set of triples. When a triple does not have any connection path in the knowledge graph of crane machinery faults that satisfies the equipment structure constraints and fault mechanism constraints, it is marked as a conflicting triple. If a conflicting triple exists, a conflict description containing the conflicting triple and related node information is generated and input into the large language model again along with the original natural language question and the runtime data summary to trigger the large language model to rewrite the diagnostic answer under the conflict constraints.

[0071] This solution improves the reliability and robustness of results by introducing a knowledge graph-based consistency verification and rewriting mechanism in step S06. After the large model generates a diagnostic answer, the answer content can be parsed into structured triples such as "equipment-component-fault-symptom-operating condition-inspection-maintenance" through information extraction. Connectivity and mechanistic constraints are then checked against the knowledge graph, and relationships that do not conform to the graph structure or are obviously contradictory are marked. When overall consistency is insufficient, conflict information is fed back to the large model in the form of a prompt, triggering it to rewrite the answer under the constraints of the knowledge graph. Compared to traditional solutions that rely solely on the output of the large model, this invention performs a "secondary verification" of the answer through the knowledge graph, effectively reducing the risk of the large model generating incorrect reasoning or illusory answers, and improving the reliability and robustness of the system in safety-critical scenarios.

[0072] Based on the above, this solution also proposes a method for querying crane machinery faults, which applies the aforementioned intelligent question-and-answer method for crane machinery faults based on the fusion of knowledge graphs and large models; it also includes: Based on the fault results confirmed in actual maintenance, these results are packaged with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. The edge weights, newly added nodes, and relationships in the crane machinery fault knowledge graph are updated based on multiple question-and-answer examples. At the same time, samples of "question-graph query-candidate explanation subgraph-diagnostic answer" are extracted from the question-and-answer examples to incrementally fine-tune the large language model or to serve as dialogue context examples, thereby improving its ability to generate graph query templates and diagnostic answers in the future.

[0073] This solution enables continuous evolution through bidirectional updates based on question-and-answer examples. A complete question-and-answer example is formed by "natural language question - runtime data summary - graph query - candidate explanation subgraph - diagnostic answer - actual maintenance result," which is used to update the entity and relation weights in the knowledge graph and serves as training or prompt samples for "large model to generate graph queries, understand graph summaries, and output structured diagnoses." Through continuous accumulation of real-world cases, this invention achieves bidirectional enhancement of the knowledge graph and the large model, allowing the system to continuously optimize as equipment models evolve, operating conditions change, and field experience increases. This results in significant engineering deployment value and long-term maintainability.

[0074] Combination Figure 4 As shown above, this solution also proposes an intelligent question-and-answer system for crane machinery faults based on the fusion of knowledge graphs and large models, which includes: The graph construction unit is used to construct a knowledge graph of crane machinery faults, which includes multiple nodes and relation edges connecting the nodes. It also extracts entities and relations related to crane machinery and writes the extracted entities and relations into the knowledge graph of crane machinery faults. The question collection unit is used to receive question and answer requests about crane malfunctions input by operators, convert the question and answer requests into natural language questions, and at the same time acquire the crane's operating data related to the question and answer requests and convert it into an operating data summary. The model loading unit loads a large language model that has been fine-tuned and trained on a crane domain corpus and is used to provide question-and-answer information processing. The information scheduling unit is used to input the natural language problem and the summary of operation data into a large language model fine-tuned by the crane domain corpus, and the large language model outputs structured instructions including task type identifiers and graph query templates, wherein the graph query templates contain pattern descriptions of the time sequence of events. The information processing unit is used to perform entity matching and multi-hop relationship search on the knowledge graph of crane machinery faults according to the graph query template, and to perform event-fault pattern matching in combination with the pattern description of event time sequence to obtain several candidate explanation subgraphs, and then to score and sort the candidate explanation subgraphs according to preset conditions. The answer generation module converts the natural language question, the runtime data summary, and the sorted candidate explanation subgraphs into a fusion context in graph summary form and inputs it into the large language model. The large language model generates a crane mechanical fault diagnosis answer under the constraints of the fusion context. The diagnosis answer includes at least one main fault mode and optional alternative fault modes, corresponding explanations, and suggested inspection steps and maintenance plans. The model optimization module is used to package the confirmed fault results from actual maintenance with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. Based on multiple question-and-answer examples, it updates the edge weights and adds new nodes and relationships in the crane machinery fault knowledge graph. At the same time, it extracts samples of "question-graph query-candidate explanation subgraph-diagnostic answer" from the question-and-answer examples to incrementally fine-tune the large language model or use them as dialogue context examples to improve its ability to generate graph query templates and diagnostic answers in the future.

[0075] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent question answering of crane machinery faults based on the fusion of knowledge graph and large model, characterized in that, It includes: S01. Construct a knowledge graph of crane machinery faults, which includes multiple nodes and relational edges connecting the nodes. Also, extract entities and relations related to crane machinery, and write the extracted entities and relations into the knowledge graph of crane machinery faults. S02. Receive the question and answer request about the crane malfunction input by the operator, convert the question and answer request into a natural language question, and at the same time obtain the crane operation data related to the question and answer request and convert it into an operation data summary. S03. Input the natural language question and the summary of the operation data into a large language model fine-tuned by the crane domain corpus. The large language model outputs a structured instruction including a task type identifier and a graph query template. The graph query template contains a pattern description of the time sequence of events. S04. Based on the graph query template, perform entity matching and multi-hop relationship search on the knowledge graph of crane machinery faults, and combine the event-fault pattern matching with the pattern description of the event time sequence to obtain several candidate explanation subgraphs. Then, score and sort the candidate explanation subgraphs according to preset conditions. S05. The natural language question, the running data summary, and the sorted candidate explanation subgraph are converted into a fusion context in graph summary form and input into the large language model. The large language model generates a crane mechanical fault diagnosis answer under the constraints of the fusion context. The diagnosis answer includes at least one main fault mode and optional alternative fault modes, corresponding explanations, and suggested inspection steps and maintenance plans.

2. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 1, characterized in that, In S01, the multiple nodes of the crane machinery fault knowledge graph include multiple nodes from equipment type nodes, configuration and component nodes, fault mode nodes, symptom nodes, operating condition nodes, inspection item nodes, maintenance plan nodes, and event and status nodes. The relationship edges are formed by connecting nodes in the crane machinery fault knowledge graph through one or more of the following corresponding relationships: "equipment-component", "component-fault mode", "fault mode-symptom", "operating condition-fault mode", "fault mode-inspection item", "fault mode-maintenance plan", and "event-fault mode". The information related to the lifting machinery includes crane manuals, standards and specifications, maintenance records, failure cases, and expert experience.

3. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 2, characterized in that, In S01, the knowledge graph of crane machinery faults adopts a hierarchical modeling approach. Equipment type nodes are located in the equipment layer, configuration and component nodes in the structure layer, fault mode nodes in the fault layer, symptom nodes in the symptom layer, operating condition nodes in the operating condition layer, inspection item nodes and maintenance plan nodes in the handling layer, and event and status nodes in the time event layer. The event and status nodes include alarm event nodes, shutdown event nodes, and sensor over-limit event nodes, and are connected to the corresponding fault mode nodes through event-fault mode relationship edges to support time-sequence-based event-fault mode matching. In S02, the operating data is obtained by extracting information from the question and answer request, or by obtaining it from the crane control system. In S02, the operating data includes the load, amplitude, lifting height, operating mode, wind speed, alarm records, and statistical characteristics of some key sensors of the lifting machinery in the current or recent period. The runtime data summary is a description of the runtime data in natural language.

4. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in any one of claims 1 to 3, characterized in that, In S03, the graph query template is used to limit the target equipment type, candidate components, candidate symptoms, operating condition constraints, and the expected output fault modes and processing suggestions. The graph query template includes multiple information slots to be filled, including at least a target equipment type slot, a target component slot, a target symptom slot, a working condition constraint slot, and an output item slot. The output item slot is used to indicate the query return fault mode, the corresponding evidence node, and the recommended inspection and maintenance nodes. The large language model fills the information in each slot according to the description in the natural language question and the running data summary, generating a graph query statement containing specific entities or limiting conditions.

5. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 4, characterized in that, In S04, each candidate explanation subgraph includes at least one suspected fault mode node and its connected component nodes, symptom nodes, operating condition nodes, event and status nodes, inspection item nodes and maintenance plan nodes. In S04, the candidate interpretation subgraphs are scored and ranked according to the degree of matching between each candidate interpretation subgraph and the current running data and observed symptoms, as well as the reliability of the knowledge source. The scoring of each candidate explanatory subgraph should consider at least the following factors: (1) The degree of matching between the symptom nodes in the candidate explanatory subgraph and the symptom expressions mentioned in the natural language problem; (2) The consistency between the working condition nodes in the candidate interpretation subgraph and the working condition features in the current operating data; (3) The reliability level of the source of the knowledge entries involved in the candidate explanation subgraph; (4) The structural complexity and path length of the candidate interpretation subgraph; The comprehensive score of the candidate explanatory subgraph is obtained by weighted summation of one or more of the above factors.

6. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 4, characterized in that, It also includes: S06. Extract information from the diagnostic response generated by the large language model, and parse it into a set of structured triples that include one or more of the following relationships: "equipment-component-fault mode", "fault mode-symptom", "fault mode-operating condition", "fault mode-inspection item", and "fault mode-maintenance plan". Compare the set of triples with the crane machinery fault knowledge graph. If all key relationships in the triplet set have connected paths in the knowledge graph of crane machinery faults and do not contradict the equipment structure or fault mechanism, then the diagnostic answer will be output as the final question and answer result. If a significant conflict is detected with the knowledge graph of crane machinery faults, the conflict information and the relevant graph node summary are input into the large language model to trigger a rewrite of the diagnostic response. If the conflict cannot be eliminated after a preset number of rewrites, a conservative diagnostic result with uncertainty prompts is output.

7. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 6, characterized in that, In S06, when extracting information from the diagnostic response, a large language model or sequence labeling model is used to parse the natural language response into a set of structured triples. Each triple in the set of triples includes at least a subject entity, a relation type, and an object entity. The subject entity and the object entity correspond to the equipment type node, configuration and component node, fault mode node, symptom node, operating condition node, inspection item node, or maintenance plan node in the knowledge graph of crane machinery faults, respectively. The relation type corresponds to the relation edge type in the knowledge graph of crane machinery faults.

8. The intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in claim 6, characterized in that, In S06, when comparing the set of triples with the knowledge graph of crane machinery faults, the knowledge graph of crane machinery faults queries whether there is a predefined type of connection path between the nodes corresponding to each triple in the set of triples. When a triple does not have any connection path in the knowledge graph of crane machinery faults that satisfies the equipment structure constraints and fault mechanism constraints, it is marked as a conflicting triple. If a conflicting triple exists, a conflict description containing the conflicting triple and related node information is generated and input into the large language model again along with the original natural language question and the running data summary to trigger the large language model to rewrite the diagnostic answer under the conflict constraints.

9. A smart question-and-answer system for crane machinery faults based on the fusion of knowledge graphs and large models, characterized in that, It includes: The graph construction unit is used to construct a knowledge graph of crane machinery faults, which includes multiple nodes and relation edges connecting the nodes. It also extracts entities and relations related to crane machinery and writes the extracted entities and relations into the knowledge graph of crane machinery faults. The question collection unit is used to receive question and answer requests about crane malfunctions input by operators, convert the question and answer requests into natural language questions, and at the same time acquire the crane's operating data related to the question and answer requests and convert it into an operating data summary. The model loading unit loads a large language model that has been fine-tuned and trained on a crane domain corpus and is used to provide question-and-answer information processing. The information scheduling unit is used to input the natural language problem and the summary of operation data into a large language model fine-tuned by the crane domain corpus, and the large language model outputs structured instructions including task type identifiers and graph query templates, wherein the graph query templates contain pattern descriptions of the time sequence of events. The information processing unit is used to perform entity matching and multi-hop relationship search on the knowledge graph of crane machinery faults according to the graph query template, and to perform event-fault pattern matching in combination with the pattern description of event time sequence to obtain several candidate explanation subgraphs, and then to score and sort the candidate explanation subgraphs according to preset conditions. The answer generation module converts the natural language question, the runtime data summary, and the sorted candidate explanation subgraphs into a fusion context in graph summary form and inputs it into the large language model. The large language model generates a crane mechanical fault diagnosis answer under the constraints of the fusion context. The diagnosis answer includes at least one main fault mode and optional alternative fault modes, corresponding explanations, and suggested inspection steps and maintenance plans. The model optimization module is used to package the confirmed fault results from actual maintenance with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. Based on multiple question-and-answer examples, it updates the edge weights and adds new nodes and relationships in the crane machinery fault knowledge graph. At the same time, it extracts samples of "question-graph query-candidate explanation subgraph-diagnostic answer" from the question-and-answer examples to incrementally fine-tune the large language model or use them as dialogue context examples to improve its ability to generate graph query templates and diagnostic answers in the future.

10. A method for diagnosing faults in lifting machinery, characterized in that, Its application includes the intelligent question-answering method for crane machinery faults based on the fusion of knowledge graph and large model as described in any one of claims 1 to 8; It also includes: Based on the fault results confirmed in actual maintenance, these results are packaged with natural language questions, operational data summaries, graph query templates, candidate explanation subgraphs, and diagnostic answers to form crane machinery fault question-and-answer examples. The edge weights, new nodes, and relationships in the crane machinery fault knowledge graph are updated based on multiple question-and-answer examples. At the same time, samples of "question-graph query-candidate explanation subgraph-diagnostic answer" are extracted from the question-and-answer examples to incrementally fine-tune the large language model or to serve as dialogue context examples, thereby improving its ability to generate graph query templates and diagnostic answers in the future.

Images