A large model-based diagnostic knowledge graph construction and application method and system

Abstract

This invention provides a method and system for constructing and applying a diagnostic knowledge graph based on a large model. The method includes: acquiring multi-source maintenance data; performing visual-semantic concatenation parsing on unstructured maintenance documents to obtain structured knowledge data; aligning knowledge based on the structured knowledge data and structured maintenance records, storing standardized entity nodes and relational edges in a graph database, storing corresponding semantic vectors and metadata in a vector database, and establishing a bidirectional mapping index; acquiring diagnostic query information based on the vehicle-side fault data and identifying the query type, performing semantic and text retrieval respectively, and fusing retrieval scores based on query type to output diagnostic results; extracting common diagnostic nodes as priority investigation nodes when multiple fault codes occur concurrently; and dynamically recommending and updating diagnostic steps based on historical diagnostic frequency and user feedback. This improves the efficiency of maintenance knowledge construction, retrieval accuracy, and fault diagnosis efficiency.

Description

Technical Field

[0001] This invention relates to the field of vehicle diagnostic technology, and in particular to a method and system for constructing and applying a diagnostic knowledge graph based on a large model. Background Technology

[0002] With the development of vehicle networking and remote diagnostic technologies, the amount of data generated during vehicle after-sales maintenance is constantly increasing. Existing automotive after-sales maintenance diagnostic systems typically use a combination of relational databases, rule tree diagnostics, and keyword retrieval, using maintenance data and historical cases as the basis for queries, and outputting troubleshooting steps to maintenance personnel according to a preset diagnostic process.

[0003] However, existing solutions still have the following shortcomings: On the one hand, unstructured documents such as maintenance manuals, circuit diagrams, and technical bulletins rely heavily on manual input or character recognition, making it difficult to accurately extract the layout structure, test logic, electrical connection relationships, and fault causal relationships from the documents, resulting in lagging knowledge updates and coarse-grained analysis. On the other hand, traditional keyword retrieval struggles to balance the precise matching of professional terms such as fault codes and part numbers with the semantic generalization of technicians' colloquial fault descriptions, easily leading to insufficient or false recalls. Traditional fixed diagnostic paths are also difficult to dynamically adjust based on historical maintenance frequency and technician feedback. When multiple fault codes are reported simultaneously, it is impossible to effectively identify common diagnostic steps or potential common root causes, requiring maintenance personnel to check each one individually, resulting in low troubleshooting efficiency. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for constructing and applying a diagnostic knowledge graph based on a large model, which aims to solve the technical problem of low troubleshooting efficiency in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides: a method for constructing and applying a diagnostic knowledge graph based on a large model, comprising the following steps: Acquire multi-source repair data in the automotive aftermarket repair field, including unstructured repair documents, structured repair records, and vehicle-side fault data; The unstructured maintenance documents are subjected to visual-semantic concatenated parsing to obtain structured knowledge data; Based on the vehicle-side fault data, diagnostic query information is obtained. Semantic retrieval and text retrieval are performed on the diagnostic query information respectively. The fusion weight of semantic retrieval score and text retrieval score is determined based on the query type. The comprehensive score of candidate diagnostic results is calculated to output the target diagnostic result. Based on the target diagnostic results, diagnostic steps are dynamically recommended to the user, and the historical diagnostic frequency and / or graph relationship weights of the corresponding diagnostic steps are updated according to the user feedback results to form a closed-loop update of the diagnostic knowledge graph.

[0006] According to one aspect of the above technical solution, the visual-semantic cascade parsing includes extracting the layout structure information of the unstructured maintenance document through a first parsing channel, extracting the semantic entities and entity relationships of the unstructured maintenance document through a second parsing channel, and then performing confidence alignment and result fusion based on the layout structure information and the semantic entities and entity relationships.

[0007] According to one aspect of the above technical solution, the steps for performing visual-semantic concatenated parsing on the unstructured maintenance document include: The first parsing channel identifies the bounding box coordinates, reading order, and table row and column structure of text blocks, tables, illustrations, and / or formulas in the unstructured maintenance document, thereby obtaining a page structure tree with spatial location attributes. The second parsing channel identifies the reference relationships of maintenance terms, the causal relationships between fault phenomena and test standards, the component symbols in the circuit diagram, the wiring harness routing, and the electrical connection relationships in the unstructured maintenance document, and obtains entity relationship triples with semantic tags. Using the bounding box coordinates as spatial reference anchor points, the entity relationship triples are projected onto the corresponding physical regions; When the first parsing channel and the second parsing channel have the same parsing result for the same physical region, the corresponding parsing confidence scores are superimposed or weighted and fused to generate high-weight structured nodes. When the first parsing channel and the second parsing channel have conflicting parsing results for the same physical region, or when the confidence level of either parsing channel is lower than the preset confidence threshold, cross-validation is triggered, and semantic completion or coordinate correction is performed based on the high-confidence parsing result and the complementary features of the other parsing channel.

[0008] According to one aspect of the above technical solution, the step of performing visual-semantic concatenated parsing on the unstructured maintenance document further includes: Select the differentiated parsing subprocess based on the content type in the unstructured maintenance document; When the content type is normal text, semantic blocks are segmented according to paragraph logic and chapter hierarchy to construct a text index tree; When the content type is a symptom table, a mapping relationship between the fault phenomenon, possible causes and troubleshooting suggestions is established based on the table row and column coordinates and column semantic alignment results. When the content type is precise test content, extract the closed-loop test relationship between test conditions, measuring instruments, standard values ​​and judgment logic; When the content type is a circuit diagram, component nodes and electrical connection edges are generated based on the coordinates of graphic elements, component symbol types, wiring harness directions, and electrical control logic.

[0009] According to one aspect of the above technical solution, the step of knowledge alignment based on the structured knowledge data and the structured maintenance records includes: The names of the parts to be aligned are encoded into semantic vectors, and the cosine similarity between the names of the parts to be aligned and the existing parts nodes is calculated. Supervised training data containing query statements, positive sample sets, and negative sample sets is constructed based on historical fault description data, and a semantic similarity model is trained using the binary cross-entropy loss function. The fault description to be aligned is encoded into a fault description semantic vector. Based on the semantic similarity model, fault descriptions with semantic similarity greater than the second similarity threshold are aggregated to generate standardized fault phenomenon nodes, and the mapping relationship between the standardized fault phenomenon nodes and the original fault description is retained. By using a domain-adapted large language model, structured maintenance records with incomplete fault descriptions and / or missing fault locations are used to summarize fault phenomena, correct typos, extract fault locations, and classify fault causes, forming a hierarchical knowledge structure.

[0010] According to one aspect of the above technical solution, the steps of performing semantic retrieval and text retrieval on the diagnostic query information according to the query type, and calculating the comprehensive score of the candidate diagnostic results, include: Extract at least one of vehicle identification information, fault codes, and natural language fault phenomenon descriptions from the diagnostic query information, and mark the query type as fault code precise query, fault phenomenon semantic query, or mixed query. The natural language description of the fault phenomenon is converted into a query semantic vector, and the semantic similarity between the query semantic vector and the document semantic vector is calculated in the vector database to obtain the semantic retrieval result. Keyword extraction and word form normalization are performed on the diagnostic query information. Exact matching, partial matching and / or fuzzy matching are performed in the full-text index of the graph database, and text retrieval results are obtained based on text relevance scores. The semantic retrieval results and the text retrieval results are mapped to a unified graph node identifier space, and duplicate candidate nodes are deduplicated and their scores are retained. Based on the query type, basic semantic weights and basic text weights are determined to calculate the comprehensive score of the candidate diagnostic results.

[0011] According to one aspect of the above technical solution, when the diagnostic query information includes multiple fault codes, the step of performing concurrent diagnostic analysis on the multiple fault codes based on the graph database includes: Using multiple fault codes as a combined query key, historical repair cases containing multiple fault codes are retrieved from the graph database. Calculate the case matching degree between historical maintenance cases and current diagnostic query information; When there are historical repair cases whose case matching degree reaches the preset matching conditions, the diagnostic path corresponding to the historical repair cases will be output first. When there are no historical repair cases that meet the preset matching conditions, standard diagnostic paths are constructed for each fault code to obtain common diagnostic steps that include the intersection nodes of multiple standard diagnostic paths. The common diagnostic steps are prioritized based on their frequency of occurrence in multiple standard diagnostic paths and / or their upstream position in the graph topology.

[0012] According to one aspect of the above technical solution, the step of dynamically recommending diagnostic steps to the user based on the target diagnostic results includes: Load the standard diagnostic topology tree corresponding to the target fault object from the graph database. The standard diagnostic topology tree includes possible cause nodes, test condition nodes, measurement standard nodes, and troubleshooting step nodes. Query the historical successful resolution frequency of each diagnostic step in the standard diagnostic topology tree, and perform initial sorting of diagnostic steps at the same level according to the historical successful resolution frequency; Determine the first-hop node and second-hop node of the current diagnostic node, and calculate the comprehensive recommendation weight of the first-hop node; Push a preset number of diagnostic steps ranked first by comprehensive recommendation weight to the user's terminal, and dynamically load the next hop node after the user confirms or clicks on the terminal. When the user reports that the diagnostic steps have resolved the fault, the historical diagnostic frequency of the diagnostic steps is incremented in the positive direction.

[0013] According to one aspect of the above technical solution, the method further includes: The multi-source maintenance data is accessed through a message queue, and the unstructured maintenance document parsing results are written into the knowledge construction queue, the structured maintenance record alignment results are written into the knowledge update queue, and the vehicle-side real-time fault code stream is written into the diagnostic trigger queue. Sliding time window aggregation is performed on the real-time fault code stream corresponding to the same vehicle identification information to deduplicate and align timestamps of repeatedly reported fault codes. When at least two fault codes or a new fault code is detected within the sliding time window, multi-fault code concurrent diagnostic analysis is triggered, and diagnostic suggestions are pushed to the user terminal.

[0014] Secondly, this solution also provides a diagnostic knowledge graph construction and application system based on a large model, including: The data acquisition module is used to acquire multi-source maintenance data in the field of automotive aftermarket repair, including unstructured maintenance documents, structured maintenance records, and vehicle-side fault data. The visual-semantic parsing module is used to perform visual-semantic concatenated parsing on the unstructured maintenance documents to obtain structured knowledge data; The output module is used to obtain diagnostic query information based on the vehicle fault data, perform semantic retrieval and text retrieval on the diagnostic query information respectively, determine the fusion weight of semantic retrieval score and text retrieval score based on query type, calculate the comprehensive score of candidate diagnostic results, and output the target diagnostic result. The update module is used to dynamically recommend diagnostic steps to the user based on the target diagnostic results, and update the historical diagnostic frequency and / or graph relationship weight of the corresponding diagnostic steps according to the user feedback results, so as to form a closed-loop update of the diagnostic knowledge graph.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By uniformly acquiring and processing unstructured repair documents, structured repair records, and vehicle-side fault data in the automotive aftermarket repair field, and utilizing a visual-semantic cascaded parsing method to extract layout structure, semantic entities, entity relationships, and circuit topology information from repair documents, the accuracy of knowledge parsing of repair manuals, circuit diagrams, and technical bulletins is improved; through knowledge alignment and graph / vector dual-database collaborative storage, the differences in the multiple names of objects such as parts and fault phenomena can be eliminated, realizing the combination of structured relationship query and semantic generalized retrieval; and through semantic retrieval based on query type... By adaptively integrating with text retrieval, it can cater to both precise query needs such as fault codes and part numbers, and fuzzy query needs such as technicians' colloquial fault descriptions, thereby improving the recall and accuracy of diagnostic results. Through concurrent diagnostic analysis of multiple fault codes, it extracts common diagnostic nodes and dynamically adjusts the recommended order of diagnostic steps based on historical diagnostic frequency and user feedback, which can reduce redundant troubleshooting and improve fault location efficiency. At the same time, through diagnostic feedback, it continuously updates the knowledge graph relationship weights and model training data, so that the diagnostic knowledge base and large model capabilities are continuously optimized with actual repair data, thereby improving the intelligence level and troubleshooting efficiency of automotive after-sales repair diagnosis. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method for constructing and applying a diagnostic knowledge graph based on a large model in the first embodiment of the present invention. Figure 2 This is a structural block diagram of the diagnostic knowledge graph construction and application system based on a large model in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1 This embodiment provides a method for constructing and applying a diagnostic knowledge graph based on a large model. This method is primarily applied to automotive aftermarket repair and diagnostic scenarios, and is particularly suitable for scenarios where repair manuals, circuit diagrams, technical bulletins, historical work orders, claim forms, repair cases, product problem records, and real-time vehicle fault code data all participate in knowledge construction and diagnostic reasoning. In this embodiment, the large model can be a multimodal visual language model and / or a large language model. The diagnostic knowledge graph is used to express the relationships between entities such as vehicle fault codes, fault phenomena, components, fault causes, diagnostic steps, vehicle models, and solutions.

[0020] Please see Figure 1 The figure shows a flowchart of the method for constructing and applying a diagnostic knowledge graph based on a large model in the first embodiment of the present invention. As shown in the figure, the method includes the following steps: Step S100: Acquire multi-source repair data in the automotive aftermarket repair field. This multi-source repair data includes unstructured repair documents, structured repair records, and vehicle-side fault data. Specifically, the system acquires multi-source repair data in the automotive aftermarket repair field, including unstructured repair documents, structured repair records, and vehicle-side fault data. The unstructured repair documents may include repair manuals (PDF), circuit diagrams, technical service announcements, testing standard documents, repair training documents, image-formatted test instructions, scanned copies, etc. The structured repair records may include historical repair work orders, claim forms, repair case libraries, product problem reports, quality problem closed-loop records, expert-annotated cases, etc. The vehicle-side fault data may include vehicle identification information, fault codes, freeze frame data, fault occurrence time, fault frequency, high-frequency sampling data from the vehicle, and fault code streams uploaded from remote diagnostic platforms.

[0021] In one embodiment, the vehicle identification information can be a VIN code, the fault code can be a standard DTC code, such as P0300, P0301, P0420, etc., and the fault description can be natural language text entered by the technician, such as "engine shaking and weak acceleration", "cannot start", "cannot start", "engine fault light on the instrument panel", etc.

[0022] Step S200: Perform visual-semantic concatenated parsing on the unstructured maintenance document to obtain structured knowledge data containing entities, attributes, relationships, coordinate anchors, and parsing confidence. Based on the structured knowledge data and the structured maintenance record, perform knowledge alignment to obtain standardized entity nodes and standardized relationship edges. Store the standardized entity nodes and standardized relationship edges in a graph database, and store the semantic vectors and metadata corresponding to the standardized entity nodes in a vector database. Simultaneously, establish a bidirectional mapping index between graph node identifiers in the graph database and vector identifiers in the vector database.

[0023] The visual-semantic cascade parsing includes extracting the layout structure information of the unstructured maintenance document through a first parsing channel, extracting the semantic entities and entity relationships of the unstructured maintenance document through a second parsing channel, and then performing confidence alignment and result fusion based on the layout structure information and the semantic entities and entity relationships.

[0024] Specifically, in this embodiment, the first parsing channel is a layout structure parsing channel, used to reconstruct the physical structure of the document. The first parsing channel can employ the MinerU high-precision document parsing model, or other document parsing models with layout structure recognition capabilities. The first parsing channel is used to identify layout elements in unstructured maintenance documents, such as text blocks, tables, illustrations, formulas, titles, paragraphs, headers, and footers, and outputs the bounding box coordinates, reading order, table row and column structure, and layout hierarchy of text blocks, tables, illustrations, and formulas, thereby obtaining a layout structure tree with spatial location attributes.

[0025] The second parsing channel is a multimodal semantic reasoning channel used to understand the logical semantics in the document. This second parsing channel can employ the Qwen2.5-VL-72B multimodal visual language model, or other multimodal semantic models capable of processing image, text, and layout information. The second parsing channel is used to identify the references to technical terms in maintenance manuals, the causal chains between fault phenomena and test standards, component symbols in circuit diagrams, wiring harness routing, electrical connections, and control logic, and outputs entity relationship triples with semantic tags.

[0026] During the parsing process, the system uses the bounding box coordinates output by the first parsing channel as the spatial reference anchor point, projecting the semantic entities and entity relationships output by the second parsing channel onto the corresponding physical region. When the parsing results of the first and second parsing channels for the same region are consistent, the system superimposes or weights and fuses the corresponding parsing confidence scores to generate high-weight structured nodes. When the parsing results of the first and second parsing channels for the same region conflict, or when the confidence score of either parsing channel is lower than a preset confidence threshold, the system triggers a cross-validation mechanism and performs semantic completion or coordinate correction based on the high-confidence parsing result and the complementary features of the other parsing channel.

[0027] For example, for a multi-page table in a maintenance manual, the first parsing channel can accurately identify the table's row and column coordinates and the multi-page connection relationship, while the second parsing channel can understand the business meaning of the table header fields, such as "fault phenomenon," "possible cause," and "troubleshooting suggestion." When a multi-page table causes a decrease in the confidence of a single model, the system uses the table coordinates as anchor points and supplements them with column semantics to restore the correspondence between the fault phenomenon, possible cause, and troubleshooting suggestion. As another example, for connectors, fuses, grounding points, and wiring harnesses in a circuit diagram, the first parsing channel locates the spatial coordinates of the graphic elements, while the second parsing channel identifies the component symbol type, connection direction, and electrical control logic. Based on this, the system generates component nodes and electrical connection relationship edges.

[0028] In one embodiment, the system selects a differentiated parsing sub-process based on the content type of the unstructured maintenance document. When the content type is normal text, the system performs semantic block segmentation according to paragraph logic and chapter hierarchy to construct a text index tree. When the content type is a symptom table, the system establishes a mapping relationship between fault phenomena, possible causes, and troubleshooting suggestions based on the table's row and column coordinates and column semantic alignment results. When the content type is precise test content, the system extracts the closed-loop test relationship between test conditions, measuring instruments, standard values, and judgment logic. When the content type is a circuit diagram, the system generates component nodes and electrical connection relationship edges based on the coordinates of graphic elements, component symbol types, wiring harness routing, and electrical control logic.

[0029] In a specific example, for the precise test content "with the ignition switch ON, use a multimeter to measure the voltage of connector A terminal. The standard value is 11V to 14V. If it is lower than 11V, check the fuse and power supply line", the system extracts the test condition as "ignition switch ON", the measuring instrument as "multimeter", the measuring object as "voltage of connector A terminal", the standard value as "11V to 14V", and the judgment logic as "check the fuse and power supply line if it is lower than 11V". The system then converts the above information into test condition nodes, measurement object nodes, standard value attributes, and judgment relationship edges.

[0030] In another specific example, for connectors, fuses, grounding points, and control modules in a circuit diagram, the system converts "graphical symbols + spatial location + electrical semantics" into standardized component nodes and electrical connection edges in a knowledge graph. For example, the fuse symbol is converted into a "fuse" node, the grounding point symbol into a "grounding point" node, the wiring harness direction into an "electrical connection" edge, and the control signal direction into a "control" edge.

[0031] After all parsing results are normalized, the output is structured JSON data conforming to a predefined ontology. This structured JSON data includes at least entity type, attribute keys, relation edges, coordinate anchors, parsing confidence, source document identifier, page number, region coordinates, and timestamp. This structured JSON data serves as input for subsequent knowledge alignment and graph database construction.

[0032] First, the system constructs a vehicle-specific terminology database. This database includes standard component names, aliases, abbreviations, dialect expressions, industry slang, mixed Chinese and English expressions, and non-standard terms from historical maintenance records. For example, "starter motor," "motor," and "starter motor" can be semantically aligned and merged into the same component node; "cannot start," "cannot start," "cannot ignite," and "cannot start" can be semantically aligned and aggregated into the same standardized fault phenomenon node.

[0033] For multi-name alignment of components, the system encodes the component name to be aligned into a component semantic vector and calculates the cosine similarity between the component name to be aligned and existing component nodes. When the cosine similarity is greater than a first similarity threshold, for example, greater than 0.85, the system determines that the two are different representations of the same component, merges the component name to be aligned into the corresponding existing component node, and uses the component name to be aligned as an alias attribute of the existing component node. When a new word appears that is not included in the terminology database, the system analyzes the similarity between the new word and existing nodes through a semantic model, and back-integrates it into the vehicle-specific terminology database when it is manually reviewed or the confidence level meets the conditions, thereby realizing dynamic expansion of the terminology database.

[0034] In one embodiment, the semantic vector model can be either a BGE model or an mGTE model. The BGE model can output a 1024-dimensional semantic vector to improve semantic matching accuracy; the mGTE model can output a flexible dimension vector of 128 to 768 dimensions to balance computational efficiency and storage space.

[0035] For semantic standardization of fault descriptions, the system constructs supervised training data containing query statements, positive sample sets, and negative sample sets based on historical fault description data, and trains a semantic similarity model using a binary cross-entropy loss function. Compared to using only triplet loss, the binary cross-entropy loss function allows the model to simultaneously learn multiple positive and negative samples corresponding to a query statement, thereby improving training efficiency and fault description aggregation accuracy. The system encodes the fault descriptions to be aligned into fault description semantic vectors, aggregates fault descriptions with semantic similarity greater than a second similarity threshold based on the semantic similarity model, generates standardized fault phenomenon nodes, and retains the mapping relationship between standardized fault phenomenon nodes and original fault descriptions.

[0036] For fault cause hierarchical classification and knowledge completion, the system uses a domain-adapted large language model to summarize fault phenomena, correct typos, extract fault locations, and classify fault causes in structured repair records with incomplete fault descriptions and / or missing fault locations. The large language model can be Qwen3-32B or other large language models with semantic understanding capabilities for automotive aftermarket repair. The system automatically extracts fault locations from the repair process or content and classifies and labels them according to a preset fault cause classification system. In one embodiment, the fault cause classification system includes 6 major categories and 19 subcategories, where the 6 major categories may include wiring harnesses, software programs, modules, power supplies and grounding, electrical problems, and other categories. The system organizes knowledge according to a hierarchical structure of "fault phenomenon → primary fault cause classification → secondary fault cause classification → fault location → solution," forming a knowledge framework for guided diagnosis.

[0037] In the collaborative storage process of graph / vector dual databases, the graph database is used to store structured entity nodes and relation edges to represent the topological network between entities. The graph database can be Neo4j, and node types include DTC codes, parts, fault phenomena, fault causes, diagnostic steps, vehicle models, and repair solutions; relation types include cause, belong, contain, investigate, associate, apply to, and connect. The graph database supports multi-hop relation queries and complex path traversal, enabling diagnostic path construction, extraction of common diagnostic steps, and root cause path inference.

[0038] The vector database stores the semantic vector features and metadata of the original text fragments. The vector database can be Milvus, and the metadata can include the original fault description, repair manual paragraphs, technical service announcements, source documents, vehicle models, timestamps, and confidence scores. The vector database supports Top-K similarity retrieval and scalar filtering to meet semantic generalization retrieval needs. In one embodiment, Milvus employs a storage-compute separation architecture, with object storage used at the underlying layer to support distributed deployment and horizontal scaling.

[0039] The system establishes a bidirectional mapping index between graph node identifiers and vector identifiers. When the knowledge alignment module adds or updates nodes, the system synchronously creates or updates nodes and relation edges in the graph database, inserts or updates vector records in the vector database, and establishes or updates the correspondence between graph node identifiers and vector identifiers in the bidirectional mapping index table. Through this bidirectional mapping index, the system can achieve a closed-loop query process of "semantic retrieval → graph association → result fusion".

[0040] Regarding dynamic updates, when new repair cases, technical service announcements, or vehicle-side fault data enter the system, the system triggers a knowledge parsing and alignment process to generate new entities and relationships. During knowledge alignment, if semantic conflicts are found between new data and existing nodes (e.g., the same component name pointing to different entities), the system triggers manual review or an automatic resolution strategy based on confidence level. The system records the version number, timestamp, update source, and update content for each update to the knowledge graph, and supports rollback to historical versions in case of data anomalies.

[0041] Step S300: Obtain diagnostic query information based on the vehicle fault data, perform semantic retrieval and text retrieval on the diagnostic query information respectively, determine the fusion weight of semantic retrieval score and text retrieval score based on query type, calculate the comprehensive score of candidate diagnostic results, and output the target diagnostic result.

[0042] Specifically, the system receives diagnostic query information from technician terminals, web interfaces, mobile apps, in-vehicle tablets, or remote diagnostic platforms, and performs intent recognition on the diagnostic query information to determine the corresponding query type. The diagnostic query information includes at least one of vehicle identification information, fault codes, and natural language descriptions of fault phenomena.

[0043] In one embodiment, the system supports two input methods: the first is a combination of VIN code and DTC code / fault symptom input; the second is a single DTC code or a single fault symptom input. The system automatically identifies the input content type and extracts the 17-digit VIN code, the standard 5-digit DTC code, and the natural language fault description text. If the input is a standard DTC code or part number, it is marked as a precise fault code query; if the input is a natural language fault symptom description, it is marked as a semantic fault symptom query; if the input contains both a DTC code and a fault symptom description, it is marked as a mixed query.

[0044] In a specific example, if a user enters "VIN: LXXXXXXXXXXXXXXX, P0300, engine shaking and weak acceleration", the system extracts the vehicle identification information, fault code P0300, and the natural language description of the fault symptom "engine shaking and weak acceleration", and marks the query type as a mixed query. If the user only enters "P0420", the system marks the query type as a precise fault code query. If the user enters "cannot start", the system marks the query type as a semantic fault symptom query.

[0045] The target diagnostic results may include the target fault object, the target diagnostic path, candidate fault causes, recommended diagnostic steps, and / or recommended maintenance solutions.

[0046] For semantic retrieval, the system converts the natural language description of the fault phenomenon into a query semantic vector, and calculates the semantic similarity between the query semantic vector and the document semantic vector in the vector database to obtain the semantic retrieval results. Semantic similarity can be calculated using the following formula: ; In the formula, For semantic similarity, To query semantic vectors, For document semantic vectors, For vector dot product, To query the magnitude of the semantic vector, The magnitude of the document semantic vector; In one embodiment, the system employs an inverted file index combined with a product quantization index structure, or a hierarchical navigable small-world graph index structure, in the vector database to calculate the cosine similarity between the query vector and the document vectors in the database. The system sets a semantic similarity threshold, for example, 0.75, retaining only candidate results with semantic similarity at or above this threshold, and sorting them in descending order of semantic similarity. The Top-K results are then used as the semantic retrieval result set, for example, K is 50. For each result in the semantic retrieval result set, the system uses a bidirectional mapping index to find the corresponding graph node in the graph database and extracts its one-hop or two-hop neighbor nodes, such as fault causes, diagnostic steps, components, and solutions, to enhance the contextual information of the retrieval results.

[0047] For text retrieval paths, the system extracts keywords and normalizes lemmatization from diagnostic query information. It then performs exact matching, partial matching, and / or fuzzy matching in the full-text index of the graph database, and obtains text retrieval results based on text relevance scores. The system can combine a vehicle-specific terminology database to normalize technical terms; for example, it normalizes "motor" to "starter" and "cannot start" to "cannot start". The full-text index can use Lucene full-text indexing. Matching modes include exact keyword matching, prefix or suffix matching, and fuzzy matching allowing 1 to 2 character edit distances.

[0048] Text relevance scores can be calculated using the BM25 algorithm, and the calculation expression is as follows: ; In the formula, To score the relevance of the text, Let m be the number of query terms and D be the number of documents to be matched. For query terms The word frequency in the document D to be matched. For query terms Inverse document frequency, and For adjustment coefficients, The length of the document to be matched. This represents the average document length. In one embodiment, You can take 1.2. A value of 0.75 is acceptable. The system sorts the results in descending order of text relevance score and uses the Top-M results as the text retrieval result set, for example, M = 50.

[0049] After obtaining the semantic search result set and the text search result set, the system maps them to a unified graph node identifier space and removes duplicate candidate nodes while retaining their scores. If the same node appears in both the semantic search result set and the text search result set, the semantic search score and the text search score of that node are retained; if the node appears in only one of the result sets, the other score is recorded as 0 or treated as a preset missing value.

[0050] The system determines the basic semantic weight and basic text weight based on the query type. For precise fault code queries, the system increases the weight of the text retrieval score, for example, setting the basic semantic weight to 0.2 and the basic text weight to 0.8, to emphasize exact matching and avoid false recalls caused by semantic generalization. For semantic queries of fault symptoms, the system increases the weight of the semantic retrieval score, for example, setting the basic semantic weight to 0.8 and the basic text weight to 0.2, to overcome differences in expression such as colloquialisms, dialects, and typos. For mixed queries, the system adopts a segmented weighting strategy, setting the basic semantic weight to 0.3 and the basic text weight to 0.7 for fault code-related nodes; and setting the basic semantic weight to 0.7 and the basic text weight to 0.3 for fault symptom-related nodes.

[0051] To avoid ranking instability caused by sudden weight changes, the system introduces a weight smoothing factor and calculates the actual semantic weights according to the following formula: ; ; In the formula, For actual semantic weights, Based on basic semantic weights, For weighted smoothing factor, This is the current query sequence number. This represents the total number of queries. This represents the actual text weight; Then, the comprehensive score of the candidate diagnostic results is calculated according to the following formula: ; In the formula, For the overall score, For semantic similarity scores, A score is given for text relevance.

[0052] Subsequently, the system performs secondary sorting and business rule optimization on the fused candidate results. If the diagnostic query information contains a VIN code, the system performs vehicle model matching filtering based on the vehicle model information in the VIN code. For example, it identifies the vehicle model based on the 4th to 8th digits of the VIN code, retaining only fault cases, diagnostic steps, and repair solutions compatible with that vehicle model. The system can also assign a timeliness bonus to repair cases added within the past year, for example, by adding 0.05 points, to prioritize recommending the latest technical solutions. The system can also remove nodes with a resolution confidence level below a preset threshold, for example, removing nodes with a resolution confidence level below 0.7 in step S200.

[0053] When multiple candidate results have the same overall score or a difference less than a preset threshold, the system introduces text similarity as a secondary ranking criterion. The text similarity can be calculated using Levenshtein edit distance and Jaccard word overlap. Candidate results with smaller edit distances and higher word overlap are ranked higher. Finally, the system outputs Top-N diagnostic results, for example, N = 20, and supports paginated display, for example, displaying 10 results per page.

[0054] Preferably, when the diagnostic query information includes multiple fault codes, the system performs concurrent diagnostic analysis on the multiple fault codes based on a graph database, extracts common diagnostic nodes in the diagnostic paths corresponding to the multiple fault codes, and uses the common diagnostic nodes as priority investigation nodes. This process is used to solve the problem that traditional systems can only perform linear investigations one by one when multiple DTCs are reported simultaneously, and cannot identify common root causes and common investigation steps.

[0055] First, the system uses multiple fault codes as a combined query key to retrieve historical repair cases containing multiple fault codes from the graph database, and calculates the case matching degree between the historical repair cases and the current diagnostic query information: ; In the formula, To score the matching degree, This is a set of DTC codes from historical cases. This is the set of DTC codes for the current fault. The number of elements in the set; When a historical repair case with a matching score meeting preset conditions exists (e.g., MatchScore=1.0), the system prioritizes outputting the diagnostic path corresponding to that historical repair case. If no completely matching historical repair case exists, the system constructs a standard diagnostic path for each fault code, for example, following the path construction method of "DTC code → Possible cause → Diagnostic steps → Solution". It then retrieves the diagnostic path corresponding to each fault code from the graph database to obtain the common diagnostic steps containing the intersection nodes of multiple standard diagnostic paths. ; In the formula, For a shared set of diagnostic steps, Let n be the diagnostic path for the i-th DTC code, and n be the total number of DTC codes. The system prioritizes common diagnostic steps based on their frequency of occurrence across multiple standard diagnostic paths and / or their upstream position in the graph topology. Nodes located upstream of multiple fault code paths can be identified as potential root cause nodes. For example, main relay failure, ECU grounding abnormality, and power supply abnormality may simultaneously cause multiple DTC codes. The system prioritizes recommending troubleshooting steps corresponding to such root cause nodes.

[0056] In a specific example, when a vehicle simultaneously reports P0300, P0301, and P0302, the system first searches its historical repair records for cases containing all of these fault codes. If no matching case is found, the system constructs diagnostic paths for P0300, P0301, and P0302 respectively and extracts the intersection nodes. If the intersection node includes "Check ignition coil power supply voltage," "Check ECU ground line," or "Check injector common power supply line," the system prioritizes pushing these shared nodes to the technician to avoid the technician repeatedly performing the same checks for each DTC.

[0057] When the shared diagnostic steps set is empty, the system breaks down multiple fault codes into individual fault codes and performs single-fault code diagnostics sequentially according to the historical occurrence frequency of each individual fault code, from high to low. After completing the troubleshooting of a certain fault code, the system reassesses whether the remaining fault codes have been automatically eliminated due to the repair of the shared root cause, thereby dynamically adjusting the remaining diagnostic tasks.

[0058] Step S400: Based on the target diagnostic results, dynamically recommend diagnostic steps to the user terminal, and update the historical diagnostic frequency and / or graph relationship weight of the corresponding diagnostic steps according to the user terminal feedback results, so as to form a closed-loop update of the diagnostic knowledge graph.

[0059] The system first loads the standard diagnostic topology tree corresponding to the target fault object from the graph database. The standard diagnostic topology tree may include nodes for possible causes, test conditions, measurement standards, troubleshooting steps, and solutions. The system queries the historical successful resolution frequency of each diagnostic step in the standard diagnostic topology tree and initially sorts the diagnostic steps at the same level according to their historical successful resolution frequency, forming a high-frequency priority benchmark recommendation list.

[0060] To reduce the cognitive load on technicians and improve the initial diagnosis accuracy, the system does not directly display the complete diagnostic tree, but instead adopts a step-by-step guided interactive mode. The system designates the current diagnostic node as hop 0, the next possible cause or test step directly associated with the current diagnostic node as hop 1, and the verification steps further associated with the hop 1 node as hop 2. When sorting hop 1 nodes, the system considers not only the historical success frequency of the hop 1 node itself, but also the historical success frequency of its downstream hop 2 nodes for comprehensive weighting, calculating the comprehensive recommendation weight of the hop 1 node according to the following formula: ; in, The overall recommendation weight for the first hop node. The historical success frequency of the first hop node. The historical success frequency of the second hop node associated with the first hop node. and The recommended weighting coefficients; in one embodiment, z can be 0.6 and f can be 0.4.

[0061] The system pushes a preset number of diagnostic steps, ranked by comprehensive recommendation weight, to the user's device, such as the first 2 to 3 diagnostic steps. After the user confirms or selects the step, the next step node is dynamically loaded, achieving a "small steps, fast progress" guidance.

[0062] When a user reports that a certain diagnostic step has resolved the fault, the system determines that the path corresponding to that diagnostic step is valid and performs a positive increment on the historical diagnostic frequency of that step: Freq = Freq + 1. At the same time, the system updates the weights of the corresponding relation edges in the graph database to improve the recommendation priority of that path in subsequent similar diagnostic scenarios.

[0063] When a user reports that a diagnostic step has not resolved a fault or is invalid, the system determines that the high-frequency path is invalid under the current vehicle operating conditions and performs a negative attenuation Freq=Freq-2 on the historical diagnostic frequency of that diagnostic step. The absolute value of the negative attenuation is greater than the absolute value of the positive accumulation, which is used to quickly reduce the recommended priority of the invalid path and avoid technicians repeatedly performing invalid troubleshooting.

[0064] When the historical diagnostic frequency of a certain diagnostic step falls below a safe threshold (e.g., Freq < 5), or when the diagnostic step receives a preset number of invalid feedbacks consecutively (e.g., two consecutive unresolved reports from different technicians), the system switches to the standard diagnostic mode, which executes according to the logical sequence in the repair manual. The standard diagnostic mode proceeds step-by-step in the order of "power supply check → line continuity test → signal waveform analysis → component replacement verification" to prevent over-reliance on historical frequency from causing missed diagnoses of less common faults. After the standard diagnostic mode is completed, the system continues to collect new repair data. When the number of new samples reaches the effective sample threshold (e.g., N ≥ 10), the system reverts to dynamic guided mode.

[0065] For inputting pure fault symptoms without DTC codes, the system employs a probabilistic progressive selection mechanism. Based on semantic retrieval results, the system maps fault symptoms to primary fault cause nodes in the knowledge graph, such as wiring harness faults, software program anomalies, module faults, power supply and grounding faults, electrical problems, and other categories, and displays them in descending order of historical occurrence probability. After the technician selects a primary fault cause, the system dynamically loads the corresponding secondary fault causes and specific fault locations, and continues to sort them by probability. As the user progresses through the selections to the final node, the system directly outputs a standardized repair plan. This repair plan may include required tools, torque standards, safety precautions, test conditions, measurement standards, and post-repair verification steps.

[0066] After the diagnostic process is completed, the system automatically generates a structured repair report. This report includes the actual troubleshooting path, time spent on each step, final resolution point, user feedback, technician notes, vehicle identification information, fault code set, recommended hit rate, and knowledge graph association nodes. The system then sends this structured repair report back to the knowledge alignment module and the graph / vector dual-database system to update node weights, relation edge weights, and vector metadata, forming a closed-loop data flywheel of "diagnosis execution → feedback recording → frequency optimization → graph update → re-diagnosis".

[0067] Preferably, in other embodiments of the present invention, after step S400, the method further includes: Step S500: Construct a dataset of instruction, query, and response triples based on historical work orders, maintenance cases, and technical bulletins. The instructions are used to define the task type of the large language model, the queries include raw maintenance text or semi-structured work order data, and the responses include structured maintenance results. To avoid catastrophic forgetting during domain data fine-tuning, the system employs an incremental fine-tuning strategy that mixes general corpora with automotive aftermarket corpora, performing domain-specific incremental training on the large language model according to the following formula: ; In the formula, For total training losses, For general corpus loss, For the loss of corpus in the automotive aftermarket, These are the weighting coefficients for a general corpus; in one embodiment, A value of 0.2 to 0.3 can be used to enhance the model's expertise in the automotive aftermarket while maintaining its ability to understand general language.

[0068] During domain-specific incremental training, dual-indicator monitoring is performed based on general test set metrics and automotive aftermarket domain test set metrics, and training stops when early stop conditions are met. The instructions define the task type for the large language model, such as "extract the vehicle model, fault symptoms, and fault location from the above repair records." The queries include raw repair text or semi-structured work order data, such as "JP360 model lacks acceleration power; inspection revealed an abnormal throttle position sensor signal." The responses include structured repair results, such as {"Vehicle Model": "JP360", "Fault Symptom": "Lack of acceleration power", "Fault Location": "Throttle Position Sensor"}. The system can employ full parameter fine-tuning combined with low-rank adaptation technology to selectively update the attention layer and feedforward network weights. The system monitors general test set metrics and automotive aftermarket domain test set metrics in real time, and, combined with early stop methods and learning rate decay strategies, stops training when general capabilities and domain capabilities reach a preset balance.

[0069] During the inference deployment phase, the key-value cache during the inference process is divided into multiple cache blocks, and the mapping relationship between logical sequences and physical cache blocks is maintained through page tables, so that cache blocks corresponding to different generated sequences can be stored non-contiguously; thus, the system does not require the key-value cache to occupy physical video memory continuously, which can reduce the degree of video memory fragmentation.

[0070] When multiple inference requests have the same prefix, they share the cache block corresponding to the same prefix. New inference requests are inserted after the cache block of completed requests is released through continuous batch processing to improve the memory reuse rate.

[0071] In one embodiment, the system uses Qwen3-32B as the base model. This base model is suitable for the professional nature of automotive aftermarket repair knowledge, the complexity of multi-turn dialogues, and the need for long text processing. It can understand technical parameters, fault tree logic, and technician's colloquial descriptions in repair manuals. In other embodiments, a large language model with similar language understanding and reasoning capabilities can also be used as the base model. The inference framework can adopt vLLM, and the pagination attention mechanism can adopt PagedAttention. The system achieves efficient utilization of GPU memory under high-concurrency inference requests by allocating cache blocks on demand, using non-contiguous storage, and sharing the same prefix. In terms of engineering deployment, the system encapsulates model weights, inference engine, dependency libraries, and API services into Docker images, supporting elastic scaling of cloud-based Kubernetes clusters and lightweight deployment on in-vehicle diagnostic terminals or edge servers. The system adopts continuous batch processing technology. When a request in a batch is generated, the cache block it occupies is immediately released and a new request is inserted, thereby improving GPU throughput. The system provides RESTful and / or gRPC interfaces to receive front-end diagnostic query requests and return structured maintenance suggestions and knowledge graph association nodes. In one embodiment, the system can control the first character generation time to within 50ms and the end-to-end inference latency to within 200ms.

[0072] Furthermore, after step S500, this method further includes: In step S600, the core algorithms and modules from steps S100 to S500 are engineered and encapsulated. Through multi-source data bus scheduling, real-time connection to the remote diagnostic platform, standardized API output, and two-way feedback mechanism, a full-link business closed loop of "data acquisition → knowledge construction → intelligent diagnosis → result feedback → model evolution" is realized.

[0073] In one embodiment, the system employs a message queue-based microservice architecture. The message queue can use Apache Pulsar or Kafka. The system writes the parsing results of unstructured maintenance documents to a knowledge building queue, the alignment results of structured maintenance records to a knowledge update queue, and the real-time fault code stream from the vehicle to a diagnostic trigger queue. The system encapsulates the visual-semantic parsing service, knowledge alignment service, retrieval service, diagnostic service, and large language model inference service as independent Docker microservices. These microservices communicate via RESTful or gRPC protocols and support elastic scaling of the Kubernetes cluster. The failure of a single service does not affect the overall system operation.

[0074] The system calculates the scheduling priority of queue data according to the following formula: ; In the formula, For scheduling priority, For data freshness, To analyze the confidence level, Depending on the urgency of the business, , , These are the weighting coefficients; For real-time DTC streams, the system can assign a higher level of business urgency; for low-confidence document parsing results, the system can reduce their automatic database entry priority or first place them in the manual review queue.

[0075] To address the issues of out-of-order DTC data, duplicate reporting, and instantaneous spikes caused by high-frequency sampling at the vehicle-side of the remote diagnostic platform, the system constructs a real-time synchronization and aggregation buffer mechanism. In one embodiment, vehicle-side fault code data can be reported at a frequency of once every 5 seconds or once every 10 seconds. The system performs sliding time window aggregation on the real-time fault code stream corresponding to the same VIN code, for example, setting a 30-second sliding time window, and deduplicating and aligning timestamps for duplicate reported fault codes. When at least two fault codes are detected within the sliding time window or a new fault code is added, the system triggers concurrent diagnostic analysis of multiple fault codes and pushes diagnostic suggestions to the user terminal. In one embodiment, the system uses WebSocket long connections and gRPC bidirectional streaming to ensure low-latency data connectivity between the vehicle-side, the platform, and the knowledge graph system.

[0076] The system has a built-in two-way feedback channel. When the user reports a resolved issue, the system records the actual troubleshooting path, troubleshooting time, final fault node, and remarks, and updates the frequency statistics table and relation edge weights in the graph database. When the user reports an unresolved issue or an invalid step, the system negatively attenuates the corresponding diagnostic node and marks the corresponding repair case as a pending review case. After manual review confirms that it is a knowledge blind spot or a parsing error, the system triggers the knowledge alignment module to split nodes, merge nodes, or redirect relations. The system can also periodically extract high-quality instruction, query, and response triplet data accumulated in the closed loop, and incrementally fine-tune the large language model according to the mixed ratio strategy of step S800, so that the professional accuracy of the model continues to improve with the increase of data used.

[0077] Regarding API output, the system employs an intent routing gateway. This gateway automatically identifies the terminal type based on the request headers, including web applications, mobile apps, in-vehicle tablets, and remote diagnostic platforms. The system dynamically adapts the response data structure according to different terminal types. The system outputs a structured diagnostic response in standard JSON format, which may include a diagnostic number, diagnostic step number, diagnostic action, confidence level, association graph node identifier, recommended solution, and overall confidence level. For example, a structured diagnostic response may include the following fields: {"diagnosis_id": "xxx", "steps": [{"step_seq": 1, "action": "measure voltage", "confidence": 0.92, "linked_node_id": "node_123"}], "recommended_solution": "replace ECU", "confidence_score": 0.87} To ensure system stability under high concurrency scenarios, the system can use a Redis second-level cache for high-frequency queries. The cached objects can include common DTC codes, standard fault phenomena, and high-frequency diagnostic paths, with a cache validity period of 24 hours. The system can use a token bucket algorithm for rate limiting, for example, limiting each node to 1000 QPS. When the knowledge graph in step S300 is updated, the system synchronously refreshes the relevant cache to ensure the timeliness of search results.

[0078] In terms of security management, the system employs a role-based access control model. Different roles can include technicians, technical experts, administrators, and external partners. The system configures data access granularity according to roles and anonymizes sensitive fields, such as the last 6 digits of the VIN code, customer information, and contact details. The system also records immutable audit logs for all diagnostic path modifications, knowledge node edits, and model parameter adjustments, and generates version snapshots for each major update to the knowledge graph to support one-click rollback and prevent the spread of erroneous knowledge from impacting business operations.

[0079] In one system embodiment, the large-model-based diagnostic knowledge graph construction and application system includes a data acquisition module, a visual-semantic parsing module, a knowledge alignment module, a graph / vector dual-database storage module, an output module, a diagnostic path module, a priority diagnosis module, an update module, a model training and deployment module, and a data scheduling closed-loop module. The data acquisition module acquires multi-source repair data from the automotive aftermarket repair field. The visual-semantic parsing module performs visual-semantic concatenated parsing on unstructured repair documents. The knowledge alignment module aligns structured knowledge data and structured repair records. The graph / vector dual-database storage module stores standardized entity nodes, standardized relation edges, semantic vectors, and metadata, and establishes a bidirectional mapping index. The output module acquires diagnostic query information based on the vehicle-side fault data and performs intent recognition. The diagnostic path module performs semantic retrieval, text retrieval, and adaptive weight fusion. The priority diagnosis module extracts common diagnostic nodes in multi-fault code scenarios and prioritizes their investigation. The update module updates historical diagnostic frequencies and graph relation weights based on user feedback. The model training and deployment module performs domain incremental training, paginated attention inference deployment, and continuous batch processing scheduling. The data scheduling closed-loop module is used to perform message queue scheduling, sliding window aggregation, feedback closed loop, API output, and security control.

[0080] Several alternative implementation methods are further described below.

[0081] In optional implementation one, the first and second parsing channels can be replaced by other models. For example, the LayoutLMv3 multimodal document understanding model can be used to replace the dual-model architecture to simultaneously process text, images, and layout information; PP-StructureV2 can be used for layout analysis and OCR parsing of technical service announcements; for semi-structured PDF documents, pdfplumber or PDFMiner can be used for text extraction. The circuit diagram key element location model can use PP-PicoDet, or YOLOv8 or RT-DETR, to locate graphic elements such as connectors, fuses, and grounding points.

[0082] In optional implementation two, the semantic vector model can be the Sentence-BERT or Contriever model, or a vector model fine-tuned based on the Qwen2 or ChatGLM series models. The graph database can be Neo4j, or other graph databases such as NebulaGraph, JanusGraph, or HugeGraph. The vector database can be Milvus, or the dense_vector field of Faiss or Elasticsearch, or the Chroma vector library. As long as semantic vector retrieval, complex entity topological representation, and bidirectional mapping between graph node identifiers and vector identifiers can be achieved, the technical objective of this invention can be realized.

[0083] In optional implementation three, the large language model base can be Qwen3-32B, or it can be the Llama series model, Baichuan series model, or ChatGLM series model. The inference acceleration framework can be vLLM, or it can be Text Generation Inference, TensorRT-LLM, or llama.cpp. If the hardware environment is fixed at NVIDIA GPU, TensorRT-LLM can be used to further improve the inference speed; if the deployment scenario is cost-sensitive and has low real-time requirements, CPU inference can be used for offline batch processing.

[0084] In optional implementation four, the dynamic diagnostic path optimization algorithm can use reinforcement learning or Markov decision processes instead of the frequency decay strategy. Reinforcement learning treats the current diagnostic node as the state, the next diagnostic step as the action, and the "solved" or "unsolved" feedback as the reward signal, optimizing the diagnostic path through Q-learning or PPO algorithms. Markov decision processes can perform globally optimal planning of the diagnostic path. However, in scenarios with insufficient cold-start data or high real-time requirements, heuristic strategies using historical frequency and asymmetric decay are more likely to take effect quickly.

[0085] In optional implementation five, the system can be deployed in the cloud, at the edge, or in a cloud-edge collaborative environment. The cloud is responsible for large-scale knowledge graph construction, incremental model fine-tuning, and centralized retrieval services, while the edge is responsible for local caching, initial screening of high-frequency DTCs, and low-latency interaction. When the network is abnormal, the edge can continue to provide basic diagnostic suggestions based on the locally cached high-frequency DTC diagnostic path; when the network recovers, the edge will synchronize the local diagnostic feedback to the cloud knowledge graph system.

[0086] Through the above implementation methods, this invention can unify the access of unstructured maintenance documents, structured maintenance records, and vehicle-side fault data for automated knowledge construction; improve the parsing accuracy of complex maintenance documents and circuit diagrams through visual-semantic cascaded parsing; break down heterogeneous data silos through semantic alignment and graph / vector dual-database collaboration; balance precise matching of professional terms with semantic generalization of colloquial descriptions through adaptive weight fusion of semantic retrieval and text retrieval; extract common diagnostic nodes and root cause paths through concurrent diagnosis of multiple fault codes; dynamically optimize diagnostic paths through historical frequency and feedback decay mechanisms; enhance the professional capabilities and inference efficiency of large models in automotive aftermarket repair scenarios through domain incremental fine-tuning and paginated attention deployment; and achieve end-to-end business closed loop through message queues, sliding windows, feedback loops, and permission auditing.

[0087] In summary, the diagnostic knowledge graph construction and application method based on a large model in the above embodiments of the present invention improves the accuracy of knowledge parsing of repair manuals, circuit diagrams, and technical bulletins by uniformly acquiring and processing unstructured repair documents, structured repair records, and vehicle-side fault data in the automotive aftermarket repair field. It utilizes a visual-semantic cascaded parsing method to extract layout structure, semantic entities, entity relationships, and circuit topology information from repair documents. Through knowledge alignment and collaborative storage of graph / vector dual databases, it eliminates the differences in the multiple names of objects such as parts and fault phenomena, achieving a combination of structured relationship querying and semantic generalized retrieval. Furthermore, through query-based... The adaptive fusion of semantic and textual retrieval by type can cater to both precise query needs such as fault codes and part numbers, and fuzzy query needs such as technicians' colloquial fault descriptions, improving the recall and accuracy of diagnostic results. By extracting common diagnostic nodes through concurrent diagnostic analysis of multiple fault codes and dynamically adjusting the recommended order of diagnostic steps based on historical diagnostic frequency and user feedback, the system can reduce redundant troubleshooting and improve fault location efficiency. At the same time, the system continuously updates the knowledge graph relationship weights and model training data through diagnostic feedback, enabling the diagnostic knowledge base and large model capabilities to be continuously optimized with actual repair data, thereby improving the intelligence level and troubleshooting efficiency of automotive after-sales repair diagnosis.

[0088] Example 2 The second embodiment of this application also provides a diagnostic knowledge graph construction and application system based on a large model. This system is used to implement the embodiments and preferred embodiments described herein, and will not be repeated hereafter. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0089] like Figure 2As shown, the system includes: a data acquisition module 100, a visual-semantic parsing module 200, an output module 300, and an update module 400.

[0090] The data acquisition module 100 is used to acquire multi-source maintenance data in the field of automotive aftermarket repair, including unstructured maintenance documents, structured maintenance records, and vehicle-side fault data. The visual-semantic parsing module 200 is used to perform visual-semantic concatenated parsing on the unstructured maintenance document to obtain structured knowledge data containing entities, attributes, relationships, coordinate anchors and parsing confidence. Based on the structured knowledge data and the structured maintenance record, knowledge alignment is performed to obtain standardized entity nodes and standardized relationship edges. The standardized entity nodes and standardized relationship edges are stored in the graph database, and the semantic vectors and metadata corresponding to the standardized entity nodes are stored in the vector database. The output module 300 is used to obtain diagnostic query information based on the vehicle fault data, perform semantic retrieval and text retrieval on the diagnostic query information respectively, determine the fusion weight of semantic retrieval score and text retrieval score based on query type, calculate the comprehensive score of candidate diagnostic results, and output the target diagnostic result. The update module 400 is used to dynamically recommend diagnostic steps to the user terminal based on the target diagnostic results, and update the historical diagnostic frequency and / or graph relationship weight of the corresponding diagnostic steps according to the user terminal feedback results, so as to form a closed-loop update of the diagnostic knowledge graph.

[0091] It should be noted that the modules can be functional modules or program modules, and can be implemented in software or hardware. For modules implemented in hardware, the modules can reside in the same processor; or the modules can be located in different processors in any combination.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for constructing and applying a diagnostic knowledge graph based on a large model, characterized in that, Includes the following steps: Acquire multi-source repair data in the automotive aftermarket repair field, including unstructured repair documents, structured repair records, and vehicle-side fault data; The unstructured maintenance documents are subjected to visual-semantic concatenated parsing to obtain structured knowledge data; Based on the vehicle-side fault data, diagnostic query information is obtained. Semantic retrieval and text retrieval are performed on the diagnostic query information respectively. The fusion weight of semantic retrieval score and text retrieval score is determined based on the query type. The comprehensive score of candidate diagnostic results is calculated to output the target diagnostic result. Based on the target diagnostic results, diagnostic steps are dynamically recommended to the user, and the historical diagnostic frequency and / or graph relationship weights of the corresponding diagnostic steps are updated according to the user feedback results to form a closed-loop update of the diagnostic knowledge graph.

2. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, The visual-semantic cascaded parsing includes extracting the layout structure information of the unstructured maintenance document through a first parsing channel, extracting the semantic entities and entity relationships of the unstructured maintenance document through a second parsing channel, and then performing confidence alignment and result fusion based on the layout structure information and the semantic entities and entity relationships.

3. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 2, characterized in that, The steps for performing visual-semantic concatenation parsing on the unstructured maintenance document include: The first parsing channel identifies the bounding box coordinates, reading order, and table row and column structure of text blocks, tables, illustrations, and / or formulas in the unstructured maintenance document, thereby obtaining a page structure tree with spatial location attributes. The second parsing channel identifies the reference relationships of maintenance terms, the causal relationships between fault phenomena and test standards, the component symbols in the circuit diagram, the wiring harness routing, and the electrical connection relationships in the unstructured maintenance document, and obtains entity relationship triples with semantic tags. Using the bounding box coordinates as spatial reference anchor points, the entity relationship triples are projected onto the corresponding physical regions; When the first parsing channel and the second parsing channel have the same parsing result for the same physical region, the corresponding parsing confidence scores are superimposed or weighted and fused to generate high-weight structured nodes. When the parsing results of the first parsing channel and the second parsing channel conflict for the same physical region, or when the confidence level of either parsing channel is lower than the preset confidence threshold, cross-validation is triggered, and semantic completion or coordinate correction is performed based on the high-confidence parsing result and the complementary features of the other parsing channel.

4. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 3, characterized in that, The steps for performing visual-semantic concatenated parsing on the unstructured maintenance documents also include: Select the differentiated parsing subprocess based on the content type in the unstructured maintenance document; When the content type is normal text, semantic blocks are segmented according to paragraph logic and chapter hierarchy to construct a text index tree; When the content type is a symptom table, a mapping relationship between the fault phenomenon, possible causes and troubleshooting suggestions is established based on the table row and column coordinates and column semantic alignment results. When the content type is precise test content, extract the closed-loop test relationship between test conditions, measuring instruments, standard values ​​and judgment logic; When the content type is a circuit diagram, component nodes and electrical connection edges are generated based on the coordinates of graphic elements, component symbol types, wiring harness directions, and electrical control logic.

5. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, The steps for knowledge alignment based on the structured knowledge data and the structured maintenance records include: The names of the parts to be aligned are encoded into semantic vectors, and the cosine similarity between the names of the parts to be aligned and the existing parts nodes is calculated. Supervised training data containing query statements, positive sample sets, and negative sample sets is constructed based on historical fault description data, and a semantic similarity model is trained using the binary cross-entropy loss function. The fault description to be aligned is encoded into a fault description semantic vector. Based on the semantic similarity model, fault descriptions with semantic similarity greater than the second similarity threshold are aggregated to generate standardized fault phenomenon nodes, and the mapping relationship between the standardized fault phenomenon nodes and the original fault description is retained. By using a domain-adapted large language model, structured maintenance records with incomplete fault descriptions and / or missing fault locations are used to summarize fault phenomena, correct typos, extract fault locations, and classify fault causes, forming a hierarchical knowledge structure.

6. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, The steps of performing semantic and textual retrieval on the diagnostic query information according to the query type, and calculating the comprehensive score of the candidate diagnostic results, include: Extract at least one of vehicle identification information, fault codes, and natural language fault phenomenon descriptions from the diagnostic query information, and mark the query type as fault code precise query, fault phenomenon semantic query, or mixed query. The natural language description of the fault phenomenon is converted into a query semantic vector, and the semantic similarity between the query semantic vector and the document semantic vector is calculated in the vector database to obtain the semantic retrieval result. Keyword extraction and word form normalization are performed on the diagnostic query information. Exact matching, partial matching and / or fuzzy matching are performed in the full-text index of the graph database, and text retrieval results are obtained based on text relevance scores. The semantic retrieval results and the text retrieval results are mapped to a unified graph node identifier space, and duplicate candidate nodes are deduplicated and their scores are retained. Based on the query type, basic semantic weights and basic text weights are determined to calculate the comprehensive score of the candidate diagnostic results.

7. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, When the diagnostic query information includes multiple fault codes, the step of performing concurrent diagnostic analysis on the multiple fault codes based on the graph database includes: Using multiple fault codes as a combined query key, historical repair cases containing multiple fault codes are retrieved from the graph database. Calculate the case matching degree between historical maintenance cases and current diagnostic query information; When there are historical repair cases whose case matching degree reaches the preset matching conditions, the diagnostic path corresponding to the historical repair cases will be output first. When there are no historical repair cases that meet the preset matching conditions, standard diagnostic paths are constructed for each fault code to obtain common diagnostic steps that include the intersection nodes of multiple standard diagnostic paths. The common diagnostic steps are prioritized based on their frequency of occurrence in multiple standard diagnostic paths and / or their upstream position in the graph topology.

8. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, Based on the target diagnostic results, the steps for dynamically recommending diagnostic steps to the user include: Load the standard diagnostic topology tree corresponding to the target fault object from the graph database. The standard diagnostic topology tree includes possible cause nodes, test condition nodes, measurement standard nodes, and troubleshooting step nodes. Query the historical successful resolution frequency of each diagnostic step in the standard diagnostic topology tree, and perform initial sorting of diagnostic steps at the same level according to the historical successful resolution frequency; Determine the first-hop node and second-hop node of the current diagnostic node, and calculate the comprehensive recommendation weight of the first-hop node; Push a preset number of diagnostic steps ranked first by comprehensive recommendation weight to the user's terminal, and dynamically load the next hop node after the user confirms or clicks on the terminal. When the user reports that the diagnostic steps have resolved the fault, the historical diagnostic frequency of the diagnostic steps is incremented in the positive direction.

9. The method for constructing and applying a diagnostic knowledge graph based on a large model according to claim 1, characterized in that, The method further includes: The multi-source maintenance data is accessed through a message queue, and the unstructured maintenance document parsing results are written into the knowledge construction queue, the structured maintenance record alignment results are written into the knowledge update queue, and the vehicle-side real-time fault code stream is written into the diagnostic trigger queue. Sliding time window aggregation is performed on the real-time fault code stream corresponding to the same vehicle identification information to deduplicate and align timestamps of repeatedly reported fault codes. When at least two fault codes or a new fault code is detected within the sliding time window, multi-fault code concurrent diagnostic analysis is triggered, and diagnostic suggestions are pushed to the user terminal.

10. A diagnostic knowledge graph construction and application system based on a large model, characterized in that, The method for constructing and applying a diagnostic knowledge graph based on a large model as described in any one of claims 1-9 includes: The data acquisition module is used to acquire multi-source maintenance data in the field of automotive aftermarket repair, including unstructured maintenance documents, structured maintenance records, and vehicle-side fault data. The visual-semantic parsing module is used to perform visual-semantic concatenated parsing on the unstructured maintenance documents to obtain structured knowledge data; The output module is used to obtain diagnostic query information based on the vehicle fault data, perform semantic retrieval and text retrieval on the diagnostic query information respectively, determine the fusion weight of semantic retrieval score and text retrieval score based on query type, calculate the comprehensive score of candidate diagnostic results, and output the target diagnostic result. The update module is used to dynamically recommend diagnostic steps to the user based on the target diagnostic results, and update the historical diagnostic frequency and / or graph relationship weight of the corresponding diagnostic steps according to the user feedback results, so as to form a closed-loop update of the diagnostic knowledge graph.

Images