Automobile intelligent diagnosis method based on knowledge enhancement large language model

Abstract

The invention discloses an automobile intelligent diagnosis method based on a knowledge-enhanced large language model, and the method is characterized in that the method comprises the following steps: S1, vehicle data collection and preprocessing: collecting a fault code set, a sensor signal, an ECU log, a CAN message and a user natural language description in real time through a vehicle bus and a diagnosis interface, performing denoising, normalization and time synchronization preprocessing on the acquired data to form a diagnosis input set Xt = {D, St, Lt}, D being a fault code set, St being a sensor signal data stream, and Lt being a language description of a user or a maintenance technician; s2, knowledge base construction and management: constructing a unified automobile knowledge graph G = (V, E) and a text knowledge base K; according to the automobile intelligent diagnosis method based on the knowledge enhancement large language model, semantic level diagnosis capability is improved, semantic association of fault codes, sensor data and natural language description can be fused, deep understanding of cross-modal data is achieved, and the problem of data isolation of a traditional model is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle diagnostics, specifically to an intelligent vehicle diagnostic method based on a knowledge-enhanced large language model. Background Technology

[0002] With the increasing intelligence of automobiles, modern vehicles integrate a large number of electronic control units (ECUs), sensors, and on-board diagnostic systems (OBD-II). Fault types are becoming increasingly complex, and traditional diagnostic methods are no longer sufficient to meet the demands for efficient and accurate diagnosis. Currently, mainstream automotive diagnostic technologies are mainly divided into two categories: rule-based fault code analysis systems (DTC-based) and model- or expert-based diagnostic systems. In recent years, although deep learning technology has been attempted to be applied to on-board data analysis, traditional models are limited by the heterogeneity of vehicle data sources (fault codes, sensors, logs) and the diversity of fault scenarios. Cross-system semantic fusion results in diagnostic conclusions lacking interpretability and failing to absorb dynamic knowledge such as new repair cases and recall information in real time. Existing methods suffer from the following shortcomings: isolated knowledge and lack of semantic integration: repair manuals, fault code descriptions, and sensor data are stored in a scattered manner, and the model cannot establish a logical chain of "fault symptoms-component association-repair solution" at the semantic level; rigid rules and weak generalization ability: diagnostic logic relies on predefined rules or static knowledge bases, which cannot adapt to changes in the electronic control system of new vehicle models or dynamic fault scenarios (such as multi-component cascading faults). There is an urgent need for an intelligent automotive diagnostic method based on a knowledge-enhanced large language model to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide a vehicle intelligent diagnostic method based on a knowledge-enhanced large language model to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a vehicle intelligent diagnostic method based on a knowledge-enhanced large language model, comprising the following steps: S1 Vehicle Data Acquisition and Preprocessing: Real-time acquisition of fault code sets, sensor signals, ECU logs, CAN messages, and user natural language descriptions via vehicle bus and diagnostic interface. The acquired data is preprocessed by noise reduction, normalization, and time synchronization to form a diagnostic input set Xt={D,St,Lt}, where D is the fault code set, St is the sensor signal data stream, and Lt is the user's or technician's language description. S2 Knowledge Base Construction and Management: Construct a unified automotive knowledge graph G=(V,E) and a text knowledge base K. The nodes V of the knowledge graph include parts, fault codes, symptoms, repair steps, and cases. The edges E represent the semantic relationships of "cause and effect", "applicable models", and "solutions". The text knowledge base K includes repair manuals, technical bulletins, recall records, manufacturer repair experience, and semantic index documents. Knowledge storage adopts a vectorized approach. S3 Semantic Retrieval and Knowledge Enhancement: The diagnostic input set Xt is processed by vector embedding to generate a query vector q. The similarity between the query vector and the knowledge base vector is calculated. The k knowledge fragments with the highest similarity are selected to form the enhanced context Caug=[Xt;Ktopk], where Ktopk is the k knowledge fragments with the highest similarity. S4 Large Language Model Diagnostic Reasoning: The enhanced context Caug is input into the large language model that has been fine-tuned for the automotive domain. Through the reasoning chain of "input → semantic understanding → knowledge retrieval → causal reasoning → fault inference → suggestion generation", multi-step reasoning is performed, and the results of fault cause identification, location of affected components, suggested repair steps and possible risk prediction are output. S5 Result Generation and Interpretation: Transforms inference results into structured output and outputs natural language interpretation text through templates or large language model generators; S6 Feedback Optimization and Knowledge Update: Based on the maintenance results or manual confirmation tags, the weights of the corresponding cases in the knowledge base are updated using the weight update formula W_i←W_i+α・(1-error). At the same time, incremental training is performed on the semantic retrieval model to achieve knowledge self-enhancement and model optimization.

[0005] Preferably, the sensor signals in step S1 include temperature, speed, oil pressure, and voltage signals, and the fault codes are OBD standard fault codes, including preset fault code types such as P0301 and P0420.

[0006] Preferably, the knowledge graph construction process in step S2 includes entity extraction, relation extraction and semantic alignment of structured data in the automotive field, and the vectorized storage of the text knowledge base adopts the Transformer class model to realize text feature extraction and vector transformation.

[0007] Preferably, the vector embedding process in step S3 is implemented using the encoding layer of a pre-trained language model, and the similarity calculation uses the cosine similarity algorithm, with the value of k ranging from 5 to 20.

[0008] Preferably, the large language model fine-tuned in the automotive domain in step S4 is selected from the GPT, Qwen, and LLaMA series models. The fine-tuning process uses automotive fault diagnosis datasets, repair case datasets, and manufacturer technical document datasets for supervised training.

[0009] Preferably, the risk prediction results in step S4 include the risk of fault propagation, the level of safety hazards, and the estimated range of maintenance costs.

[0010] Preferably, the incremental training in step S6 adopts an online fine-tuning method, the training data is valid diagnostic cases verified by maintenance, the underlying parameters of the model are fixed during the training process, and only the model parameters of the semantic retrieval layer and the inference layer are updated.

[0011] Preferably, the structured output in step S5 is in JSON format and includes fields for fault code, fault level, list of affected components, maintenance priority, and estimated maintenance time.

[0012] Preferably, in step S2, the text knowledge base also includes real-time updated vehicle recall information and fault handling experience documents from third-party repair organizations, with an update frequency of once a day.

[0013] Preferably, in step S3, after selecting the k knowledge fragments with the highest similarity, the knowledge fragments are further subjected to semantic deduplication processing to remove redundant and conflicting knowledge content before forming an enhanced context.

[0014] Compared with existing technologies, this solution designs a vehicle intelligent diagnostic method based on a knowledge-enhanced large language model, which has the following beneficial effects: (1) Enhanced semantic-level diagnostic capabilities: It can integrate the semantic associations of fault codes, sensor data, and natural language descriptions to achieve deep understanding of cross-modal data and solve the problem of "data isolation" in traditional models; (2) High diagnostic accuracy: By introducing a professional knowledge base through the retrieval enhancement generation (RAG) mechanism, it can be adapted to different vehicle models and complex fault scenarios, and the accuracy is improved by more than 30% compared with the traditional rule system; (3) The results are highly interpretable: the reasoning basis is output in natural language form, and the maintenance technicians can trace the diagnostic logic, reducing the reliance on "expert experience"; (4) Self-learning ability: New cases and new TSB information are absorbed in real time through the feedback mechanism, eliminating the need for frequent manual rule updates and extending the system life cycle; (5) User-friendly interaction: Supports natural language input and output, allowing technicians to describe faults directly in colloquial language, reducing the operational threshold. Attached Figure Description

[0015] Figure 1 This is a system overall framework diagram of the present invention; Figure 2 This is a system workflow diagram of the present invention. Detailed Implementation

[0016] Please see Figure 1 and Figure 2The present invention provides a technical solution: a vehicle intelligent diagnostic method based on a knowledge-enhanced large language model, comprising the following steps: S1 Vehicle Data Acquisition and Preprocessing: Real-time acquisition of fault code sets, sensor signals, ECU logs, CAN messages, and user natural language descriptions via vehicle bus and diagnostic interface. The acquired data undergoes noise reduction, normalization, and time synchronization preprocessing to form a diagnostic input set Xt={D,St,Lt}, where D is the fault code set, St is the sensor signal data stream, and Lt is the user's or technician's natural language description. S2 Knowledge Base Construction and Management: Construction of a unified automotive knowledge graph G=(V,E) and a text knowledge base K. Nodes V in the knowledge graph include parts, fault codes, symptoms, repair steps, and cases. Edges E represent semantic relationships such as "cause and effect," "applicable vehicle models," and "solutions." The text knowledge base K includes repair manuals, technical bulletins, recall records, manufacturer repair experience, and semantic index documents. Knowledge storage uses a vectorized approach. S3 Semantic Retrieval and Knowledge Enhancement: Vector embedding processing is applied to the diagnostic input set Xt to generate a query vector q. The query vector is then calculated. Based on the similarity with the knowledge base vectors, the k knowledge fragments with the highest similarity are selected to form the enhanced context Caug=[Xt;Ktopk], where Ktopk is the k knowledge fragments with the highest similarity; S4 Large Language Model Diagnostic Reasoning: The enhanced context Caug is input into the large language model fine-tuned for the automotive domain, and multi-step reasoning is performed through the reasoning chain of "input → semantic understanding → knowledge retrieval → causal reasoning → fault inference → suggestion generation", outputting the fault cause identification result, the location result of the affected parts, the suggested repair steps, and the possible risk prediction result; S5 Result Generation and Interpretation: The reasoning result is converted into structured output, and natural language explanation text is output through templates or large language model generators; S6 Feedback Optimization and Knowledge Update: Based on the repair result or manual confirmation label, the weight of the corresponding case in the knowledge base is updated using the weight update formula W_i←W_i+α・(1-error), and incremental training is performed on the semantic retrieval model to achieve knowledge self-enhancement and model optimization.

[0017] The sensor signals mentioned in step S1 include temperature, speed, oil pressure, and voltage signals. The fault codes are OBD standard fault codes, including preset fault code types such as P0301 and P0420.

[0018] The knowledge graph construction process described in step S2 includes entity extraction, relation extraction, and semantic alignment of structured data in the automotive field. The vectorized storage of the text knowledge base adopts the Transformer-type model to realize text feature extraction and vector transformation.

[0019] The vector embedding process described in step S3 is implemented using the encoding layer of a pre-trained language model, and the similarity calculation uses the cosine similarity algorithm, with the value of k ranging from 5 to 20.

[0020] The large language model fine-tuned in the automotive domain in step S4 is selected from the GPT, Qwen, and LLaMA series models. The fine-tuning process uses automotive fault diagnosis datasets, repair case datasets, and manufacturer technical document datasets for supervised training.

[0021] The risk prediction results mentioned in step S4 include the risk of fault propagation, the level of safety hazards, and the estimated range of maintenance costs.

[0022] The incremental training described in step S6 adopts an online fine-tuning method. The training data consists of valid diagnostic cases that have been verified through maintenance. During the training process, the underlying parameters of the model are fixed, and only the model parameters of the semantic retrieval layer and the inference layer are updated.

[0023] The structured output in step S5 uses JSON format and includes fields for fault code, fault level, list of affected parts, maintenance priority, and estimated maintenance time.

[0024] In step S2, the Chinese text knowledge base also contains real-time updated vehicle recall information and fault handling experience documents from third-party repair organizations, updated daily.

[0025] Step S3, after selecting the k knowledge fragments with the highest similarity, also includes semantic deduplication of the knowledge fragments, removing redundant and conflicting knowledge content before forming an enhanced context.

[0026] Example 1: Professional Repair and Diagnosis of Passenger Vehicles at 4S Dealerships 1. Application Scenarios A 4S store of a certain brand received a 2023 family sedan. The owner reported that "the engine vibrates at idle and acceleration is weak". The technician needs to quickly locate the fault and generate a repair plan.

[0027] 2. Implementation Steps (1) Vehicle data collection and preprocessing Data collection: The fault code D={P0301} (cylinder 1 misfire) was read by the OBD diagnostic tool. The sensor signal St includes the idle speed (fluctuation range 800-950r / min, standard value 850±50r / min) and the ignition coil voltage of cylinder 1 (11.2V, standard value 12.0±0.5V). The technician entered the natural language description Lt=“Idle speed vibration, delayed power response during acceleration, no history of fault light flashing”. Preprocessing: Filter out the 1200r / min abnormal value of instantaneous fluctuation in the speed signal, normalize the voltage data to "V" unit, and synchronize the P0301 fault code trigger time (30s after idle start) with the sensor data timing to form Xt={P0301,[speed: 800-950r / min, voltage: 11.2V],"idle vibration...no history of fault light flashing"}.

[0028] (2) Knowledge base construction and management Knowledge Graph G: Nodes include “P0301”, “Cylinder 1 ignition coil”, “Engine vibration”, “Replace ignition coil”, and “2023 XX sedan”, with edge relationships of “P0301-Cause-Cylinder 1 ignition coil aging” and “Replace ignition coil-Applicable model-2023 XX sedan”. Text knowledge base K: Stores the vehicle's repair manual (Chapter 5 "Ignition System Fault Diagnosis"), manufacturer's TSB (number TSB-2024-005, mentioning "2023 model 1 cylinder ignition coil is prone to aging, resistance value needs to be checked"), 100+ similar misfire fault cases. All texts have been converted into semantic vectors using the BERT model.

[0029] (3) Semantic retrieval and knowledge enhancement Generate query vector q: Use a pre-trained BERT model to encode Xt to capture the semantic association of "P0301 + idling vibration + low voltage"; Similarity calculation: By comparing the vectors in q and K with cosine similarity, the 10 knowledge fragments with the highest similarity k are selected (including TSB-2024-005, 3 P0301 fault cases of 2023 models, and the core content of Chapter 5 of the repair manual). Construct the enhanced context Caug=[Xt;TSB-2024-005,Case 1-3,Maintenance Manual Section 5.1,...].

[0030] (4) Diagnostic reasoning based on large language models Model selection: The GPT-4 model, which has been fine-tuned with fault data of this brand's vehicles (the fine-tuning data includes 5000+ repair cases of this brand); Reasoning process: The model identifies "P0301 + low voltage" as the core feature through semantic understanding, retrieves TSB and cases in Caug, and infers that "the aging of the ignition coil in cylinder 1 leads to insufficient voltage, causing misfire." The affected components are located as "the ignition coil and spark plug in cylinder 1." The repair steps are generated as follows: "1. Remove the engine cover and take out the ignition coil in cylinder 1; 2. Use a multimeter to test the primary resistance of the coil (standard value 0.5-0.8Ω); 3. If the resistance is outside the range, replace it with the original ignition coil; 4. After installation, start the engine and verify the idle stability." The risk prediction is "If continued use occurs, it may lead to carbon buildup on the spark plug in cylinder 1, increasing repair costs."

[0031] (5) Results generation and interpretation Structured output (JSON format): Fault code: "P0301", Fault Level: "Medium Risk" "Affected Components":["Ignition coil for cylinder 1","Spark plug for cylinder 1"], "Repair Priority": "High" Estimated repair time: 30 minutes Natural Language Explanation: "Based on fault code P0301 (cylinder 1 misfire), idle speed fluctuation (800-950r / min), low ignition coil voltage in cylinder 1 (11.2V), combined with manufacturer's TSB-2024-005 (2023 model cylinder 1 ignition coil is prone to aging) and 3 similar cases, the cause of the fault is inferred to be the aging of the cylinder 1 ignition coil; repair should refer to section 5.1 of the vehicle's repair manual, first check the coil resistance, if it exceeds the standard value of 0.5-0.8Ω, then replace it; failure to repair in time may lead to spark plug carbon buildup, which will require additional cleaning later." (6) Feedback optimization and knowledge update Repair verification: The technician measured the coil resistance at 1.2Ω (exceeding the standard). After replacement, the idle speed stabilized (840r / min), and the fault was eliminated. Weight update: The weights of TSB-2024-005 and Cases 1-3 in Caug are adjusted by W_i←W_i+0.2×(1-0) (α=0.2, error=0) to improve their subsequent retrieval priority; Incremental training: Using this case as training data, the semantic retrieval model is fine-tuned to make the model more accurately match the knowledge fragment "2023 model + P0301 + low voltage".

[0032] Example 2: Remote Fault Monitoring and Diagnosis for Commercial Vehicle Fleets 1. Application Scenarios A logistics company manages a fleet of 50 heavy trucks and needs to monitor vehicle malfunctions in real time through a remote system to provide early warnings and guide drivers / repair stations to handle the issues and avoid breakdowns on the road.

[0033] 2. Implementation Steps (1) Vehicle data collection and preprocessing Data collection: Vehicle data is acquired in real time through the fleet remote monitoring platform (TSP). Fault code D={P0420} (three-way catalytic converter efficiency is below the threshold). Sensor signals St include exhaust temperature (450℃, standard value 350-400℃) and engine load (80%, normal driving load 60%-70%). The system automatically generates the description Lt=“Truck reports P0420 while driving, exhaust temperature is too high, engine load is abnormal”. Preprocessing: Filter the instantaneous peak value of 550℃ in the exhaust temperature signal (caused by rapid acceleration), normalize the engine load to "%" unit, and synchronize the P0420 trigger time (after 2 hours of driving) with the sensor data to form Xt={P0420,[exhaust temperature: 450℃, load: 80%],"P0420 reported during driving...load abnormality"}.

[0034] (2) Knowledge base construction and management Knowledge Graph G: Nodes include “P0420”, “three-way catalytic converter”, “high exhaust temperature”, “catalytic converter cleaning”, and “heavy truck”, with edge relationships of “P0420-causality-three-way catalytic converter carbon deposit” and “catalytic converter cleaning-applicable models-heavy truck”. Text knowledge base K: Stores heavy truck repair manuals (Chapter 8 "Aftertreatment System"), manufacturer recall records (catalytic converters in a certain batch in 2022 were prone to carbon buildup), fleet history P0420 failure cases (30+), and text vectors are converted by the lightweight model DistilBERT (adapted to the computing power of remote platforms).

[0035] (3) Semantic retrieval and knowledge enhancement Generate query vector q: Use DistilBERT encoding Xt to capture the semantics of "P0420 + high exhaust temperature + commercial vehicle"; Similarity calculation: k=8 knowledge fragments were selected by cosine similarity (including recall records, 5 similar cases from different fleets, and section 8.3 "Catalyst Cleaning Process" in the maintenance manual). Construct the enhanced context Caug=[Xt;Recall record, Cases 1-5, Repair Manual Section 8.3, ...].

[0036] (4) Diagnostic reasoning based on large language models Model selection: The lightweight Qwen-7B model was adopted (adapted to the low computing power requirements of remote platforms), and fine-tuned with commercial vehicle fault data (including 2000+ truck cases). Reasoning process: The model identifies "P0420 + high exhaust temperature" as a characteristic of catalytic converter carbon buildup. Combining this with recall records, it infers that "the three-way catalytic converter is experiencing reduced efficiency due to carbon buildup caused by poor fuel quality." The affected components are identified as "three-way catalytic converter and oxygen sensor." The model generates the following repair recommendations: "1. Go to the nearest repair shop to clean the three-way catalytic converter; 2. Check if the oxygen sensor signal is normal; 3. Refill with high-octane diesel fuel to avoid carbon buildup." The risk prediction is that "if driving continues for more than 500 kilometers, it may trigger torque limiting (engine power reduction), affecting transportation efficiency."

[0037] (5) Results generation and interpretation Structured output (JSON format): Fault code: "P0420" Fault Level: "Low-Medium Risk" "Affected Components": ["Three-way catalytic converter", "Oxygen sensor"], "Repair Priority": "Medium" Suggested processing time: Within 24 hours Natural language interpretation: "Based on fault code P0420 (low efficiency of three-way catalytic converter), high exhaust temperature (450℃, standard 350-400℃), abnormal engine load (80%), combined with the 2022 recall record for catalytic converters of this batch of trucks and similar cases from 5 different fleets, the cause of the fault is inferred to be carbon buildup in the catalytic converter (possibly related to low-octane diesel); the catalytic converter needs to be cleaned and the oxygen sensor checked at a repair shop within 24 hours; continued driving may trigger torque limiting, and long-distance transportation is recommended." (6) Feedback optimization and knowledge update Repair verification: After the driver went to the repair shop to clean the catalytic converter, P0420 was eliminated and the exhaust temperature returned to 380℃; Weight update: The weights of recall records and cases 1-5 are adjusted by W_i←W_i+0.15×(1-0) (α=0.15) to improve the retrieval priority in commercial vehicle scenarios; Incremental training: Fine-tune the semantic retrieval model using this case to optimize the vector matching accuracy of "commercial vehicle + P0420 + exhaust temperature" and adapt it to more truck models.

[0038] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0039] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent diagnosis of a vehicle based on a knowledge-enhanced large language model, characterized in that, Comprising the following steps: S1 Vehicle data collection and preprocessing: Collecting fault code sets, sensor signals, ECU logs, CAN messages and user natural language descriptions in real time through vehicle bus and diagnostic interface, denoising, normalizing and time synchronization preprocessing of the collected data to form a diagnostic input set Xt={D,St,Lt}, where D is the fault code set, St is the sensor signal data stream, and Lt is the language description of the user or maintenance technician; S2 Knowledge base construction and management: Constructing a unified automobile knowledge graph G=(V,E) and a text knowledge base K. The nodes V of the knowledge graph include parts, fault codes, symptoms, maintenance steps, and cases. The edges E represent semantic relationships such as "cause and effect", "applicable vehicle model", and "solution". The text knowledge base K includes maintenance manuals, technical bulletins, recall records, manufacturer maintenance experience, and semantic index documents. Knowledge storage uses vectorization method; S3 Semantic retrieval and knowledge enhancement: Perform vector embedding processing on the diagnostic input set Xt to generate query vector q, calculate the similarity between query vector and knowledge base vector, select the top k pieces of knowledge fragments with the highest similarity to form the enhanced context Caug=[Xt;Ktopk], where Ktopk is the top k pieces of knowledge fragments with the highest similarity; S4 Large language model diagnosis reasoning: Input the enhanced context Caug into the large language model fine-tuned in the automobile field, and perform multi-step reasoning through the reasoning chain of "input→semantic understanding→knowledge retrieval→causal reasoning→fault inference→suggestion generation", output the fault cause identification result, influence component positioning result, recommended maintenance step and possible risk prediction result; S6 Feedback optimization and knowledge update: According to the maintenance result or artificial confirmation label, update the weight of the corresponding case in the knowledge base using the weight update formula W_i←W_i+α・(1-error), and perform incremental training on the semantic retrieval model to realize knowledge self-enhancement and model optimization. The sensor signals in step S1 include temperature, speed, oil pressure, and voltage signals. The fault codes are OBD standard fault codes, including P0301, P0420, and other preset fault code types.

2. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. The construction process of the knowledge graph in step S2 includes entity extraction, relationship extraction and semantic alignment of structured data in the automobile field. The vectorization storage of the text knowledge base uses a Transformer-based model to realize text feature extraction and vector conversion.

3. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. The vector embedding processing in step S3 uses the encoding layer of a pre-trained language model. The similarity calculation uses the cosine similarity algorithm, and the value range of k is 5-20.

4. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. The large language model fine-tuned in the automobile field in step S4 is selected from GPT, Qwen, and LLaMA series models. The fine-tuning process uses automobile fault diagnosis data sets, maintenance case data sets, and manufacturer technical document data sets for supervised training.

5. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. The risk prediction result in step S4 includes fault diffusion risk, safety hazard level, and maintenance cost estimation range.

6. The method of claim 5, wherein the method further comprises: ​ 7. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. The incremental training in step S6 adopts an online fine-tuning manner, the training data is the valid diagnosis case verified by maintenance, and the model parameters of the bottom layer are fixed during the training process, and only the model parameters of the semantic retrieval layer and the reasoning layer are updated. 8.The method of claim 1, wherein the method further comprises: The structured output in step S5 adopts a JSON format, and contains a fault code, a fault level, an affected component list, a maintenance priority and a predicted maintenance time field. 9.The method of claim 1, wherein the method further comprises: The text knowledge base in step S2 also contains real-time updated automobile recall information and fault handling experience documents of third-party maintenance agencies, and the updating frequency is once a day.

10. The method of claim 1, wherein the method is based on a knowledge-enhanced large language model. After selecting the k pieces of knowledge segments with the highest similarity in step S3, the knowledge segments are subjected to semantic deduplication processing, and after removing redundant and conflicting knowledge contents, an enhanced context is formed.

Images