Automobile intelligent question and answer method and system based on large language model

Abstract

The present application relates to the technical field of intelligent question and answer, and more particularly to a car intelligent question and answer method and system based on a large language model. The method comprises the following steps: deploying a large language model enhanced with car field knowledge, and constructing a multi-source information fusion module; receiving a mixed query text to obtain current vehicle state parameters, user historical preference features and current dialogue session vectors for joint coding, and generating an enhanced query semantic vector; inputting the enhanced query semantic vector into the large language model enhanced with car field knowledge for multi-hop reasoning, to generate a candidate answer set and corresponding confidence scores; performing personalized sorting, content adaptation and instantiation replacement on the candidate answer set to generate executable answers; presenting the executable answers through a vehicle-mounted human-computer interaction interface in a multi-modal manner, and performing incremental reinforcement learning fine-tuning on the strategy network of the large language model. The present application can greatly improve the logicality of car intelligent question and answer.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, and in particular to an intelligent question answering method and system for automobiles based on a large language model. Background Technology

[0002] With the rapid iteration of automotive intelligence and connectivity technologies, the level of intelligence in in-vehicle human-machine interaction has become one of the core indicators for measuring the competitiveness of automotive products. Users' needs for automotive-related Q&A have evolved from simple information queries to comprehensive interactive demands that are professional, personalized, and actionable, covering multiple automotive scenarios such as vehicle maintenance, function operation, and driving behavior guidance. Traditional automotive Q&A methods often rely on fixed-script matching or simple keyword retrieval, which presents numerous challenges, including inaccurate semantic understanding, inability to integrate with real-time vehicle status, and a lack of targeted answers, making it difficult to meet users' diverse and in-depth interactive needs. Against this backdrop, intelligent Q&A systems that deeply integrate large language models with automotive domain knowledge have emerged. By introducing technologies such as domain knowledge enhancement and multi-source information fusion, they achieve accurate identification of question and answer intent and generate high-quality answers. Currently, many automakers and technology companies are investing in this field, launching various in-vehicle intelligent Q&A solutions, achieving significant results in improving user interaction experience and optimizing service efficiency.

[0003] In addition, a similar patent, CN121524196A, discloses a method and system for intelligent question answering in the automotive channel based on a large language model. This method includes: receiving a user's natural language query request; inputting the natural language query request into a large language model fine-tuned with data from the automotive channel domain, where the large language model obtains a structured function call instruction corresponding to the query request; parsing the structured function call instruction and extracting at least one business parameter; dynamically generating an access link for a front-end business page based on the business parameter; and returning a response to the user, the response including at least a natural language text answer based on the execution result of the structured function call instruction and an access link. This invention achieves real-time querying and analysis of automotive channel data through natural language interaction, simultaneously returning text answers and access links to the user during the query process. This breaks the limitation of traditional intelligent question answering systems that only provide text replies, improving the efficiency and business conversion rate of the question answering system. However, its description of the professionalism of the question answering, context awareness, personalized adaptation, and continuous model optimization is relatively simple, mainly focusing on the querying of automotive channel data and the generation of access links, without fully combining the professional knowledge characteristics of the automotive field and the advantages of multi-source data in the in-vehicle scenario. Summary of the Invention

[0004] To address the aforementioned technical problems in existing automotive intelligent question answering processes, this invention provides a method for automotive intelligent question answering based on a large language model. This method integrates automotive domain knowledge with multi-source information to understand mixed query intents and generate context-aware semantic vectors by combining real-time vehicle status and user preferences. It enhances the professionalism of answers through multi-hop reasoning based on knowledge graphs, outputs executable results through personalized sorting and parameter instantiation, and achieves incremental reinforcement learning of the model through user feedback. This effectively solves the problems of inaccurate semantic understanding, lack of real-time performance and personalization, and difficulty in continuous optimization in traditional question answering methods. The method includes the following steps: During the system initialization phase, a large language model with enhanced automotive domain knowledge is deployed, and a multi-source information fusion module is built, which includes a real-time vehicle status data interface, a user personalized profile library, and a multi-turn dialogue context set. In the query understanding and information fusion stage, the system receives mixed query text input by the user, which includes natural language and vehicle system instructions. The system obtains the current vehicle status parameters, user historical preference features and current dialogue session vector in real time through the multi-source information fusion module. The system then jointly encodes the mixed query text, vehicle status parameters, user historical preference features and current dialogue session vector to generate a context-aware enhanced query semantic vector. In the domain knowledge-guided reasoning stage, the enhanced query semantic vector is input into the large language model with automotive domain knowledge enhancement. The large language model performs multi-hop reasoning based on its internal structured knowledge graph of automotive repair, driving behavior, and functional configuration and the enhanced query semantic vector to generate a preliminary candidate answer set containing reasoning logic chains and corresponding confidence scores. In the answer generation and personalization adaptation stage, the candidate answer set is personalized and adapted based on the user's historical preference features. The candidate answer with the highest ranking is selected as the basic answer. Then, the specific operation suggestions or parameter descriptions in the basic answer are instantiated and replaced in combination with the vehicle status parameters to generate the final executable answer. In the answer presentation and feedback learning phase, the executable answer is presented in a multimodal manner through the in-vehicle human-machine interface, and the user's interaction feedback signals to the presented answer are monitored in real time. A reward value is calculated based on the interaction feedback signals and the executable answer, and the reward value is used to perform incremental reinforcement learning fine-tuning on the policy network of the large language model for automotive domain knowledge enhancement.

[0005] Preferably, the multi-source information fusion module includes a real-time vehicle data acquisition submodule, a dynamic user profile construction submodule, and a dialogue context management submodule; The real-time vehicle data acquisition submodule communicates with the vehicle sensor network via the vehicle bus to continuously acquire and standardize vehicle status parameters, including engine speed, fault codes, driving range, and the working status of each subsystem. The user profile dynamic construction submodule extracts and periodically updates the user's historical preference features by analyzing the user's historical Q&A records, in-vehicle function settings preferences, and driving behavior data. These features are vectorized to represent the user's tendencies in technical depth, expression style, and functional focus. The dialogue context management submodule is responsible for maintaining the dialogue history of the current session and encoding the question-answer pairs of the most recent N rounds of dialogue through a sequence model to generate the current dialogue session vector that represents the topic and depth of the current session.

[0006] Preferably, the joint encoding to generate context-aware enhanced query semantic vectors includes: The mixed query text is segmented and word vector embedded to obtain a query text vector sequence; The vehicle status parameters at the current moment are read from the real-time vehicle data acquisition submodule, and the relevance weight of each vehicle status parameter is calculated based on the query text vector sequence through a parameter importance evaluation network. The parameter values ​​are then weighted and fused based on the relevance weight to generate a contextualized vehicle status vector. The user's historical preference features are obtained from the user profile dynamic construction submodule, and the attention distribution of the features and the query text vector sequence in the latent semantic space is calculated. Based on the attention distribution, the user's historical preference features are selectively enhanced to generate a query-oriented user preference vector. Obtain the current dialogue session vector from the dialogue context management submodule; The query text vector sequence, the contextualized vehicle state vector, the query-oriented user preference vector, and the current dialogue session vector are concatenated to form a fusion feature tensor. The fused feature tensor is input into a cross-modal attention encoder, which interacts and aligns features from different sources through a multi-head attention mechanism, and finally outputs the context-aware enhanced query semantic vector.

[0007] Preferably, the process of calculating the relevance weights in the parameter importance evaluation network includes: The parameter importance evaluation network receives the query text vector sequence as input and extracts its globally pooled semantic summary vector. The semantic summary vector is matched with a learnable vehicle state parameter type embedding table to generate a preliminary type relevance score for each type of vehicle state parameter. Based on the type-related scores, the real-time collected original vehicle state parameter values ​​of the corresponding type are normalized. The normalized parameter values ​​are combined with the corresponding parameter type embeddings, and then a non-linear transformation layer is used to generate the final, fine-grained correlation weights for weighting.

[0008] Preferably, the generation of the final, fine-grained correlation weights used for weighting includes: The parameter importance evaluation network first performs self-attention operation on the query text vector sequence, focusing on key entities and verbs in the query, and extracting the core vector of query intent; Then, the cosine similarity between the core vector of the query intent and each parameter pattern vector in the predefined vehicle state parameter pattern library is calculated, where each parameter pattern vector is associated with one or a set of vehicle state parameters. The calculated cosine similarity is normalized using the softmax function to obtain the initial attention score for each parameter pattern; The initial attention score is multiplied by the data reliability coefficient of the corresponding parameter obtained in real time from the vehicle network. The data reliability coefficient is determined by the sensor health status and update frequency of the data source of the parameter to obtain the adjusted pattern correlation. Finally, based on the aforementioned pattern correlation degree and considering the prior dependency graph model between different types of parameters, a graph attention network is used for propagation and aggregation to output the fine-grained correlation weights that are ultimately used to weight the original parameter values.

[0009] Preferably, the personalized sorting and content adaptation of the candidate answer set based on the user's historical preference features includes: Analyze each candidate answer in the candidate answer set and extract its answer type label, technical depth level, and expression complexity index; The answer type label, technical depth level, and expression complexity index of each candidate answer are matched with the user's acceptance of each answer type, the preferred technical depth range, and the comprehensible expression complexity threshold recorded in the user's historical preference features to calculate the matching degree and generate a personalized adaptation score for each candidate answer. The personalized adaptation score is weighted and fused with the confidence score of the candidate answer set. The weights are dynamically adjusted according to whether the current dialogue session vector contains the initial consultation or the in-depth follow-up question, to obtain the comprehensive ranking score. The candidate answer set is sorted in descending order based on the comprehensive ranking score, and the candidate answer ranked first is selected as the base answer.

[0010] Preferably, the instantiation and replacement of specific operation suggestions or parameter descriptions in the basic answer based on vehicle state parameters includes: Semantic role labeling is performed on the basic answer to identify the abstract operation instructions, general parameter descriptions, and conditional statements contained therein; Construct an instantiated mapping rule library that defines the mapping relationships between various abstract operation instructions, general parameter descriptions and specific vehicle subsystems and readable real-time status parameters; Based on the instantiation mapping rule base, the abstract elements identified in the basic answer are traversed. If an abstract element matches a rule, the specific instance value is calculated or found based on the latest vehicle status parameters obtained from the vehicle data real-time acquisition submodule. Replace the corresponding abstract operation instructions or general parameter descriptions in the basic answer with the calculated specific instance values, and compare the assumptions in the conditional statements with the current actual state of the vehicle. If they are met, the suggestion is retained; if they are not met, the suggestion is marked as invalid or replaced with alternative suggestions. After replacing all instantiable elements and validating conditions, the final executable answer is generated.

[0011] Preferably, the construction and application of the instantiated mapping rule base includes: The instantiation mapping rule base consists of a manually defined initial rule set and a system-learned extended rule set; When performing instantiation replacement, the system first attempts to find a matching mapping rule in the manually defined initial rule set; If no perfectly matching rule is found in the initial rule set, a rule self-learning process is triggered. This process, based on the enhanced query semantic vector and the basic answer, calls the automotive domain knowledge-enhanced large language model to generate a possible mapping hypothesis. Next, the system simulates and verifies the logical rationality and safety of the mapping hypothesis in a safe sandbox environment based on vehicle state parameters. If the verification is successful, the mapping hypothesis is temporarily adopted for this instantiation replacement. After the interaction is successful, and following a manual review process, the mapping hypothesis may be selectively added to the system's self-learning extended rule set.

[0012] Preferably, the rule self-learning process includes: Separate the entity vector related to the operation object and the intent vector related to the operation target from the enhanced query semantic vector; The entity vector, intent vector, and abstract description fragment to be instantiated in the basic answer are input into a rule reasoning submodule. This submodule is a lightweight neural network used for path finding and reasoning on the structured knowledge subgraph generated by the large language model that enhances knowledge in the automotive domain. The rule reasoning submodule outputs one or more possible mapping paths from abstract descriptions to specific vehicle subsystems or parameters, along with their confidence levels. The mapping path with the highest confidence is selected, and combined with the specific value range and unit of the vehicle state parameters, it is transformed into a structured and executable mapping rule draft by the big language model enhanced with automotive domain knowledge, namely the mapping hypothesis. The draft mapping rules clearly define the triggering conditions, abstract description patterns, target subsystem or parameter identifiers, and instantiation value calculation methods.

[0013] As a preferred embodiment, the second technical solution of the present invention is a car intelligent question answering system based on a large language model, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the car intelligent question answering method based on a large language model as described above.

[0014] The present invention has the following specific beneficial effects: (1) During the system initialization phase, a large language model with enhanced automotive domain knowledge was deployed, and a multi-source information fusion module was constructed. This solved the problems of traditional in-vehicle question-answering systems lacking domain-specific support, being unable to coordinate the use of multi-source information, and being unable to adapt to the full-scenario question-answering needs of in-vehicle systems. The large language model with enhanced automotive domain knowledge integrates structured professional knowledge such as automotive maintenance, driving behavior, and functional configuration, breaking the limitations of the traditional model's generality and lack of domain specificity, and providing solid knowledge support for subsequent professional question-answering and reasoning. The multi-source information fusion module integrates the vehicle's real-time status data interface, the user's personalized archive, and the multi-turn dialogue context set, realizing centralized management and collaborative calling of multiple key information in in-vehicle scenarios, avoiding the drawbacks of fragmented information and lack of linkage in traditional systems. Compared with the traditional fixed dialogue or keyword retrieval mode, this step builds a professional and integrated question-answering infrastructure from the source, which not only makes up for the shortcomings of the general large language model in terms of automotive domain professionalism, but also provides a data channel for subsequent integration with vehicle real-time status and user preferences, effectively improving the professionalism and adaptability of in-vehicle intelligent question answering, and laying the foundation for meeting users' deep-level and full-scenario interaction needs.

[0015] (2) In the query understanding and information fusion stage, the system receives mixed query text from users and performs multi-source information joint encoding to generate context-aware enhanced query semantic vectors. This solves the problem of traditional question-answering systems having one-sided semantic understanding and being unable to combine multi-source scenario information, leading to biased intent recognition. This step breaks the limitation of traditional systems that only process single natural language queries and can effectively receive mixed query texts containing natural language and vehicle system instructions, adapting to diverse user input habits. Through the multi-source information fusion module, the system simultaneously acquires current vehicle status parameters, user historical preference features, and dialogue session vectors, and jointly encodes multiple types of information with mixed query texts. This fully integrates the vehicle's real-time operating conditions, user past habits, and dialogue context, allowing the semantic vectors to more comprehensively reflect the user's true query intent and avoiding the drawbacks of traditional keyword retrieval that ignores context and cannot be combined with the actual vehicle status. The generated enhanced query semantic vectors achieve deep analysis and scenario-based adaptation of query intent, effectively solving the problem of incomplete semantic understanding in traditional systems. This provides high-quality semantic support for subsequent domain knowledge reasoning and accurate answer generation, further enhancing the scenario adaptation capability of the question-answering system.

[0016] (3) In the domain knowledge-guided reasoning stage, the enhanced query semantic vector is input into the automotive domain knowledge-enhanced large language model. Through multi-hop reasoning, a set of candidate answers with reasoning logic chains and confidence scores are generated, solving the problems of traditional question-answering systems lacking professionalism and clear reasoning basis, and similar solutions being insufficient in terms of question-answering professionalism and context awareness. This large language model has a built-in structured automotive domain knowledge graph, which can combine enhanced query semantic vectors for multi-hop reasoning. It can not only generate candidate answers that meet user needs, but also output complete reasoning logic chains, making the answers traceable and breaking the limitations of traditional systems where answers lack basis and professionalism. At the same time, the generated confidence scores can quantify the reliability of each candidate answer, providing a clear basis for subsequent answer selection and avoiding the problems of ambiguous reasoning processes and inability to judge the reliability of answers in similar solutions. Compared with the traditional fixed-phrase matching, simple keyword retrieval, and the single-channel query mode of similar solutions, this step greatly improves the professionalism and logic of question answering, and can accurately respond to professional demands in multiple scenarios such as vehicle repair and functional operation, effectively making up for the shortcomings of existing solutions in the application of domain professional knowledge.

[0017] (4) In the answer generation and personalization adaptation stage, candidate answers are optimized by combining user preferences and vehicle status to generate executable answers. This solves the problems of traditional question-and-answer systems lacking personalization, failing to incorporate real-time vehicle status, and having insufficient personalization adaptation for similar solutions. This step personalizes and adapts the candidate answer set based on user historical preference characteristics, abandoning the monotonous pattern of traditional system answers and making the answers more in line with user habits and needs, thus improving the user interaction experience. At the same time, the operation suggestions and parameter descriptions in the basic answers are instantiated and replaced by combining the current vehicle status parameters, so that the answers can adapt to the current working conditions of the vehicle and have strong executableness, breaking the drawbacks of traditional system answers being abstract and unable to directly guide operations. Compared with the traditional question-and-answer system's lack of specificity in answers and simple personalization adaptation for similar solutions, this step achieves the integration of personalization and scenario-based answers, which not only meets the diverse interaction needs of users but also provides users with directly applicable operation guidance, effectively solving the problem of insufficient practicality of traditional system answers and further improving the user experience and practical value of in-vehicle intelligent question answering.

[0018] (5) In the answer presentation and feedback learning stage, the answer is presented in a multimodal manner, and the model is incrementally fine-tuned based on user feedback. This solves the problems of traditional question-answering systems, such as single answer presentation, inability to continuously optimize the model, and lack of continuous iteration capability in similar solutions. The multimodal presentation method breaks the limitations of traditional single text reply. The answer is presented in a more intuitive and easy-to-understand form through the in-vehicle human-machine interface, improving the efficiency and experience of users in obtaining information. At the same time, the user interaction feedback signal is monitored in real time, the reward value is calculated, and the model policy network is incrementally fine-tuned, realizing the continuous optimization of the model. This allows the model to continuously adapt to changes in user interaction habits and supplement domain knowledge gaps, breaking the drawbacks of traditional system models being fixed and unable to be iteratively upgraded. Compared with the traditional question-answering system's lack of feedback mechanism and the similar solutions' lack of emphasis on continuous model optimization, this step constructs a complete closed loop of "presentation-feedback-optimization". This not only improves the intuitiveness of answer presentation, but also allows the model to continuously improve the quality of question and answer in long-term use, continuously adapt to users' deep and diverse interaction needs, and further enhance the competitiveness and practicality of the in-vehicle intelligent question-answering system. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of the intelligent question-answering method for automobiles based on a large language model, as described in this invention. Figure 2 This is a schematic diagram of the module connection of the multi-source information fusion module in this invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0021] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for intelligent question answering in automobiles based on a large language model, such as... Figure 1 As shown, it includes the following steps: S01: In the system initialization phase, deploy a large language model with knowledge enhancement in the automotive field, and build a multi-source information fusion module that includes a real-time vehicle status data interface, a user personalized profile library, and a multi-turn dialogue context set; In this embodiment of the invention, a large language model augmented with automotive domain knowledge is deployed to an in-vehicle computing unit via a hardware deployment unit. The in-vehicle computing unit utilizes the NVIDIA Jetson AGX Orin embedded AI computing platform, featuring a 2048-core Ampere architecture GPU and a 12-core ARM Cortex-A78AE CPU, equipped with 32GB of LPDDR5 video memory, meeting the computing power and storage requirements for large language model inference. The deployment process includes the following steps: First, the trained model parameter file (approximately 13GB) is downloaded to the solid-state drive of the in-vehicle computing unit via OTA; then, the model is quantized and compressed using the TensorRT framework, converting FP32 parameters to FP16 precision, reducing the model size to approximately 6.5GB; finally, the model is loaded into the GPU video memory, and after loading, a forward inference verification is performed (the input is a preset test query text "engine malfunction light on"). Once the model output is confirmed to be normal, the deployment is complete, with a total deployment time of approximately 45 seconds. The large language model is based on an open-source base model (such as LLaMA-2-7B or ChatGLM-6B) for domain adaptation. The specific construction process includes: (1) Construction of a structured knowledge system: From the automotive repair manual, on-board diagnostic standard (OBD-II), vehicle electronic control unit (ECU) communication protocol, user manual and after-sales fault case library, the system structure knowledge, on-board fault mechanism knowledge, vehicle operation and maintenance specification knowledge and on-board electronic control logic knowledge are extracted by rule-based entity relation extraction method (such as dependency parsing and pattern matching). The system is stored in a graph database (such as Neo4j) in the form of triples (entity-relation-entity) to construct a structured knowledge graph containing more than 15,000 nodes and 30,000 relations, and is classified and indexed according to 8 subsystems such as power system, braking system and comfort system. (2) Model Domain Weight Fine-tuning: A two-stage fine-tuning strategy was adopted. The first stage was domain pre-training, using approximately 5 million Chinese automotive domain corpora collected from automotive forums, repair manuals, and technical documents, and continuing pre-training using the Masked Language Model (MLM) task. The second stage was instruction fine-tuning, constructing approximately 200,000 automotive question-and-answer instruction pairs, covering scenarios such as fault diagnosis, function operation, and driving guidance. The LoRA (Low-Rank Adaptation) method was adopted, with the loss function being cross-entropy loss, the learning rate set to 5e-5, the number of training rounds being 3, and the batch size being 16, ultimately generating a large language model with enhanced automotive domain knowledge. This large language model embeds a structured knowledge graph related to automotive repair, driving behavior, and function configuration, covering 10,000 core automotive domain knowledge points, which are stored in eight subsystems, including the power system, braking system, and comfort system. Each knowledge point is associated with specific operational logic and parameter ranges.Simultaneously, a multi-source information fusion module is constructed, which includes a real-time vehicle status data interface, a user personalized profile library, and a multi-turn dialogue context set. The vehicle real-time status data interface establishes a fixed connection with the vehicle sensor network via the vehicle bus. This interface uses a combination of the CAN controller chip MCP2515 and the CAN transceiver TJA1050, communicating with the main control chip via the SPI protocol. The communication rate is set to 500kbps, and the data refresh frequency is uniformly set to 100Hz. This frequency is determined based on the typical communication rate of the vehicle CAN bus and the physical characteristics of the sensor signals, ensuring the real-time performance of high-frequency signals such as engine speed without overloading the bus. It is used to receive various vehicle status parameters in real time. The user personalized profile library is built using a hardware storage unit. This storage unit is a 256MB NOR Flash chip, connected to the main control chip via a Quad SPI interface. It is divided into 1000 storage blocks, each 256KB in size, corresponding to one user. A wear leveling algorithm is used to extend its service life. It stores the user's historical question and answer records, in-vehicle function settings preferences, and driving behavior data for the past 3 months. The multi-turn dialogue context set is implemented through a 64MB session cache unit. The SRAM chip connects to the main control chip via a parallel bus and employs a ring buffer data structure. The cache capacity is set to 50 rounds of dialogue. Each round of dialogue stores the semantic vectors and timestamps of the user query and system response, ensuring that all modules can work together normally after initialization. After initialization, the system performs a self-test process: sequentially checking the large language model inference response time (should be less than 500ms), CAN bus communication status (should have no packet loss), Flash storage read / write correctness (checksum matching), and SRAM cache available space (should be greater than 80%). Once all checks pass, an initialization success flag is output, and the system enters standby mode.

[0022] S02: Query understanding and information fusion stage. The system receives a mixed query text input by the user, which contains natural language and vehicle system instructions. The system obtains the current vehicle status parameters, user historical preference features and current dialogue session vector in real time through the multi-source information fusion module. The system then jointly encodes the mixed query text, vehicle status parameters, user historical preference features and current dialogue session vector to generate a context-aware enhanced query semantic vector. In this embodiment of the invention, the in-vehicle human-machine interface receives mixed query text input by the user. The mixed query text contains natural language and in-vehicle system commands, such as "Engine speed 6500 rpm is too high, how to reduce the speed and check the fault," where "engine speed 6500 rpm" is the in-vehicle system command, and the rest is natural language. The mixed query text first passes through a command parsing circuit, which identifies the in-vehicle system command format through regular expression matching. The regular expression pattern is "". The successfully matched portions are marked as system commands, and their parameter names and values ​​are extracted. The unmatched portions are marked as natural language. The extracted system command parameter name "engine speed" and value "6500" are temporarily stored in the command register for subsequent encoding. The multi-source information fusion module collects current vehicle status parameters at 100ms intervals through the vehicle's real-time status data interface, including engine speed 6500rpm, ignition voltage 11.8V, driving range 420km, and the operating status of each subsystem. The collected parameters are converted into standardized data after isolation amplification and level shifting. The module also reads the user's historical preference features from the user's personalized profile database. These features are iteratively generated fixed 64-dimensional vectors. The selection of the 64 dimensions is based on an empirical balance principle, which can effectively distinguish the user's tendencies across different technical depths, expression styles, and functional focuses. To avoid the storage and computational overhead of excessively high dimensionality, the current conversation is generated by reading the first three rounds of dialogue records from a multi-turn dialogue context set. The 128-dimensional vector is chosen based on the information capacity requirements of the dialogue context encoding. For the most recent five rounds, approximately 25 dimensions are encoded per round, with a small amount of redundant dimensions used to mark the dialogue stage, totaling 128 dimensions. This approach fully preserves the dialogue topic and depth information while avoiding dimensional redundancy. The mixed query text is split using a word segmentation circuit and processed by word vector embedding. The word segmentation circuit employs a dictionary-based maximum forward matching algorithm, containing 100,000 Chinese words and 5,000 automotive-specific terms, with a segmentation speed of at least 1000 characters per second. Each segmented word is converted into a 32-dimensional vector by a word vector embedding circuit, which incorporates a look-up table. The system (Table) stores 32-dimensional embedding vectors corresponding to 100,000 words, which can be directly read via word ID index, with a single lookup taking no more than 10 ns. A 6×32 query text vector sequence is generated, and this vector sequence, along with vehicle state parameters, user historical preference features, and the current dialogue session vector, is input into a joint encoding circuit. This joint encoding circuit is implemented using an FPGA and integrates eight parallel attention computation units, each containing an independent matrix multiplier and softmax computation module, supporting pipelined operation with a single encoding computation latency of no more than 2 ms. Feature interaction and alignment are performed through a multi-head attention mechanism to generate a 256-dimensional context-aware enhanced query semantic vector, completing information fusion and encoding.

[0023] S03: Domain knowledge-guided reasoning stage, the enhanced query semantic vector is input into the automotive domain knowledge-enhanced large language model, the large language model performs multi-hop reasoning based on its internal structured automotive repair, driving behavior, and functional configuration knowledge graph and the enhanced query semantic vector, and generates a preliminary candidate answer set containing reasoning logic chains and corresponding confidence scores; In this embodiment of the invention, a 256-dimensional enhanced query semantic vector is input into a large language model for automotive domain knowledge enhancement. The structured knowledge graph within this large language model contains three main categories of related knowledge: automotive repair, driving behavior, and functional configuration. Each node in the knowledge graph corresponds to a knowledge point, and connections are established between nodes through association relationships. The knowledge graph is stored in an in-memory database in a graph-structured data format, employing a combination of adjacency lists and hash indexes to support efficient node and neighbor queries. Multi-hop inference employs a path search and scoring mechanism based on Graph Attention Network (GAT): First, the core entity in the enhanced query semantic vector (e.g., "engine speed too high") is used as the starting node; then, a breadth-first search (BFS) is performed in the knowledge graph, expanding neighboring nodes layer by layer, with each layer expansion considered a hop; in each hop, the semantic correlation between the current node and its neighbors is calculated using the attention mechanism, and paths with correlation higher than a threshold (e.g., 0.6) are selected for further expansion; the hop count is limited to a maximum of 3 hops to prevent semantic drift caused by excessively long inference paths; finally, the node sequences and relation sequences in all reachable paths are combined to form multiple candidate inference chains, which are then sorted based on the confidence of the paths, and the top K candidate answers and their corresponding inference logic chains are output. The inference logic chains are output in JSON format, containing the following fields: chain_id (inference chain number), nodes (a list of nodes, each containing node_name and node_type), edges (a list of edges, each containing source_node, target_node, and relation_type), and confidence (confidence score of the path). For example, the output of the reasoning logic chain for candidate answer 1 is: {"chain_id": "001", "nodes":[{"name":"Engine speed too high","type":"symptom"}, {"name":"Throttle body carbon buildup","type":"cause"}, {"name":"Check throttle body","type":"operation"}], "edges": [{"from":"Engine speed too high","to":"Throttle body carbon buildup","relation":"may be caused by..."}, {"from":"Throttle body carbon buildup","to":"Check throttle body","relation":"suggestion..."}], "confidence": 0.85}.The large language model performs multi-hop inference based on the knowledge graph and enhanced query semantic vectors. First, it extracts the core requirement "engine speed too high" from the enhanced query semantic vectors. The first hop inference associates common causes of high engine speed, including throttle body carbon buildup, abnormal ignition voltage, and insufficient fuel supply. The second hop infers the troubleshooting logic for each cause, and the third hop infers specific operational suggestions for each troubleshooting logic. Finally, it generates a candidate answer set containing three candidate answers and their corresponding confidence scores. The confidence score for each candidate answer is obtained by weighting two parts: inference path confidence (weight 0.7) and answer text fluency (weight 0.3). The inference path confidence is obtained by multiplying the attention scores of all points in the path, and the answer text fluency is calculated using the average log probability during decoding by the large language model. Candidate answer 1 is "High engine speed may be due to carbon buildup in the throttle body. It is recommended to check and clean the throttle body, and then retest the speed after cleaning," with a confidence level of 0.85; Candidate answer 2 is "Abnormal ignition voltage causes high speed. It is recommended to check the ignition system and adjust the ignition voltage to the standard range of 12V," with a confidence level of 0.8; Candidate answer 3 is "Insufficient fuel supply causes high speed. It is recommended to check the fuel line, add fuel, and try again," with a confidence level of 0.75. Each candidate answer contains a complete logical chain of reasoning.

[0024] S04: In the answer generation and personalized adaptation stage, the candidate answer set is personalized and adapted based on the user's historical preference features. The candidate answer with the highest ranking is selected as the basic answer. Then, the specific operation suggestions or parameter descriptions in the basic answer are instantiated and replaced in combination with the vehicle status parameters to generate the final executable answer. In this embodiment of the invention, the candidate answer set is personalized and content-adapted based on user historical preference features. Each candidate answer is parsed using a semantic parsing circuit, extracting answer type tags, technical depth level, and expression complexity index. This semantic parsing circuit is based on a lightweight Transformer model implemented on FPGA, containing a 2-layer encoder and a 1-layer classification head, with approximately 5MB of model parameters and a single parsing latency of no more than 1ms. Candidate answer 1 has the type tag of fault handling, basic technical depth level, and expression complexity index of 0.6; candidate answer 2 has the type tag of fault handling, advanced technical depth level, and expression complexity index of 0.8; candidate answer 3 has the type tag of fault handling, basic technical depth level, and expression complexity index of 0.5. Calculate the matching degree between each candidate answer and the user's preference. The matching degree calculation formula is: Matching degree = 0.4 × Type acceptance + 0.3 × Technical depth fit + 0.3 × Expression complexity fit. The user's acceptance degree for troubleshooting answers is 0.9, the basic level technical depth fit is 1, the advanced level fit is 0, and the expression complexity ≤ 0.7 fit is 1. The calculated matching degree of candidate answer 1 is 0.4 × 0.9 + 0.3 × 1 + 0.3 × 1 = 0.96, the matching degree of candidate answer 2 is 0.4 × 0.9 + 0.3 × 0 + 0.3 × 0 = 0.36, and the matching degree of candidate answer 3 is 0.4 × 0.9 + 0.3 × 1 + 0.3 × 0.71 = 0.87. The matching degree and the confidence of candidate answers are weighted and combined. The current dialogue is the initial consultation, and the weights are set to matching degree 0.6 and confidence degree 0.4. The comprehensive ranking score = 0.6 × matching degree + 0.4 × confidence degree. The comprehensive score of candidate answer 1 = 0.6 × 0.96 + 0.4 × 0.85 = 0.916, candidate answer 2 = 0.6 × 0.36 + 0.4 × 0.8 = 0.576, and candidate answer 3 = 0.6 × 0.87 + 0.4 × 0.75 = 0.802. The candidate answer 1 with the highest ranking is selected as the base answer. The basic answer is instantiated and replaced by combining vehicle status parameters. The abstract operation command "check throttle body and clean" in the basic answer is identified. Based on the instantiation mapping rule library, the engine speed and throttle opening parameters are associated. The current throttle opening is 85%, and the standard opening is 60%. The carbon deposit level is calculated as (85%-60%) / 60%×100%≈41.7%. The abstract command is replaced with "Check throttle body (current carbon deposit level is about 41.7%), it is recommended to clean it immediately, and re-check the engine speed after cleaning", generating the final executable answer.During the instantiation and replacement process, if a vehicle status parameter fails to be read or its value exceeds a reasonable range (e.g., throttle opening value is -5%), the system activates a fault-tolerant mechanism: First, it attempts to reread the parameter, retrying a maximum of 3 times with 10ms intervals; if it still fails, it replaces the parameter with its default value (a standard value provided by the vehicle manufacturer), and adds a note at the end of the answer: "[This parameter is calculated based on the default value; it is recommended to verify the actual vehicle status]"; if the default value is also unavailable, the instantiation of the parameter is skipped, and the original abstract description is retained. When multiple abstract elements in the basic answer need to be replaced, they are processed in the following order: first, replace the operation instructions, then replace the parameter descriptions, and finally verify the conditional statements to ensure the consistency of the replacement logic.

[0025] S05: Answer presentation and feedback learning stage, the executable answer is presented in a multimodal manner through the in-vehicle human-machine interface, and the user's interaction feedback signal to the presented answer is monitored in real time. A reward value is calculated based on the interaction feedback signal and the executable answer, and the reward value is used to perform incremental reinforcement learning fine-tuning on the policy network of the large language model for automotive domain knowledge enhancement.

[0026] In this embodiment of the invention, the final executable answer is presented in a multimodal manner through the in-vehicle human-machine interface. The interface simultaneously displays text, voice broadcast, and simple graphic prompts. The coordination rules for multimodal presentation are as follows: the text display is located in the central area of ​​the central control screen, with a font size of 24px, a white color, and a semi-transparent black background. The display duration matches the user's reading speed (default 8 seconds, adjustable according to user settings); the voice broadcast is output through the in-vehicle audio system, with a speech rate of 240 words per minute and a volume consistent with the current media volume. The media volume is automatically reduced to 30% at the start of the broadcast and restored after the broadcast ends; the graphic prompts are displayed in the lower right corner of the screen, using simple line-drawn icons (such as engine icons and throttle icons), with an icon size of 64×64 pixels, corresponding to the key operations in the text content, and the icons remain constantly lit during text display. All three modalities are presented simultaneously, but the voice broadcast takes precedence over the text display, and the graphic prompts flash once when the corresponding operation is reached during the voice broadcast. Real-time monitoring of user interaction feedback signals includes three scenarios: user click of the confirmation button, voice feedback indicating "valid" or "invalid," and no operation. Voice feedback is collected via an in-vehicle microphone array and processed by Voice Activity Detection (VAD) and Automatic Speech Recognition (ASR) modules to obtain text. The ASR module employs an end-to-end model, achieving a recognition accuracy of no less than 95% in an in-vehicle environment, with a false wake-up rate of less than 1 time per hour. The monitoring cycle is 5 seconds. The reward value is calculated based on the interactive feedback signal and the executable answer. The reward value calculation formula is: Reward Value = 0.7 × Feedback Score + 0.3 × Answer Suitability. If the user clicks the confirmation button or provides voice feedback "Valid", the feedback score is 1. The answer suitability is calculated by the correlation between the executable answer and the vehicle status parameters. Currently, the correlation is 0.9, and the reward value is 0.7 × 1 + 0.3 × 0.9 = 0.97. If the user provides voice feedback "Invalid", the feedback score is 0, the answer suitability is 0.9, and the reward value is 0.7 × 0 + 0.3 × 0.9 = 0.27. If there is no operation, the feedback score is 0.5, and the reward value is 0.7 × 0.5 + 0.3 × 0.9 = 0.62. The calculated reward value is input into the policy network of the large language model for automotive domain knowledge enhancement. Incremental reinforcement learning is used to fine-tune the network parameters. During the fine-tuning process, the network weights are adjusted using the gradient descent algorithm. The weight adjustment amount = reward value × learning rate. The learning rate is set to 0.01. For example, when the reward value is 0.97, the weight adjustment amount = 0.97 × 0.01 = 0.0097. After the fine-tuning is completed, the model parameters are updated and stored. The persistent storage and version management mechanism of the model parameters is as follows: After each fine-tuning, the updated policy network parameters (only about 5% of the total number of parameters) are stored in differential form in a dedicated partition of the vehicle solid-state drive. At the same time, a version number (in the format YYYYMMDD_HHMMSS_seq) is generated, and the fine-tuning time, reward value, and query text summary that triggered the fine-tuning are recorded.The system retains snapshots of the parameters from the most recent 100 versions. When a new version is written, causing insufficient storage space, the oldest version is automatically deleted. Upon the next system startup, the latest version of the parameters is loaded, and users can roll back to any historical version via the diagnostic interface. This is used for reasoning and answer generation for subsequent user queries.

[0027] Furthermore, the multi-source information fusion module includes a real-time vehicle data acquisition submodule, a dynamic user profile construction submodule, and a dialogue context management submodule; The real-time vehicle data acquisition submodule communicates with the vehicle sensor network via the vehicle bus to continuously acquire and standardize vehicle status parameters, including engine speed, fault codes, driving range, and the working status of each subsystem. The user profile dynamic construction submodule extracts and periodically updates the user's historical preference features by analyzing the user's historical Q&A records, in-vehicle function settings preferences, and driving behavior data. These features are vectorized to represent the user's tendencies in technical depth, expression style, and functional focus. The dialogue context management submodule is responsible for maintaining the dialogue history of the current session and encoding the question-answer pairs of the most recent N rounds of dialogue through a sequence model to generate the current dialogue session vector that represents the topic and depth of the current session.

[0028] In this embodiment of the invention, the multi-source information fusion module consists of a real-time vehicle data acquisition submodule, a dynamic user profile construction submodule, and a dialogue context management submodule. These three submodules work collaboratively to provide multi-dimensional fusion information for subsequent semantic encoding. The three submodules interact via a shared data bus and follow the following collaborative working mechanism: (1) Data interaction interface: The vehicle data real-time acquisition submodule connects to the vehicle network through the CAN bus driver interface and writes the standardized vehicle status parameters into the shared cache in the form of key-value pairs. The key is the parameter type identifier (such as "engine_speed") and the value is a normalized floating-point number. The user profile dynamic construction submodule communicates with the hardware storage unit through the I2C interface and writes the user's historical preference feature vector into the user feature storage area. The address offset is calculated by the user ID hash. The dialogue context management submodule communicates with the session cache unit through the SPI interface and writes the current dialogue session vector into the tail of the circular queue of the session cache area.

[0029] (2) Timing Coordination: The vehicle data real-time acquisition submodule continuously updates the parameter values ​​in the shared cache at a frequency of 100Hz; the user profile dynamic construction submodule triggers an update every 7 days, during which write permissions to the user feature storage area are locked, and read operations are not affected; the dialogue context management submodule updates the session cache immediately after each round of dialogue, with an update time of no more than 10ms. When the joint encoding circuit receives a new mixed query text, it reads the latest data from the three storage areas simultaneously through the data bus to ensure the timing consistency of multi-source information.

[0030] (3) Data consistency guarantee: The shared buffer adopts a double buffering mechanism. The real-time vehicle data acquisition submodule writes to the backup buffer, and the joint encoding circuit reads from the main buffer. The buffer pointer is switched through atomic operations to avoid read-write conflicts. The user feature storage area and session buffer adopt a read-write lock mechanism, which allows multiple read operations to be performed concurrently, and write operations have exclusive locks to ensure data consistency.

[0031] The vehicle data real-time acquisition submodule establishes stable communication with the vehicle sensor network via the vehicle CAN bus. The communication rate is set to 500kbps, and the data refresh frequency is uniformly set to 100Hz. This frequency is determined based on the typical communication rate of the vehicle CAN bus and the physical characteristics of the sensor signals. This ensures the real-time performance of high-frequency signals such as engine speed without overloading the bus. It continuously acquires various status parameters during vehicle operation, including engine speed, fault codes, driving range, and the operating status of each subsystem. The engine speed acquisition cycle is 100ms, the fault code acquisition cycle is 1s, the driving range acquisition cycle is 5s, and the operating status of each subsystem covers the operating status of the power system, braking system, and steering system. The collected raw parameters are standardized by the hardware processing unit. Engine speed is mapped from 0-8000rpm to a normalized range of 0-1. Fault codes are converted into numerical data according to preset coding rules. Driving range is mapped from 0-600km to a normalized range of 0-1. The operating status of each subsystem is represented by 0 for normal and 1 for abnormal. The processed data is stored in the module's internal cache for later use. The user profile dynamic construction submodule reads the user's historical Q&A records, in-vehicle function settings preferences, and driving behavior data from the past 3 months through the hardware storage unit. The historical Q&A records include nearly 100 user inquiries. In-vehicle function settings preferences include default air conditioning temperature, default audio volume, and seat adjustment position. Driving behavior data includes average vehicle speed, braking frequency, and throttle control amplitude. The data is analyzed using a feature extraction circuit to extract user preferences in terms of technical depth, expression style, and functional focus. Technical depth is divided into three dimensions: basic operation, advanced settings, and troubleshooting. Expression style is divided into two dimensions: concise and straightforward, and detailed inquiry. Functional focus is divided into three dimensions: power performance, comfort configuration, and troubleshooting. The extracted features are vectorized to generate a fixed 64-dimensional user historical preference feature vector. The selection of 64 dimensions is based on the principle of empirical balance, which can effectively distinguish user preferences in different technical depths, expression styles, and functional focuses, while avoiding the storage and computational overhead caused by excessive dimensionality. This feature is periodically updated every 7 days to ensure that the feature is consistent with the latest user preferences.The dialogue context management submodule maintains the dialogue history of the current session through the session caching unit, setting N=5, that is, retaining the question-and-answer pairs of the most recent 5 rounds of dialogue. These 5 rounds of question-and-answer pairs are encoded through a sequence encoding circuit. During the encoding process, semantics are extracted from each round of question-and-answer pairs, and then the extracted semantic information is converted into a vector through linear transformation, finally generating a 128-dimensional current dialogue session vector. The choice of 128 dimensions is based on the information capacity requirements of dialogue context encoding. For the most recent 5 rounds of dialogue, each round is encoded with about 25 dimensions, plus a small number of redundant dimensions for marking the dialogue stage (initial consultation / deep follow-up question), for a total of 128 dimensions. This can not only fully retain the dialogue topic and depth information, but also avoid dimensional redundancy. This vector is updated in real time with the generation of each new round of dialogue.

[0032] Furthermore, the joint encoding to generate context-aware enhanced query semantic vectors includes: The mixed query text is segmented and word vector embedded to obtain a query text vector sequence; The vehicle status parameters at the current moment are read from the real-time vehicle data acquisition submodule, and the relevance weight of each vehicle status parameter is calculated based on the query text vector sequence through a parameter importance evaluation network. The parameter values ​​are then weighted and fused based on the relevance weight to generate a contextualized vehicle status vector. The user's historical preference features are obtained from the user profile dynamic construction submodule, and the attention distribution of the features and the query text vector sequence in the latent semantic space is calculated. Based on the attention distribution, the user's historical preference features are selectively enhanced to generate a query-oriented user preference vector. Obtain the current dialogue session vector from the dialogue context management submodule; The query text vector sequence, the contextualized vehicle state vector, the query-oriented user preference vector, and the current dialogue session vector are concatenated to form a fusion feature tensor. The fused feature tensor is input into a cross-modal attention encoder, which interacts and aligns features from different sources through a multi-head attention mechanism, and finally outputs the context-aware enhanced query semantic vector.

[0033] In this embodiment of the invention, joint encoding generates context-aware enhanced query semantic vectors. First, the mixed query text input by the user is processed. The mixed query text is a car-related question asked by the user, such as "What is the reason for the engine speed being too high and the driving range decreasing?" The text is segmented by a word segmentation circuit, splitting it into six words: "engine," "speed," "too high," "driving range," "decreasing," and "reason." Then, each word is embedded by a word vector embedding circuit, and each word is converted into a 32-dimensional word vector, forming a 6×32 query text vector sequence. Subsequently, the vehicle status parameters at the current moment are read from the internal cache of the real-time vehicle data acquisition submodule, including the current engine speed of 6500 rpm (normalized to 0.8125), fault code 0001 (numerically encoded as 1), driving range of 420 km (normalized to 0.7), power system status 0, braking system status 0, and steering system status 0. These parameters are then input into the parameter importance evaluation network. The parameter importance assessment network calculates the relevance weights of each vehicle state parameter based on the query text vector sequence. The weights are: engine speed 0.4, fault code 0.3, driving range 0.2, and the operating status of each subsystem 0.033. The parameter values ​​are then weighted and fused based on these relevance weights. The calculation process is: 0.8125×0.4+1×0.3+0.7×0.2+0×0.033+0×0.033+0×0.033=0.325+0.3+0.14=0.765, generating a 128-dimensional contextualized vehicle state vector. The system obtains 128-dimensional user historical preference features from the user profile dynamic construction submodule. An attention calculation circuit calculates the attention distribution of these features and the query text vector sequence in the latent semantic space. The attention weights for "engine," "RPM," and "range" in the query text are 0.3, 0.25, and 0.2, respectively. Based on this attention distribution, the user historical preference features are selectively enhanced, strengthening dimensions related to power performance and fault handling, generating a 128-dimensional query-oriented user preference vector. The system also obtains a 64-dimensional current dialogue session vector from the dialogue context management submodule. The query text vector sequence, contextualized vehicle state vector, query-oriented user preference vector, and current dialogue session vector are concatenated. The query text vector sequence is flattened to 192 dimensions, and the concatenated vectors form a fusion feature tensor of 192 + 128 + 128 + 64 = 512 dimensions. The 512 dimensions of the fused feature tensor consist of four parts: the first part, 192 dimensions, comes from the semantic information of the query text, representing the content of the user's current question; the second part, 128 dimensions, comes from vehicle state parameters, representing the real-time operating condition of the vehicle; the third part, 128 dimensions, comes from user preferences, representing the user's personalized needs; and the fourth part, 64 dimensions, comes from the dialogue context, representing the current stage of the conversation. The different dimensions of each part reflect the differences in information capacity of the different source features.The fused feature tensor is input into a cross-modal attention encoder containing eight attention heads. The number of heads is chosen based on an empirical rule: in Transformer models, the number of heads is usually set to a divisor of the feature dimension and a power of 2, i.e., 512 / 8=64. Each attention head processes a 64-dimensional subspace, enabling the capture of feature interactions across different semantic subspaces while maintaining computational efficiency. The multi-head attention mechanism facilitates interaction and alignment of features from different sources. Each attention head extracts semantics from the fused feature tensor in different dimensions. The outputs of the eight attention heads are then concatenated, and a linear transformation layer converts the concatenated features into a 256-dimensional vector. Finally, a context-aware enhanced query semantic vector is output, providing accurate semantic support for subsequent large-scale language model question answering.

[0034] Furthermore, the process of calculating the relevance weights in the parameter importance evaluation network includes: The parameter importance evaluation network receives the query text vector sequence as input and extracts its globally pooled semantic summary vector. The semantic summary vector is matched with a learnable vehicle state parameter type embedding table to generate a preliminary type relevance score for each type of vehicle state parameter. Based on the type-related scores, the real-time collected original vehicle state parameter values ​​of the corresponding type are normalized. The normalized parameter values ​​are combined with the corresponding parameter type embeddings, and then a non-linear transformation layer is used to generate the final, fine-grained correlation weights for weighting.

[0035] In this embodiment of the invention, the process of calculating relevance weights by the parameter importance assessment network is executed step-by-step by the hardware computing circuit. First, the parameter importance assessment network receives a sequence of query text vectors as input. This sequence is a 6×32 matrix. The matrix is ​​then subjected to global average pooling through a global pooling circuit. The calculation process involves averaging the 32-dimensional vectors in each column to generate a 32-dimensional semantic summary vector, which represents the core semantics of the query text. The 32-dimensional semantic summary vector is then matched with a learnable vehicle state parameter type embedding table. The vehicle state parameter type embedding table includes six parameter types: engine speed, fault code, driving range, powertrain status, braking system status, and steering system status. Each type corresponds to a 32-dimensional embedding vector. The embedding table is randomly initialized, and each dimension of each embedding vector is sampled from a normal distribution with a mean of 0 and a standard deviation of 0.1. The embedding table is learned through backpropagation during the offline training phase of the system. The training data consists of 100,000 manually labeled query-parameter association samples. The loss function is cross-entropy loss, the learning rate is 1e-3, and the training epochs are 20. After training, the parameters of the embedding table are fixed and are not updated during the online inference phase. The similarity between the semantic summary vector and the embedding vector of each parameter type is calculated using vector dot product operation. The dot product operation formula is: similarity = semantic summary vector · parameter type embedding vector. The calculated similarity is 0.8 for engine speed, 0.7 for fault code, 0.6 for driving range, and 0.2 for the operating status of each subsystem. These similarities are used as the preliminary type relevance scores for each type of vehicle state parameter. The original vehicle status parameters collected in real time are normalized based on the type-related scores. The normalization formula is: Normalized value = (Original value - Minimum value) / (Maximum value - Minimum value). The original engine speed is 6500 rpm, the minimum is 0 rpm, and the maximum is 8000 rpm, so the normalized value is 6500 / 8000 = 0.8125. The original fault code is 0001, the minimum is 0000, and the maximum is 0010, so the normalized value is 1 / 10 = 0.1. The original driving range is 420km, the minimum is 0km, and the maximum is 600km, so the normalized value is 420 / 600 = 0.7. The original value of the operating status of each subsystem is 0, and the normalized value is 0. The normalized parameter values ​​are combined with the corresponding parameter type embeddings using vector concatenation. The normalized values ​​are converted into 32-dimensional vectors and then concatenated with the 32-dimensional embedding vectors of the corresponding parameter types to generate a 64-dimensional combined vector. This 64-dimensional combined vector is then input into a nonlinear transformation layer, which consists of two fully connected network layers. The first layer maps the 64-dimensional input to 128 dimensions with the ReLU activation function; the second layer maps the 128-dimensional input to 32 dimensions with the Tanh activation function, and the output range is [-1, 1].Through this nonlinear transformation, the network can learn the nonlinear interaction between parameter values ​​and parameter types, generating more accurate fine-grained weights. Ultimately, six fine-grained relevance weights are generated: engine speed 0.4, fault code 0.3, driving range 0.2, powertrain status 0.033, braking system status 0.033, and steering system status 0.033. These weights are used for subsequent weighted fusion of vehicle status parameters.

[0036] Furthermore, the generation of the final, fine-grained relevance weights for weighting includes: The parameter importance evaluation network first performs self-attention operation on the query text vector sequence, focusing on key entities and verbs in the query, and extracting the core vector of query intent; Then, the cosine similarity between the core vector of the query intent and each parameter pattern vector in the predefined vehicle state parameter pattern library is calculated, where each parameter pattern vector is associated with one or a set of vehicle state parameters. The calculated cosine similarity is normalized using the softmax function to obtain the initial attention score for each parameter pattern; The initial attention score is multiplied by the data reliability coefficient of the corresponding parameter obtained in real time from the vehicle network. The data reliability coefficient is determined by the sensor health status and update frequency of the data source of the parameter to obtain the adjusted pattern correlation. Finally, based on the aforementioned pattern correlation degree and considering the prior dependency graph model between different types of parameters, a graph attention network is used for propagation and aggregation to output the fine-grained correlation weights that are ultimately used to weight the original parameter values.

[0037] In this embodiment of the invention, the final fine-grained relevance weights used for weighting are generated and executed step-by-step by the hardware computing circuit of the parameter importance evaluation network. The parameter importance evaluation network first receives a sequence of query text vectors, which is a 6×32 matrix. The matrix is ​​then subjected to self-attention operation by a self-attention operation circuit. The self-attention operation focuses on key entities and verbs in the query by calculating the correlation between each word segmentation vector and all word segmentation vectors. For example, in the query text "What are the reasons for excessive engine speed and decreased driving range?", "engine", "speed", and "driving range" are key entities, and "excessively high" and "decreased" are key verbs. The semantic information of these key elements is extracted, and a 32-dimensional query intent core vector is generated through linear transformation. This vector accurately represents the core needs of the query. Then, the cosine similarity between the 32-dimensional query intent core vector and the vector of each parameter pattern in the predefined vehicle state parameter pattern library is calculated. The vehicle state parameter pattern library contains 5 parameter patterns, each corresponding to a set of vehicle state parameters: engine-related parameter patterns, range-related parameter patterns, fault-related parameter patterns, braking-related parameter patterns, and steering-related parameter patterns. Each parameter pattern vector is 32-dimensional. The pattern library is constructed as follows: first, automotive experts define the set of parameter types included in each pattern according to vehicle subsystems; then, the frequency of co-occurrence of each parameter type is counted from a large number of historical query logs, and parameter types with a co-occurrence frequency higher than 0.3 are grouped into the same pattern; finally, a 32-dimensional prototype vector is generated for each pattern, which is the mean of the embedding vectors of all parameter types in that pattern. The cosine similarity calculation formula is: Similarity = (Query Intent Core Vector · Parameter Pattern Vector) / (||Query Intent Core Vector|| × ||Parameter Pattern Vector||). The calculated similarity for engine-related parameters is 0.85, for range-related parameters 0.78, for fault-related parameters 0.72, for braking-related parameters 0.2, and for steering-related parameters 0.18. The calculated cosine similarities are then normalized using the softmax function. The normalization formula is: Normalized Score = e^Similarity / Σe^All Similarities. The calculated initial attention scores for engine-related parameters are 0.38, for range-related parameters 0.31, for fault-related parameters 0.25, for braking-related parameters 0.03, and for steering-related parameters 0.03.The initial attention score is multiplied by the data reliability coefficient of the corresponding parameter obtained in real time from the vehicle network. The data reliability coefficient is determined by the health status and update frequency of the sensor from which the parameter's data comes. The engine speed sensor is in good health with an update frequency of 100ms and a reliability coefficient of 0.95; the range sensor is in good health with an update frequency of 5s and a reliability coefficient of 0.9; the fault code sensor is in good health with an update frequency of 1s and a reliability coefficient of 0.92; and the braking and steering system parameter sensors are in good health with an update frequency of 1s and both have a reliability coefficient of 0.9. The adjusted mode correlation is calculated as follows: engine correlation 0.38×0.95=0.361, range correlation 0.31×0.9=0.279, fault correlation 0.25×0.92=0.23, braking correlation 0.03×0.9=0.027, and steering correlation 0.03×0.9=0.027. Finally, based on the adjusted pattern correlation, combined with the prior dependency graph model between different types of parameters, this graph model is a directed acyclic graph (DAG) containing 6 nodes (corresponding to 6 parameter types) and 7 directed edges, defined as follows: engine speed → fault code (engine abnormality may lead to fault code), engine speed → driving range (engine operating condition affects energy consumption), fault code → power system status (fault code indicates subsystem abnormality), fault code → braking system status, fault code → steering system status, driving range → power system status (driving range affects power distribution), power system status → braking system status (power and braking work together). The weight of each edge is set by domain experts based on experience, with a value range of [0.1, 0.9], representing the dependency strength. Propagation and aggregation are performed through a graph attention network, which contains 2 layers, each containing 4 attention heads, with a hidden layer dimension of 16 and an output layer dimension of 1. The first layer receives the initial pattern correlation as node features and aggregates neighbor node information through the attention mechanism; the second layer maps the aggregated features to the final relevance weights. The formula for calculating the attention coefficient is as follows. ,in The weight matrix is ​​a learnable matrix. For attention vectors, and They are nodes and The feature vector is generated. The correlation of engine-related modes is propagated to fault and range parameters, and the correlation of fault-related modes is propagated to parameters of each subsystem. After aggregation, six fine-grained correlation weights are output: engine speed 0.38, fault code 0.25, range 0.22, powertrain status 0.05, braking system status 0.05, and steering system status 0.05. These weights are used for subsequent weighted fusion of the original vehicle status parameter values.

[0038] Furthermore, the personalized sorting and content adaptation of the candidate answer set based on the user's historical preference features includes: Analyze each candidate answer in the candidate answer set and extract its answer type label, technical depth level, and expression complexity index; The answer type label, technical depth level, and expression complexity index of each candidate answer are matched with the user's acceptance of each answer type, the preferred technical depth range, and the comprehensible expression complexity threshold recorded in the user's historical preference features to calculate the matching degree and generate a personalized adaptation score for each candidate answer. The personalized adaptation score is weighted and fused with the confidence score of the candidate answer set. The weights are dynamically adjusted according to whether the current dialogue session vector contains the initial consultation or the in-depth follow-up question, to obtain the comprehensive ranking score. The candidate answer set is sorted in descending order based on the comprehensive ranking score, and the candidate answer ranked first is selected as the base answer.

[0039] In this embodiment of the invention, the candidate answer set is personalized and its content is adapted based on the user's historical preference features. First, the semantic parsing circuit parses each candidate answer in the candidate answer set. The candidate answer set contains 3 candidate answers, namely: Answer 1 "The engine speed is too high, which may be due to carbon buildup in the throttle body. It is recommended to check and clean the throttle body. The decrease in range is related to the high speed"; Answer 2 "The engine speed is too high, so the throttle body and spark plugs need to be checked. The decrease in range needs to be checked, and the specific operation should be performed by a professional"; Answer 3 "The high speed is abnormal, and the decrease in range is abnormal. It is recommended to go to a 4S store for inspection". For each candidate answer, we extracted the answer type label, technical depth level, and expression complexity index. Specifically: the answer type label was used to classify candidate answers into four categories: "troubleshooting," "functional operation," "driving guidance," and "maintenance advice" using a pre-trained text classification model (based on BERT-base and fine-tuned on 5000 labeled data points). The technical depth level was determined by statistically analyzing the number of domain-specific terms and the complexity of the operation steps in the answer, classifying answers into three levels: "basic operation" (terms ≤ 3, steps ≤ 2), "advanced settings" (terms 4-7, steps 3-5), and "troubleshooting" (terms ≥ 8, steps ≥ 6). The expression complexity index was obtained by weighting the average sentence length, clause nesting depth, and terminology density of the answer, with weights of 0.4, 0.3, and 0.3, respectively. Answer 1 is tagged as fault handling, with a technical depth level of basic operations and a complexity index of 0.6; Answer 2 is tagged as fault handling, with a technical depth level of advanced settings and a complexity index of 0.8; Answer 3 is tagged as fault handling, with a technical depth level of basic operations and a complexity index of 0.4. A 128-dimensional user historical preference feature is obtained from the user profile dynamic construction submodule. This feature records the user's acceptance of fault handling answers (0.9), the preferred technical depth range of basic operations (0.4-0.7), and the comprehensible complexity threshold of 0.7. The matching degree calculation circuit calculates the matching degree between each candidate answer and the user's preference. The matching degree calculation formula is: Matching Degree = 0.4 × Answer Type Acceptance + 0.3 × Technical Depth Fit + 0.3 × Expression Complexity Fit. The weights 0.4, 0.3, and 0.3 are based on user survey data. A questionnaire survey of 100 users was conducted to statistically analyze the importance users placed on answer type, technical depth, and expression complexity when choosing answers, and the normalized mean was used as the weight. Match score for answer 1 = 0.4 × 0.9 + 0.3 × 1 + 0.3 × 1 = 0.36 + 0.3 + 0.3 = 0.96; Match score for answer 2 = 0.4 × 0.9 + 0.3 × 0 + 0.3 × 0 = 0.36; Match score for answer 3 = 0.4 × 0.9 + 0.3 × 1 + 0.3 × 0.57 = 0.36 + 0.3 + 0.171 = 0.831. A personalized fit score is generated for each candidate answer.The personalized fit score and the confidence score inherent in the candidate answer set are weighted and fused. The confidence scores of candidate answer 1 are 0.85, 2 are 0.9, and 3 are 0.7. The current dialogue session vector implies the initial consultation stage. The weights are dynamically adjusted to 0.6 for the personalized fit score and 0.4 for the confidence score. The formula for calculating the comprehensive ranking score is: Comprehensive Score = 0.6 × Fit Score + 0.4 × Confidence. The comprehensive score for answer 1 is 0.6 × 0.96 + 0.4 × 0.85 = 0.576 + 0.34 = 0.916; the comprehensive score for answer 2 is 0.6 × 0.36 + 0.4 × 0.9 = 0.216 + 0.36 = 0.576; and the comprehensive score for answer 3 is 0.6 × 0.831 + 0.4 × 0.7 = 0.4986 + 0.28 = 0.7786. The candidate answer set is sorted in descending order based on the comprehensive ranking score. The ranking results are answer 1, answer 3, and answer 2. Answer 1, which ranks first, is selected as the base answer.

[0040] Furthermore, the instantiation and replacement of specific operation suggestions or parameter descriptions in the basic answer based on vehicle state parameters includes: Semantic role labeling is performed on the basic answer to identify the abstract operation instructions, general parameter descriptions, and conditional statements contained therein; Construct an instantiated mapping rule library that defines the mapping relationships between various abstract operation instructions, general parameter descriptions and specific vehicle subsystems and readable real-time status parameters; Based on the instantiation mapping rule base, the abstract elements identified in the basic answer are traversed. If an abstract element matches a rule, the specific instance value is calculated or found based on the latest vehicle status parameters obtained from the vehicle data real-time acquisition submodule. Replace the corresponding abstract operation instructions or general parameter descriptions in the basic answer with the calculated specific instance values, and compare the assumptions in the conditional statements with the current actual state of the vehicle. If they are met, the suggestion is retained; if they are not met, the suggestion is marked as invalid or replaced with alternative suggestions. After replacing all instantiable elements and validating conditions, the final executable answer is generated.

[0041] In this embodiment of the invention, specific operational suggestions or parameter descriptions in the basic answer are instantiated and replaced by combining vehicle state parameters. First, the semantic role labeling circuit is used to perform semantic role labeling on the basic answer "Excessive engine speed may be due to carbon buildup in the throttle body; it is recommended to check and clean the throttle body. Reduced range is related to excessive engine speed." The semantic role labeling uses a sequence labeling model based on BiLSTM-CRF, trained on 10,000 manually labeled automotive repair statements. It can identify semantic roles in the text such as "operation instructions" (e.g., "check," "clean," "replace"), "parameter descriptions" (e.g., "speed," "voltage," "opening"), "conditional statements" (e.g., "if...then...", "when..."), and "causal relationships" (e.g., "because...therefore..."). The abstract operation instruction is identified as "check and clean the throttle body," while there are no general parameter descriptions or conditional statements. An instantiated mapping rule library is constructed, defining the mapping relationship between abstract operation commands and specific vehicle subsystems and real-time status parameters. For example, "Check throttle body and clean" is mapped to the throttle body subsystem of the vehicle's powertrain, with associated parameters being engine speed and throttle opening. The mapping rule is that when engine speed > 6000 rpm, it is recommended to check and clean the throttle body for carbon buildup, with the cleaning cycle referencing the current cumulative engine runtime. Based on the instantiated mapping rule library, the abstract element "Check throttle body and clean" identified in the basic answers is traversed. This abstract element matches the corresponding rule, and the latest vehicle status parameters are obtained from the real-time vehicle data acquisition submodule, including the current engine speed of 6500 rpm, throttle opening of 85%, and cumulative engine runtime of 20000 km. Based on these parameters, specific instance values ​​are calculated. The degree of carbon buildup in the throttle body is determined by a combination of throttle opening and engine speed. The formula is: Carbon buildup degree = (Throttle opening - Standard opening) / Standard opening × 100%. With a standard opening of 60%, the carbon buildup degree is approximately 41.7% (85% - 60%) / 60% × 100%. The cleaning cycle is based on the engine's cumulative runtime; 20,000 km corresponds to an immediate cleaning cycle. Replace the corresponding abstract operation command in the basic answer with the calculated instance value. The replacement will read: "High engine speed (currently 6500 rpm) may indicate throttle body carbon buildup (approximately 41.7%). It is recommended to check and clean the throttle body immediately. Reduced range (currently 420 km, a 15% decrease from normal operating conditions) is related to high engine speed." The basic answer does not contain empty conditional statements, so condition validation is unnecessary.If a conditional statement exists, such as "If the engine speed remains above 6000 rpm, it is recommended to check the throttle body," then this condition is compared with the current actual vehicle status: if the current engine speed is 6500 rpm, the condition "speed remains above 6000 rpm" is met, and the suggestion is retained; if the condition is not met, a invalidation annotation "[Current condition not met, suggestion for reference only]" is added to the suggestion section, or it is replaced with another suggestion according to alternative rules. After completing the replacement and condition validation of all instantiable elements, the final executable answer is generated. This answer combines the vehicle's real-time status parameters, clarifies the timing and specific parameters of the operation, and facilitates user understanding and execution.

[0042] Furthermore, the construction and application of the instantiated mapping rule base include: The instantiation mapping rule base consists of a manually defined initial rule set and a system-learned extended rule set; When performing instantiation replacement, the system first attempts to find a matching mapping rule in the manually defined initial rule set; If no perfectly matching rule is found in the initial rule set, a rule self-learning process is triggered. This process, based on the enhanced query semantic vector and the basic answer, calls the automotive domain knowledge-enhanced large language model to generate a possible mapping hypothesis. Next, the system simulates and verifies the logical rationality and safety of the mapping hypothesis in a safe sandbox environment based on vehicle state parameters. If the verification is successful, the mapping hypothesis is temporarily adopted for this instantiation replacement. After the interaction is successful, and following a manual review process, the mapping hypothesis may be selectively added to the system's self-learning extended rule set.

[0043] In this embodiment of the invention, the instantiation mapping rule base consists of a manually defined initial rule set and a system-learned extended rule set. These two sets work together to provide a mapping basis for instantiation replacement. The manually defined initial rule set contains 100 fixed mapping rules, covering common abstract operations and specific parameter mappings of core subsystems such as the vehicle's powertrain, braking system, and comfort system. Each rule clearly defines an abstract description, associated subsystem, parameter type, and instantiation calculation method. For example, "check the throttle body and clean it" maps to the throttle body subsystem of the powertrain, associated with engine speed and throttle opening parameters, and the instantiation calculation method is: carbon buildup level = (throttle opening - standard opening) / standard opening × 100%; "check battery charge" maps to the power system, associated with driving range and remaining charge parameters, and the instantiation calculation method is: remaining charge percentage = current driving range / rated driving range × 100%. When instantiation replacement is performed, the system first searches for matching mapping rules in a manually defined initial rule set using a rule matching circuit. The matching method involves semantically comparing the abstract elements identified in the basic answer with the abstract descriptions in the initial rule set. This comparison is achieved by calculating the semantic similarity between the two; a semantic similarity ≥ 0.9 indicates a complete match. For example, if the abstract operation instruction in the basic answer is "check spark plug ignition status," the system searches for rules with the same abstract description in the initial rule set. If the initial rule set does not contain a mapping rule for this abstract operation, and the semantic similarity is < 0.9, then the rule self-learning process is triggered. This process, based on the context-aware enhanced query semantic vector and basic answer generated earlier, calls a large language model enhanced with automotive domain knowledge through a hardware interface to generate a possible mapping hypothesis. This mapping hypothesis must clearly define the correspondence between the abstract description and specific vehicle subsystems and parameters, as well as the instantiation calculation method. For example, the mapping hypothesis generated for "checking spark plug ignition status" is: "Checking spark plug ignition status maps to the powertrain ignition subsystem, associating it with engine speed and ignition voltage parameters. The ignition status evaluation method is: ignition voltage ≥12V is normal, <12V is abnormal." Next, in a safe sandbox environment, the system simulates and verifies the logical rationality and safety of this mapping hypothesis based on the vehicle status parameters currently obtained from the real-time vehicle data acquisition submodule. The safe sandbox environment is an independent virtual machine instance partitioned within the onboard computing unit, running the same ECU simulation software as a real vehicle (such as the CANoe simulation environment), but isolated from the actual vehicle actuators by a hardware firewall to prevent simulation operations from affecting the real vehicle. After receiving the mapped hypothesis, the sandbox environment virtually executes the corresponding operation instructions and monitors the status codes and parameter changes returned by the simulated ECU to verify whether the logic is sound, whether the parameters match, and whether there are any risks of exceeding limits or conflicts. The verification time is 200ms; if the timeout occurs, the verification is considered a failure.If the simulation verification passes (i.e., ignition voltage 11.8V < 12V), indicating an abnormal ignition state consistent with the mapping hypothesis evaluation method and without affecting the operation of other subsystems, the mapping hypothesis is temporarily adopted for this instantiation replacement. "Check spark plug ignition state" is instantiated as "Check spark plug ignition state (current ignition voltage 11.8V, determined to be abnormal)". After successful interaction, i.e., the user confirms the instantiated answer meets the requirements, the mapping hypothesis enters the manual review process. The manual review process is as follows: the system packages the mapping hypothesis and its verification log and sends it to the cloud review platform. Automotive experts review the triggering conditions, abstract description mode, target subsystem, parameter identifiers, and instantiation value calculation method of the mapping hypothesis through the review interface, and verify it against real vehicle test data. Experts can choose "Pass", "Reject", or "Pass after modification". The review result is returned within 24 hours. After passing the review, the mapping hypothesis is added to the system's self-learning extended rule set, and the review time, reviewer ID, and applicable vehicle model are marked. When the review is rejected, the system records the reason for rejection and discards the mapping hypothesis. The extended rule set is stored according to subsystems. Each extended rule is marked with the review time and applicable working conditions, which are used for the instantiation and replacement of similar abstract elements in the future.

[0044] Furthermore, the rule self-learning process includes: Separate the entity vector related to the operation object and the intent vector related to the operation target from the enhanced query semantic vector; The entity vector, intent vector, and abstract description fragment to be instantiated in the basic answer are input into a rule reasoning submodule. This submodule is a lightweight neural network used for path finding and reasoning on the structured knowledge subgraph generated by the large language model that enhances knowledge in the automotive domain. The rule reasoning submodule outputs one or more possible mapping paths from abstract descriptions to specific vehicle subsystems or parameters, along with their confidence levels. The mapping path with the highest confidence is selected, and combined with the specific value range and unit of the vehicle state parameters, it is transformed into a structured and executable mapping rule draft by the big language model enhanced with automotive domain knowledge, namely the mapping hypothesis. The draft mapping rules clearly define the triggering conditions, abstract description patterns, target subsystem or parameter identifiers, and instantiation value calculation methods.

[0045] In this embodiment of the invention, the rule self-learning process relies on hardware computing circuits and lightweight neural networks to execute step by step. First, a feature separation circuit separates entity vectors related to the operation object and intent vectors related to the operation target from the enhanced query semantic vector. The enhanced query semantic vector is 256-dimensional. The entity vector extracts semantic features related to the operation object to generate a 64-dimensional vector. For example, in the query text "What to do if the engine speed is too high and the spark plug ignition is abnormal", the entity vector focuses on operation objects such as "engine" and "spark plug". The intent vector extracts semantic features related to the operation target such as "check the cause of the abnormality" and "solve the abnormality" to generate a 64-dimensional vector. The separation process is implemented through a linear transformation circuit to ensure that the two types of vectors have no semantic overlap. The 64-dimensional entity vector, the 64-dimensional intent vector, and the abstract description fragment to be instantiated in the basic answer (such as "check the spark plug ignition status") are input into the rule reasoning submodule. This submodule is a lightweight neural network containing an input layer, a hidden layer, and an output layer. The input layer receives the entity vector, the intent vector, and the embedding vector of the abstract description fragment. The hidden layer contains 32 neurons and performs semantic fusion through nonlinear transformation. The output layer outputs the mapping path and confidence. The specific structure of this lightweight neural network is as follows: the input layer has a dimension of 64+64+32=160 (the abstract description fragment is mapped to a 32-dimensional vector through an independent word embedding layer); the hidden layer is a fully connected network containing 32 neurons, with ReLU as the activation function; the output layer has a two-branch structure, the first branch outputs the path selection probability (normalized by softmax), and the second branch outputs the confidence score of each path (mapped to [0,1] by sigmoid). The network is trained on 10,000 manually labeled abstract description-mapping path pairs, with the loss function being a weighted sum of cross-entropy loss and mean squared error loss, with weights of 0.7 and 0.3 respectively, a learning rate of 1e-4, and 50 training epochs. This submodule performs pathfinding and reasoning on a structured knowledge subgraph generated by a large language model enhanced with automotive domain knowledge. The structured knowledge subgraph is a subset extracted from the main knowledge graph, containing only nodes and edges related to operational instructions. Specifically, it includes four types of nodes: "subsystem" nodes (e.g., powertrain, braking system), "component" nodes (e.g., throttle body, spark plug), "parameter" nodes (e.g., engine speed, voltage), and "operation" nodes (e.g., inspection, replacement, adjustment), along with three types of relationships between them: "belongs to," "associates with," and "acts on." This subgraph is approximately 20% the size of the main knowledge graph, aiming to improve pathfinding efficiency.The rule reasoning submodule traverses the knowledge subgraph to find the association path between the abstract description fragment and the specific vehicle subsystem and parameters, and outputs two possible mapping paths and their confidence scores. Path 1 is "Check spark plug ignition status → power system → ignition subsystem → ignition voltage, engine speed", with a confidence score of 0.92; Path 2 is "Check spark plug ignition status → power system → engine subsystem → cylinder pressure", with a confidence score of 0.68. Path 1, with the highest confidence, is selected. Considering the specific value range and unit of the vehicle status parameters (ignition voltage range 8-15V, engine speed range 0-8000rpm, units V and rpm respectively), a large language model augmented with automotive domain knowledge transforms this mapping path into a structured, executable mapping rule draft, i.e., the mapping hypothesis. This rule draft explicitly specifies the trigger condition as "abstract description: checking spark plug ignition status," with the abstract description pattern being "check + component name + status." The target subsystem is the powertrain ignition subsystem, and the parameters are identified as ignition voltage and engine speed. The instantiation value calculation method is "ignition status evaluation value = ignition voltage / standard ignition voltage × 100%, standard ignition voltage is 12V, evaluation value ≥ 90% is normal, < 90% is abnormal." This fully presents the core elements of the mapping rule, ensuring it can be directly used for instantiation replacement.

[0046] Furthermore, Embodiment 2 of the present invention also provides an intelligent question-answering system for automobiles based on a large language model, including a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the intelligent question-answering method for automobiles based on a large language model as described above.

[0047] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, 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 the equivalents of the application are intended to be included within the invention.

[0048] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A car intelligent question answering method based on a large language model, characterized in that, Includes the following steps: During the system initialization phase, a large language model with enhanced automotive domain knowledge is deployed, and a multi-source information fusion module is built, which includes a real-time vehicle status data interface, a user personalized profile library, and a multi-turn dialogue context set. In the query understanding and information fusion stage, the system receives mixed query text input by the user, which includes natural language and vehicle system instructions. The system obtains the current vehicle status parameters, user historical preference features and current dialogue session vector in real time through the multi-source information fusion module. The system then jointly encodes the mixed query text, vehicle status parameters, user historical preference features and current dialogue session vector to generate a context-aware enhanced query semantic vector. In the domain knowledge-guided reasoning stage, the enhanced query semantic vector is input into the large language model with automotive domain knowledge enhancement. The large language model performs multi-hop reasoning based on its internal structured knowledge graph of automotive repair, driving behavior, and functional configuration and the enhanced query semantic vector to generate a preliminary candidate answer set containing reasoning logic chains and corresponding confidence scores. In the answer generation and personalization adaptation stage, the candidate answer set is personalized and adapted based on the user's historical preference features. The candidate answer with the highest ranking is selected as the basic answer. Then, the specific operation suggestions or parameter descriptions in the basic answer are instantiated and replaced in combination with the vehicle status parameters to generate the final executable answer. In the answer presentation and feedback learning phase, the executable answer is presented in a multimodal manner through the in-vehicle human-machine interface, and the user's interaction feedback signals to the presented answer are monitored in real time. A reward value is calculated based on the interaction feedback signals and the executable answer, and the reward value is used to perform incremental reinforcement learning fine-tuning on the policy network of the large language model for automotive domain knowledge enhancement.

2. The intelligent question-answering method for automobiles based on a large language model according to claim 1, characterized in that, The multi-source information fusion module includes a real-time vehicle data acquisition submodule, a dynamic user profile construction submodule, and a dialogue context management submodule. The real-time vehicle data acquisition submodule communicates with the vehicle sensor network via the vehicle bus to continuously acquire and standardize vehicle status parameters, including engine speed, fault codes, driving range, and the working status of each subsystem. The user profile dynamic construction submodule extracts and periodically updates the user's historical preference features by analyzing the user's historical Q&A records, in-vehicle function settings preferences, and driving behavior data. These features are vectorized to represent the user's tendencies in technical depth, expression style, and functional focus. The dialogue context management submodule is responsible for maintaining the dialogue history of the current session and encoding the question-answer pairs of the most recent N rounds of dialogue through a sequence model to generate the current dialogue session vector that represents the topic and depth of the current session.

3. The intelligent question-answering method for automobiles based on a large language model according to claim 2, characterized in that, The joint encoding generates context-aware enhanced query semantic vectors, including: The mixed query text is segmented and word vector embedded to obtain a query text vector sequence; The vehicle status parameters at the current moment are read from the real-time vehicle data acquisition submodule, and the relevance weight of each vehicle status parameter is calculated based on the query text vector sequence through a parameter importance evaluation network. The parameter values ​​are then weighted and fused based on the relevance weight to generate a contextualized vehicle status vector. The user's historical preference features are obtained from the user profile dynamic construction submodule, and the attention distribution of the features and the query text vector sequence in the latent semantic space is calculated. Based on the attention distribution, the user's historical preference features are selectively enhanced to generate a query-oriented user preference vector. Obtain the current dialogue session vector from the dialogue context management submodule; The query text vector sequence, the contextualized vehicle state vector, the query-oriented user preference vector, and the current dialogue session vector are concatenated to form a fusion feature tensor. The fused feature tensor is input into a cross-modal attention encoder, which interacts and aligns features from different sources through a multi-head attention mechanism, and finally outputs the context-aware enhanced query semantic vector.

4. The intelligent question-answering method for automobiles based on a large language model according to claim 3, characterized in that, The process of calculating the relevance weights in the parameter importance evaluation network includes: The parameter importance evaluation network receives the query text vector sequence as input and extracts its globally pooled semantic summary vector. The semantic summary vector is matched with a learnable vehicle state parameter type embedding table to generate a preliminary type relevance score for each type of vehicle state parameter. Based on the type-related scores, the real-time collected original vehicle state parameter values ​​of the corresponding type are normalized. The normalized parameter values ​​are combined with the corresponding parameter type embeddings, and then a non-linear transformation layer is used to generate the final, fine-grained correlation weights for weighting.

5. The intelligent question-answering method for automobiles based on a large language model according to claim 4, characterized in that, The generation of the final, fine-grained correlation weights used for weighting includes: The parameter importance evaluation network first performs self-attention operation on the query text vector sequence, focusing on key entities and verbs in the query, and extracting the core vector of query intent; Then, the cosine similarity between the core vector of the query intent and each parameter pattern vector in the predefined vehicle state parameter pattern library is calculated, where each parameter pattern vector is associated with one or a set of vehicle state parameters. The calculated cosine similarity is normalized using the softmax function to obtain the initial attention score for each parameter pattern; The initial attention score is multiplied by the data reliability coefficient of the corresponding parameter obtained in real time from the vehicle network. The data reliability coefficient is determined by the sensor health status and update frequency of the data source of the parameter to obtain the adjusted pattern correlation. Finally, based on the aforementioned pattern correlation degree and considering the prior dependency graph model between different types of parameters, a graph attention network is used for propagation and aggregation to output the fine-grained correlation weights that are ultimately used to weight the original parameter values.

6. The intelligent question-answering method for automobiles based on a large language model according to claim 1, characterized in that, The personalized sorting and content adaptation of the candidate answer set based on the user's historical preference features includes: Analyze each candidate answer in the candidate answer set and extract its answer type label, technical depth level, and expression complexity index; The answer type label, technical depth level, and expression complexity index of each candidate answer are matched with the user's acceptance of each answer type, the preferred technical depth range, and the comprehensible expression complexity threshold recorded in the user's historical preference features to calculate the matching degree and generate a personalized adaptation score for each candidate answer. The personalized adaptation score is weighted and fused with the confidence score of the candidate answer set. The weights are dynamically adjusted according to whether the current dialogue session vector contains the initial consultation or the in-depth follow-up question, to obtain the comprehensive ranking score. The candidate answer set is sorted in descending order based on the comprehensive ranking score, and the candidate answer ranked first is selected as the base answer.

7. The intelligent question-answering method for automobiles based on a large language model according to claim 1, characterized in that, The instantiation and replacement of specific operation suggestions or parameter descriptions in the basic answer based on vehicle status parameters includes: Semantic role labeling is performed on the basic answer to identify the abstract operation instructions, general parameter descriptions, and conditional statements contained therein; Construct an instantiated mapping rule library that defines the mapping relationships between various abstract operation instructions, general parameter descriptions and specific vehicle subsystems and readable real-time status parameters; Based on the instantiation mapping rule base, the abstract elements identified in the basic answer are traversed. If an abstract element matches a rule, the specific instance value is calculated or found based on the latest vehicle status parameters obtained from the vehicle data real-time acquisition submodule. Replace the corresponding abstract operation instructions or general parameter descriptions in the basic answer with the calculated specific instance values, and compare the assumptions in the conditional statements with the current actual state of the vehicle. If they are met, the suggestion is retained; if they are not met, the suggestion is marked as invalid or replaced with alternative suggestions. After replacing all instantiable elements and validating conditions, the final executable answer is generated.

8. The intelligent question-answering method for automobiles based on a large language model according to claim 7, characterized in that, The construction and application of the instantiation mapping rule base include: The instantiation mapping rule base consists of a manually defined initial rule set and a system-learned extended rule set; When performing instantiation replacement, the system first attempts to find a matching mapping rule in the manually defined initial rule set; If no perfectly matching rule is found in the initial rule set, a rule self-learning process is triggered. This process, based on the enhanced query semantic vector and the basic answer, calls the automotive domain knowledge-enhanced large language model to generate a possible mapping hypothesis. Next, the system simulates and verifies the logical rationality and safety of the mapping hypothesis in a safe sandbox environment based on vehicle state parameters. If the verification is successful, the mapping hypothesis is temporarily adopted for this instantiation replacement. After the interaction is successful, and following a manual review process, the mapping hypothesis may be selectively added to the system's self-learning extended rule set.

9. The intelligent question-answering method for automobiles based on a large language model according to claim 8, characterized in that, The rule self-learning process includes: Separate the entity vector related to the operation object and the intent vector related to the operation target from the enhanced query semantic vector; The entity vector, intent vector, and abstract description fragment to be instantiated in the basic answer are input into a rule reasoning submodule. This submodule is a lightweight neural network used for path finding and reasoning on the structured knowledge subgraph generated by the large language model that enhances knowledge in the automotive domain. The rule reasoning submodule outputs one or more possible mapping paths from abstract descriptions to specific vehicle subsystems or parameters, along with their confidence levels. The mapping path with the highest confidence is selected, and combined with the specific value range and unit of the vehicle state parameters, it is transformed into a structured and executable mapping rule draft by the big language model enhanced with automotive domain knowledge, namely the mapping hypothesis. The draft mapping rules clearly define the triggering conditions, abstract description patterns, target subsystem or parameter identifiers, and instantiation value calculation methods.

10. A car intelligent question-answering system based on a large language model, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the intelligent question-answering method for automobiles based on a large language model as described in any one of claims 1-9.

Images