A large model-based vehicle live risk assessment system and method

Abstract

The application belongs to the field of automobile evaluation, and provides a vehicle live risk evaluation system and method based on a large model, which comprises receiving voice input information of an evaluator during on-site inspection of a vehicle; extracting component state information of a target vehicle based on structured inspection item text; obtaining historical maintenance record data of the target vehicle, and associating and fusing the historical maintenance record data with the component state information; based on a large language model, jointly reasoning the currently input inspection label and the historical maintenance record to construct a cause-and-effect chain related to accident risk and generate at least one subsequent inspection item; generating voice guidance content and feeding back to the evaluator to prompt him / her to continue inspecting the target vehicle on the target item; gradually improving the risk inspection path of the target vehicle until the risk chain loop is closed; calling an evaluation formula to calculate the residual value of the target vehicle, and outputting a structured risk evaluation report.

Description

Technical Field

[0001] The present invention belongs to the field of automobile assessment, and in particular relates to a large model-based automobile real-time risk assessment system and method. Background Art

[0002] With the rapid development of the used car market, residual value assessment and accident risk assessment are becoming increasingly important in insurance claims, financing risk management, and vehicle transactions. To ensure reasonable vehicle pricing and accurate risk management, assessors typically need to conduct on-site inspections of each vehicle and make comprehensive assessments based on multiple factors, including historical maintenance data, age, and structural condition. The existing assessment process relies heavily on manual experience, resulting in low efficiency and poor consistency, making it difficult to meet the requirements of large-scale transactions and automated risk management.

[0003] Existing technologies include voice recognition, image acquisition, and access to repair records to assist with information collection and process management. Research has also attempted to use rule-based engines or shallow machine learning methods for residual value estimation and accident identification. For example, these methods use regression models to fit mileage and age, or employ image recognition models to detect cosmetic damage. However, these methods typically only cover static features and lack dynamic judgment capabilities. Furthermore, they are unable to automatically identify hidden risks or provide interactive guidance based on semantic logic. Summary of the Invention

[0004] In order to solve the problems in the prior art, the present invention provides a large-scale model-based automobile real-time risk assessment system, comprising:

[0005] A speech recognition module is used to receive speech input information from the assessor during the on-site vehicle inspection process and convert the speech input information into corresponding structured inspection item text;

[0006] An inspection item parsing module, configured to extract component status information of a target vehicle based on the structured inspection item text, and map the component status information into a standard inspection tag;

[0007] A multi-source information fusion module is used to obtain historical maintenance record data of the target vehicle and associate and fuse the historical maintenance record data with the component status information;

[0008] The large model reasoning module, based on the trained large language model, conducts joint reasoning on the currently input inspection labels and the historical maintenance records, constructs a causal chain related to the accident risk, and generates at least one subsequent item to be inspected;

[0009] A guidance prompt module, configured to generate voice guidance content based on the output result of the large model reasoning module and feed it back to the evaluator to prompt him to continue to inspect the target items of the target vehicle;

[0010] A multi-round interaction module is used to repeat the speech recognition, inspection item analysis, information fusion and reasoning process after each round of inspection, gradually improving the risk inspection path of the target vehicle until the risk chain is closed;

[0011] The residual value assessment module is used to call the valuation formula to calculate the estimated residual value of the target vehicle based on the final generated structured risk label, component status information, maintenance records, vehicle parameters and regional market parameters, and output a structured risk assessment report.

[0012] Furthermore, the speech recognition module includes:

[0013] A voice acquisition module is used to collect the voice input signal of the assessor at the vehicle site and convert the voice input signal into audio format data;

[0014] A speech recognition engine module, configured to transcribe the audio format data into text information;

[0015] A terminology disambiguation module is used to perform standardized replacement and semantic disambiguation of professional terms, homophones, or synonyms in the transcribed text;

[0016] The structured generation module is used to convert standardized text content into structured inspection item data, including part name, status description, location attribute and inspection type fields.

[0017] Furthermore, the inspection item parsing module includes:

[0018] Semantic decomposition module, used to perform syntactic analysis on the structured inspection item text and extract the subject, predicate and description components;

[0019] Field normalization module, used to perform semantic matching and normalization processing on the extracted fields and the preset vocabulary;

[0020] The standard label mapping module is used to map the normalized field combination into a predefined standard inspection label, wherein the label includes the fields of component, location, status, and risk level.

[0021] Furthermore, the multi-source information fusion module includes:

[0022] A historical data acquisition module is used to obtain the historical maintenance records of the target vehicle through the vehicle's unique identification code;

[0023] Data preprocessing module, used to standardize the format and structure the fields of maintenance records;

[0024] Semantic association module, used to semantically match maintenance information with the current component status, including multi-level matching methods at the component level, time level, and behavior level;

[0025] The fusion construction module is used to construct a semantic fusion graph containing maintenance events and current status, and represent the potential causal path between event nodes in the form of a graph structure.

[0026] Furthermore, the guidance prompt module includes:

[0027] The instruction generation module is used to convert the items to be checked in the large model inference results into natural language prompts;

[0028] A multimodal output module, configured to present the prompts to the assessor's device terminal via voice broadcast or graphic display;

[0029] The status synchronization module is used to record the response status of the prompt items and update the completion status of the check items;

[0030] The priority scheduling module is used to dynamically arrange the output order of prompt content based on risk score or task urgency.

[0031] Furthermore, the multi-round interaction module includes:

[0032] Round control module, used to maintain inspection rounds and the input and output relationships corresponding to each round;

[0033] The result confirmation module is used to confirm whether each round of voice input covers the guidance content and standardize it into a check label;

[0034] Dynamic path construction module, used to call the reasoning module to update the risk path map;

[0035] The closed-loop judgment module is used to determine whether the risk chain has reached the closed-loop condition and control the termination of the guidance process.

[0036] Furthermore, the residual value assessment module includes:

[0037] Parameter aggregation module, which aggregates structured labels, vehicle parameters, and market reference data to generate valuation input vectors;

[0038] Valuation rule configuration module, used to set the residual value estimation formula and the weights of influencing factors;

[0039] The residual value calculation module is used to call the regression model or valuation logic to generate residual value estimates and confidence intervals;

[0040] The evaluation report generation module is used to output a structured evaluation report containing valuation results, deduction reasons, risk level and recommended strategies.

[0041] Furthermore, the large language model used by the large model reasoning module includes:

[0042] A unified input encoding layer, which converts the speech transcription text, structural labels, historical records, and numerical parameters into a unified embedding vector;

[0043] The structured semantic adaptation layer is used to semantically model the relationship between structural labels using the graph attention mechanism;

[0044] The task routing layer is used to dynamically schedule different task paths based on the input semantic intent;

[0045] Transformer backbone network, used for context information fusion;

[0046] Multi-task output head, used to generate multiple semantic outputs separately.

[0047] Furthermore, the multi-task output head includes:

[0048] Natural language output head, used to output diagnostic suggestions, prompts, or question-and-answer results;

[0049] The causal chain generation header is used to build the logical path between check items;

[0050] Multi-label classification head, used to judge the dimensions of structural damage, historical accidents, and maintenance traces;

[0051] The residual value estimation regression head is used to output the estimated value and credible interval.

[0052] The present invention also provides a large-model-based vehicle real-time risk assessment method, which uses any of the aforementioned large-model-based vehicle real-time risk assessment systems to perform vehicle real-time risk assessment.

[0053] The present invention provides a large-model-based real-time automobile risk assessment system. Utilizing a large language model with causal reasoning and semantic memory capabilities, it can integrate multi-source information such as voice input, structured labels, historical maintenance records, and vehicle parameters. During the reasoning process, it dynamically generates accident risk chains and provides follow-up inspection recommendations, thereby improving the logic, comprehensiveness, and pertinence of risk assessments.

[0054] This invention implements a multi-round interactive risk identification process, continuously refining the inspection path through guided verbal feedback until a closed loop is formed, thereby supporting residual value estimation and risk level output. While improving inspection efficiency and accuracy, this system also structured, intelligent, and standardized the assessment process, demonstrating its potential for widespread application. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 is a flow chart of the system usage of the present invention;

[0057] Figure 2 It is a structural diagram of the large language model of the present invention;

[0058] Figure 3 It is a unified input coding layer structure diagram;

[0059] Figure 4 This is the SSE-Encoder layer structure diagram;

[0060] Figure 5 It is a diagram of the task routing layer structure;

[0061] Figure 6 This is the Transformer backbone network structure diagram;

[0062] Figure 7 This is the natural language output head structure diagram;

[0063] Figure 8 It is the cause and effect chain generating head structure diagram; structure diagram;

[0064] Figure 9 It is a diagram of the multi-label classification head structure;

[0065] Figure 10 It is the regression head structure diagram of residual value valuation. DETAILED DESCRIPTION

[0066] Below, the invention is preferably described with reference to the accompanying drawings and specific embodiments.

[0067] This embodiment solves the above problem through the following steps:

[0068] like Figure 1 As shown, in one embodiment, the present invention discloses a large-model-based real-time automobile risk assessment system. Utilizing a large language model with multi-round semantic understanding and causal reasoning capabilities, combined with real-time vehicle component inspection information, historical maintenance records, and regional market data, it can dynamically construct a vehicle risk causal chain, guide a multi-round interactive on-site inspection process, and ultimately achieve intelligent real-time automobile risk assessment with residual value estimation and risk level determination. The system includes:

[0069] A speech recognition module is used to receive speech input information from the assessor during the on-site vehicle inspection process and convert the speech input information into corresponding structured inspection item text;

[0070] An inspection item parsing module, configured to extract component status information of a target vehicle based on the structured inspection item text, and map the component status information into a standard inspection tag;

[0071] A multi-source information fusion module is used to obtain historical maintenance record data of the target vehicle and associate and fuse the historical maintenance record data with the component status information;

[0072] The large model reasoning module, based on the trained large language model, conducts joint reasoning on the currently input inspection labels and the historical maintenance records, constructs a causal chain related to the accident risk, and generates at least one subsequent item to be inspected;

[0073] A guidance prompt module, configured to generate voice guidance content based on the output result of the large model reasoning module and feed it back to the evaluator to prompt him to continue to inspect the target items of the target vehicle;

[0074] A multi-round interaction module is used to repeat the speech recognition, inspection item analysis, information fusion and reasoning process after each round of inspection, gradually improving the risk inspection path of the target vehicle until the risk chain is closed;

[0075] The residual value assessment module is used to call the valuation formula to calculate the estimated residual value of the target vehicle based on the final generated structured risk label, component status information, maintenance records, vehicle parameters and regional market parameters, and output a structured risk assessment report.

[0076] In the present invention, real-time automotive risk assessment refers to a system that identifies potential faults, accident signs, or repair histories that affect the value and safety of vehicles during on-site vehicle condition inspections of used vehicles (especially used vehicles), and assesses their value and risk level based on the current vehicle condition, historical data, and market conditions.

[0077] A large model refers to a deep neural network model with a parameter scale of more than 100 million, especially a large language model with the capabilities of natural language understanding, reasoning generation, and instruction following. In the present invention, improvements are made based on the existing large model to make it suitable for real-time automobile risk assessment.

[0078] The implementation of the real-time automobile risk assessment system will be explained one by one below.

[0079] The speech recognition module is used to receive speech input information from the assessor during the on-site vehicle inspection process and convert the speech input information into corresponding structured inspection item text.

[0080] In a large-scale model-based real-time automotive risk assessment system, in order to achieve efficient input collection and automated processing of subsequent reasoning guidance during assessors' on-site operations, it is necessary to convert verbal information into a structured data format that the system can recognize, process, and reason with. Since assessments are mostly conducted in the actual vehicle environment, it is difficult for assessors to quickly record fault descriptions, component status, or observations through manual input. Therefore, the use of voice input, combined with voice recognition and semantic parsing technology, can significantly improve inspection efficiency. Converting voice content into structured inspection item text not only achieves the conversion between natural input and standard output, but also provides a clear, unified, machine-understandable data structure for the subsequent large-scale model reasoning module, ensuring the continuity, accuracy, and intelligence of the entire risk assessment process.

[0081] In an optional specific implementation, the speech recognition module in this system specifically includes the following submodules:

[0082] The voice acquisition module is used to collect voice data emitted by the assessor at the vehicle site. The voice data is picked up by a device with a microphone component, such as a handheld terminal, a clamp-on microphone, etc. and converted into an audio signal. The audio signal can optionally be in a 16kHz, single-channel PCM encoding format to facilitate subsequent recognition engine processing.

[0083] A speech recognition engine module is used to transcribe the audio signal into text. Preferably, the speech recognition engine is built based on a deep neural network acoustic model and a CTC decoding algorithm. It supports the recognition of spoken expressions, industry terminology, and component proper nouns, and has adaptive speech model training capabilities. The engine can run locally on the mobile device or through a cloud service. Cloud solutions offer greater model capacity and update capabilities, making them suitable for complex speech environments.

[0084] The term disambiguation module is used to standardize and semantically disambiguate professional terms, homophones, and synonyms in the recognized text, avoiding label mismatches caused by misidentification. Preferably, this module performs joint judgment based on industry dictionaries, a vehicle component knowledge base, and a context-dependent syntax tree. For example, the "right front door" in "scratches on the right front door" is mapped to the standard component name "right front side door," and "slightly deformed" is parsed as the risk status "slightly deformed."

[0085] A structured data generation module maps the semantically parsed speech text content into standardized structured data, including but not limited to four fields: component name, location attributes, status description, and damage level. This module preferably utilizes a combination of rule-based template matching and entity recognition to ensure a uniform output format, supporting structured representations such as JSON and XML, and interfacing with subsequent inference modules.

[0086] Through the multi-level processing flow of the speech recognition module, this system achieves highly accurate and semantically consistent conversion of evaluators' natural language input. This not only avoids the inefficiency and irregularities of manual entry, but also ensures the input quality of large-scale model reasoning, enhancing the overall system's intelligence, automation, and structured capabilities. Compared to traditional field entry methods, this module offers greater real-time assessment performance, stronger semantic integrity, and significantly reduces data bias caused by discrepancies in manual input.

[0087] In a specific example, an assessor utters a voice description during an on-site inspection of a target vehicle: "The left front shock absorber seat is slightly deformed." The voice acquisition module inputs the audio signal into the voice recognition engine module, which transcribes it into the text "The left front shock absorber seat is slightly deformed." The term disambiguation module normalizes "somewhat deformed" to "slightly deformed," and the structured generation module converts it into structured inspection item text:

[0088] {

[0089] "Part Name": "Left Front Shock Absorber Mount",

[0090] "Component Location": "Front Suspension System",

[0091] "Status Description": "Slightly deformed",

[0092] "Inspection Type": "Structural Damage"

[0093] }

[0094] The above structured text will serve as one of the inputs of the large model reasoning module to trigger the subsequent inspection guidance and residual value estimation process.

[0095] The inspection item parsing module is used to extract component status information of the target vehicle based on the structured inspection item text and map the component status information into standard inspection labels.

[0096] In a large-scale model-based real-time automotive risk assessment system, the structured inspection item text entered by the assessor via voice input contains a large amount of descriptive information in natural language. Although this information has a certain semantic structure, the way it is expressed may vary between assessors, and the semantic granularity, professional terminology, and expression order in the text may be non-standardized. In order for the system to accurately identify key elements such as the status, location, and risk level of vehicle components and standardize them into a unified label format that can be processed by downstream modules, it is necessary to set up an inspection item parsing module to perform semantic extraction, field normalization, and label mapping operations on the structured inspection item text to ensure the stability, accuracy, and consistency of the entire system assessment process.

[0097] In an optional specific implementation, the inspection item parsing module specifically includes the following submodules:

[0098] The semantic decomposition module is used to perform syntactic structure analysis of the subject, predicate, object and modifying components of the structured inspection item text, and identify key fields such as "component name", "status description", and "position relationship". Preferably, the module adopts a method based on the joint modeling of dependency syntactic analysis and named entity recognition, and can accurately identify technical entities and semantic relationships in colloquial descriptions after combined with domain-specific corpus training.

[0099] The field normalization module is used to match and normalize the field content output by the semantic decomposition module with the system's preset component vocabulary, state vocabulary, and position vocabulary. The module maps multiple descriptions into unified terms through fuzzy matching, spelling correction, and semantic distance calculation. For example, "front wheel left shock absorber" is normalized to "left front shock absorber seat", and "slight dent" is normalized to "slight deformation".

[0100] The standard label mapping module is used to map normalized field combinations into predefined standard inspection labels. The labels include fields such as component name, inspection type, risk level, damage type, component location, etc. Preferably, the module is constructed based on a rule mapping table and a context dynamic correction mechanism, and can adjust the confidence level and label combination form of the output label according to the context information to ensure that the label results are highly consistent and highly interpretable.

[0101] Through the collaborative processing of these submodules, the inspection item parsing module converts the semantically rich but formatted inspection item descriptions entered by assessors into structured, standardized, and clearly labeled data. This provides accurate and computable input support for subsequent large-scale model reasoning, risk path construction, and residual value calculation. This module boasts strong input fault tolerance, good contextual semantic adaptability, and stable label output, significantly improving the system's parsing accuracy and label consistency in complex speech scenarios.

[0102] In a specific example, an assessor's voice input, "This left front shock absorber seat is slightly deformed," is transcribed by the speech recognition module into the structured inspection item text, "The left front shock absorber seat is slightly deformed." The semantic decomposition module of the inspection item parsing module identifies the main component, "left front shock absorber seat," and the state description, "some deformation." The field normalization module classifies "some deformation" as the standard state, "mild deformation." The standard label mapping module maps this to the following label output:

[0103] {

[0104] "Part Name": "Left Front Shock Absorber Mount",

[0105] "Inspection Type": "Structural Parts Inspection",

[0106] "Component Location": "Front Suspension System",

[0107] "Injury Level": "Mild"

[0108] }

[0109] This structured labeled data will be passed as input to the large model reasoning module to further infer relevant risk causal chains and subsequent inspection recommendations.

[0110] The multi-source information fusion module is used to obtain the historical maintenance record data of the target vehicle and associate and fuse the historical maintenance record data with the component status information.

[0111] In a large-scale model-based real-time automotive risk assessment system, relying solely on component status information obtained from on-site vehicle inspections makes it difficult to fully assess a vehicle's actual usage and maintenance risks. Since a vehicle may have a history of repairs, accidents, and replacements, these historical events significantly impact the current component status and residual value assessment. Therefore, it is necessary to integrate historical maintenance record data and integrate it with component status information obtained from current inspections to enhance the system's comprehensive assessment of accident risks, structural hazards, and failure probabilities. By establishing a correspondence between the vehicle's historical behavior and its current real-time status, the system can reconstruct the causal chain of failures and accurately track risk evolution paths, providing a richer and more accurate data foundation for large-scale model risk reasoning and residual value estimation.

[0112] In an optional specific implementation, the multi-source information fusion module specifically includes the following submodules:

[0113] A historical data acquisition module is used to query and retrieve historical maintenance data based on the target vehicle's unique identification information. This unique identification information includes, but is not limited to, the vehicle identification number, license plate number, or VIN code. Preferably, this module establishes a data interface with a third-party maintenance service system, a 4S dealership repair platform, and an insurance claims database. Through an API or intermediate data service, this module retrieves the corresponding vehicle's historical maintenance records, including repair time, repair location, repair item, repair method, work order number, and executing agency.

[0114] The data preprocessing module is used to standardize the format and structure the fields of acquired historical maintenance records. This module applies unified field mapping to data from different sources and with different field naming conventions, removes redundant entries, fills in missing fields, and extracts the core elements of maintenance actions, including repair parts, operation types, and maintenance dates, through regular expressions, dictionary matching, and text parsing.

[0115] The semantic association module is used to perform semantic hierarchical matching between standardized historical maintenance information and component status information obtained from the current inspection. This matching includes three methods: component-level matching, time-level matching, and operational behavior-level matching. Component-level matching is based on component names, component aliases, and synonymous expression rules in the system component library; time-level matching determines the timeliness of maintenance information by comparing the interval between the repair time and the current assessment time; and operational behavior-level matching identifies whether there is logical consistency between the repair type and the current component status, for example, determining the causal relationship between "replacing the left front shock absorber seat" and "deformation of the left front shock absorber seat."

[0116] A fusion construction module is used to construct a fused representation of historical behavior and current status based on semantic association results. This module integrates maintenance records and live inspection information into an event sequence along a timeline, constructing a vehicle maintenance behavior trajectory. It also annotates the potential impact of each event on the current status, generating a structured fused semantic graph. Preferably, this module uses a graph structure representation method, with each node representing a maintenance behavior or current status, and each edge representing a potential logical association path for subsequent inference by the large model.

[0117] This multi-source information fusion module uses structured and semantic methods to link historical maintenance records with current component status information, enhancing the system's integrity and robustness in complex fault identification, risk chain reasoning, and value impact modeling. This module not only deepens risk identification but also significantly enhances the system's ability to detect dishonest behavior, false vehicle conditions, and potential accidents. This module has important applications in data-driven residual value assessment and financial risk control.

[0118] In a specific example, for the previously identified "minor deformation of the left front shock absorber seat" condition, the historical data acquisition module retrieves a maintenance record for the vehicle in June 2023, which states "Replacement of the left front shock absorber assembly." The data preprocessing module parses the record as: "Repair component: left front shock absorber seat, Operation type: Replacement, Time: June 15, 2023." The semantic association module determines, based on the current date and repair time, that the replacement occurred less than six months ago and that the operation type is correlated with the current component condition. The fusion construction module establishes a causal path between the current state and the historical behavior, generating the following fusion output:

[0119] {

[0120] "Fusion event node": [

[0121] { "Time": "2023-06-15", "Action": "Replace the left front shock absorber seat"},

[0122] { "Time": "2023-11-12", "Status": "Left front shock absorber seat slightly deformed"}

[0123] ],

[0124] "Presumed Impact Relationship": "Deformation persists after recent replacement, suggesting the front collision has not been fully repaired or there are structural stress hazards."

[0125] }

[0126] The fusion result is sent to the large model reasoning module to further determine whether a moderate collision has occurred in the front, and prompt the evaluator to check related components such as the "left A-pillar" or "subframe structure."

[0127] The large model reasoning module, based on the trained large language model, performs joint reasoning on the currently input inspection labels and the historical maintenance records, constructs a causal chain related to the accident risk, and generates at least one subsequent item to be inspected.

[0128] In a real-time automotive risk assessment system based on a large model, the status of a vehicle's components and its historical maintenance behavior often have implicit correlations, making it impossible to determine whether the vehicle has accident hazards through static judgments or simple rules. Accident damage has a chain-like propagation characteristic, and the initially damaged component often involves the subsequent deformation or repair of multiple structural components, and these indirect effects are often difficult to detect during the initial inspection stage. Therefore, it is necessary to introduce a large language model with deep semantic understanding and structured causal reasoning capabilities to semantically fuse current inspection labels with historical maintenance records. Based on the embedded knowledge graph and reasoning path, the possible risk chain of the vehicle is constructed, and then components or systems that have not yet been inspected but have potential risks are identified to guide the generation of subsequent inspection paths and improve the completeness and foresight of risk assessments.

[0129] In an optional specific implementation, the large model reasoning module specifically includes the following submodules:

[0130] The context construction module organizes structured inspection labels, integrated historical maintenance records, current assessment time, basic vehicle information, and other data into a standardized input format recognizable by the large language model. Preferably, this module abstracts each inspection label and maintenance record into an event object, constructing an event sequence based on chronological order and logical associations. This module also incorporates context variables such as vehicle model, age, and mileage to generate an input context document that is both temporally and structurally compatible.

[0131] The causal chain reasoning module is used to call upon a pre-trained and fine-tuned large language model, taking the context document as input and outputting information on possible causal impact paths between components and their logical explanations. Preferably, the large language model has multi-hop reasoning capabilities, supports cross-event causal connections, and can automatically identify the fault's origin, impact path, and potential expansion area. For example, from "left front shock absorber seat deformation," it can infer "left A-pillar stress anomaly" or "subframe hidden cracks" as risk diffusion results.

[0132] The inspection item generation module automatically selects high-priority components that have yet to be inspected as the next items to be inspected, based on the potentially impacted components and systems output by the causal chain reasoning module, combined with the vehicle configuration and existing inspection records. Preferably, this module incorporates an information gain evaluation mechanism, assigning a risk information density score to each candidate inspection item to prioritize and output suggested paths, ensuring maximum risk coverage efficiency for subsequent inspections.

[0133] The structured output module formats causal chains, logical reasoning results, and next inspection recommendations into a combination of structured data and natural language. This module represents the generated inspection path as a node-connected graph, complete with semantic descriptions and model confidence. It also outputs natural language prompts for the voice prompt module, such as "Recommend checking the front bumper for signs of replacement," to support interactive guidance.

[0134] This large-scale model reasoning module expands static inspection labels into dynamic risk chains through a four-step process: event context construction, causal path derivation, inspection path planning, and structured result output. This provides the assessment system with a core intelligent engine capable of deep semantic understanding and proactive decision-making. This module not only enhances the system's ability to identify damage in complex structures but also significantly increases assessors' efficiency in identifying hidden dangers, determining incident severity, and developing inspection paths. It is a key component in implementing intelligent, full-process risk assessment.

[0135] In a specific example, for the identified "minor deformation of the left front shock absorber seat" and the historical maintenance record of "replacement of the left front shock absorber assembly," the context building module generates an input document containing information such as the component's current status, relevant maintenance records, vehicle make and model, and mileage. Based on this input, the causal chain reasoning module identifies the possible risks of "minor displacement of the left A-pillar" and "subframe stretching." The inspection item generation module lists these uninspected components as next recommended items and prioritizes the "left A-pillar" based on the risk impact score. The structured output module ultimately generates the following content:

[0136] {

[0137] "Inference chain": ["Left front shock absorber seat deformation" → "Left A-pillar stress abnormality"],

[0138] "Recommended inspection items": "Status of the left A-pillar structure",

[0139] "Prompt": "The inspection history shows that the left front has been impacted. It is recommended to further check the left A-pillar for slight displacement or spray marks."

[0140] }

[0141] The above content will be used in the next round of voice prompts and interactive input to guide assessors to conduct more in-depth structural inspections and support subsequent residual value assessments and risk level determinations.

[0142] The guidance prompt module is used to generate voice guidance content based on the output results of the large model reasoning module and feed it back to the evaluator to prompt him to continue to inspect the target items of the target vehicle.

[0143] In a large-scale model-based real-time automotive risk assessment system, in order to achieve the integrity and logical closure of the inspection process, relying solely on the subjective experience of the assessor for component inspection often carries the risk of omissions or misjudgments. By introducing a guidance prompt module, the system can generate clear and targeted inspection guidance content based on the potential risk paths identified by the large-scale model reasoning module, and provide feedback to the assessor via voice or text. This can significantly improve the inspection coverage, reduce the risk of missed inspections, and maintain the consistency and standardization of the on-site assessment process. The introduction of this module enables the intelligent reasoning capabilities of the large model to be actually implemented in the on-site operation process in the form of human-computer interaction, promoting efficient collaboration between human judgment and intelligent systems.

[0144] In an optional specific implementation, the guidance prompt module specifically includes the following submodules:

[0145] The instruction generation module is responsible for generating natural language expressions suitable for assessors based on the subsequent inspection items output by the large model inference module. Based on a preset prompt language template library, this module maps structured risk warning information into semantically coherent and linguistically standardized guiding statements. Preferably, this module also dynamically selects appropriate language modes based on different vehicle types, inspection item categories, and contextual status to enhance the adaptability of language expression and the efficiency of instruction acceptance.

[0146] The multimodal output module is used to provide the generated guidance content to the evaluator in the form of voice, text, or a combination of text and images. This module supports speech synthesis technology, preferably using a device-based model or cloud-based TTS (Text-to-Speech) service to complete natural language voice broadcast. It also supports visual display of guidance content on the handheld terminal interface, with auxiliary information such as color highlighting and step-by-step markings to improve recognition.

[0147] The status synchronization module records the feedback status of each guided content and synchronously updates the system's assessment progress status table. This module receives subsequent voice input from the speech recognition module and automatically compares it to see if the guided inspection items are covered. If completed, the status is updated to "Executed". If not, the prompt remains in a pending state and supports a re-prompt mechanism to ensure that key inspection items are not missed.

[0148] The priority scheduling module manages the display order and output cadence of multiple guidance prompts. This module dynamically prioritizes prompts based on each inspection item's risk score, component importance, and inspection path continuity. It also automatically adjusts the interval between the next round of prompts based on the assessor's response speed to avoid information overload or delayed prompts. Optional implementation strategies include polling output, centralized prompts, or multi-level filtering, switching between them based on the complexity of the vehicle inspection.

[0149] This guidance and prompting module applies the intelligent judgment results generated by the large model to the assessment site through natural interaction. This not only improves operators' acceptance and response speed to complex reasoning results, but also optimizes the assessment process from a "human-led" to a "system-assisted" one. This module offers technical advantages such as clear prompts, flexible feedback methods, stable control mechanisms, and traceable execution status. It is a key link in achieving a closed-loop risk identification and assessment decision-making linkage.

[0150] In a specific example, in response to the subsequent suggestion item "Check whether the left A-pillar has structural deformation" output by the large model reasoning module, the instruction generation module generates the sentence "It is recommended to check whether the left A-pillar has displacement, deformation or paint marks". The multimodal output module synthesizes the sentence into voice through the TTS engine, and automatically plays it on the mobile terminal used by the evaluator. The "A-pillar part" is highlighted in the graphic interface at the same time. After the status synchronization module records the voice broadcast, the evaluator's voice reply "No obvious abnormality is found in the left A-pillar". The system automatically recognizes that the item has been completed and removes it from the prompt queue. After confirming that the risk of the current inspection path is not closed, the priority scheduling module continues to schedule the next high-risk item "Check whether the subframe has stretched" and issues corresponding guidance prompts to realize an intelligent-led multi-round guidance closed-loop process.

[0151] The multi-round interaction module is used to repeat the speech recognition, inspection item analysis, information fusion and reasoning process after each round of inspection is completed, gradually improving the risk inspection path of the target vehicle until the risk chain is closed.

[0152] In large-scale model-based real-time automotive risk assessment systems, vehicle fault information is often revealed gradually and deduced layer by layer. A single round of interaction can only obtain limited risk information, making it difficult to fully assess structural hazards or underlying accident paths. Therefore, to ensure the systematic and closed-loop nature of risk assessment, it is necessary to establish a mechanism that supports multiple rounds of human-computer interaction. This allows assessors to continuously input new inspection results. The system dynamically updates the reasoning path, risk map, and guidance content based on each round of input, advancing the risk chain construction process round by round. Through the cyclic execution of speech recognition, structural analysis, multi-source fusion, and causal reasoning, the system can gradually construct a logically self-consistent and comprehensive fault causal path with minimal operator intervention, until the system determines that the risk chain is closed, meaning that there are no high-risk inspection items requiring further guidance.

[0153] In an optional specific implementation, the multi-round interaction module specifically includes the following submodules:

[0154] The round control module tracks the round number of the current interaction and maintains a mapping between each round's inspection results and reasoning status. After initialization, the module establishes an inspection round index table. Each time a new inspection input and reasoning guidance process is completed, it records a round and updates a snapshot of the vehicle's risk status, which is used for comparison with the previous round's status and path evolution assessment. Preferably, the module supports round-by-round backtracking of previous inputs and outputs, facilitating recovery of process status during field interruptions or reassessments.

[0155] The Results Receiving and Confirmation Module receives voice input from assessors following system prompts during each round of inspections, converts it into structured inspection labels, and updates the current status information. This module standardizes the input results by invoking the Voice Recognition Module and the Inspection Item Parsing Module. It then correlates and matches the results with the previous round's reasoning prompts to confirm whether the inspection items have been executed and whether the results constitute new risk points.

[0156] The dynamic path building module triggers the large-scale model inference module to perform updated reasoning based on the input results of each inspection round and historical maintenance information. This module generates a new causal path, potential risk points, and follow-up inspection recommendations, and generates an updated vehicle risk chain structure diagram. This module analyzes the difference between the current diagram and the previous diagram to determine whether a closed loop has been formed and whether there are any uncovered risk nodes. It also assesses the path completeness based on the diagram topology.

[0157] A closed-loop determination module is used to determine whether the current risk path has met the closed-loop conditions after each round of interaction. Preferably, these conditions include: all high-risk nodes in the check chain have been assessed or checked, the confidence level of potential inspection items generated by the model falls below a set threshold, or the assessor proactively terminates the interaction process. Meeting any of these conditions determines the closed-loop completion. After the loop is closed, the system automatically terminates the guided process and outputs the complete path for residual value assessment and report generation.

[0158] The multi-round interaction module, through a linkage mechanism of round management, result input, path iteration, and closed-loop judgment, enables intelligent control and state adaptation of the vehicle inspection process, significantly improving the system's responsiveness to complex risk structures. This module frees the system from relying on fixed, manually defined inspection processes and instead enables it to continuously adjust its next strategy based on known information. This demonstrates remarkable flexibility, closed-loop functionality, and intelligent judgment, making it particularly suitable for real-time assessment scenarios involving complex vehicles, multiple component maintenance histories, and uncertain multi-source inputs.

[0159] In a specific example, the assessor has input "left front shock absorber seat slightly deformed" via voice and confirmed that the "large model reasoning module" has output "recommend inspection of left A-pillar". After the first round of interaction, the assessor continues to input "left A-pillar has slight traces of paint". The system identifies this as a new possible risk item through the "result reception and confirmation module". Based on this, the "dynamic path construction module" infers that the left front of the vehicle may have experienced a moderate collision and further recommends checking "whether the front bumper beam has any signs of replacement". At this time, the round control module records the start of the second round, and the "closed-loop judgment module" remains in the interactive state because the current reasoning has not yet terminated. The assessor continues to input "front bumper beam has no abnormalities". In the third round, the system determines that there are no new risk prompts and the confidence levels of the inspection items are all below the threshold. It ultimately determines that the risk path is closed and ends the interaction, generating a complete inspection path map and assessment report.

[0160] The residual value assessment module is used to call the valuation formula to calculate the estimated residual value of the target vehicle based on the final generated structured risk label, component status information, maintenance records, vehicle parameters and regional market parameters, and output a structured risk assessment report.

[0161] In large-scale model-based real-time automotive risk assessment systems, component status information obtained on-site and structured risk labels generated through reasoning are often used only to describe the vehicle's current technical condition. Without a quantitative assessment of the vehicle's value, key business processes such as financial pricing, residual value guarantees, and resale calibration cannot be supported. Because a vehicle's age, maintenance activities, component damage, and regional market conditions all significantly impact its residual value, a residual value assessment module is required. This module integrates structured risk output with the vehicle's static parameters, dynamic history, and external market data. Using a unified valuation logic framework, it calculates the current vehicle's estimated residual value and outputs an assessment report with data structuring capabilities and business interpretation capabilities, effectively connecting technical judgment with financial pricing.

[0162] In an optional specific implementation, the residual value assessment module specifically includes the following submodules:

[0163] The parameter aggregation module aggregates the various input parameters that form the basis of the valuation. These parameters include structured risk tags, component status information, historical maintenance records, vehicle registration information, age, mileage, and regional market residual value coefficients. Preferably, this module standardizes and normalizes all input fields, removes missing and redundant values, and performs time series alignment on temporal data to construct a complete parameter vector for residual value calculation.

[0164] The valuation rule configuration module maintains the formula models and business rules used for residual value assessment. This module supports the configuration and switching of multiple valuation methods, including pricing logic based on cost depreciation, market comparison, and model regression. Preferably, this module allows parameter configuration to adjust the weights of various risk factors, such as the impact of accident level, damage to key components, and historical repair frequency on residual value. Optional implementation options include using a static valuation model, a historical transaction sample regression model, or a large model inference score weighted model for residual value estimation.

[0165] The residual value calculation module is used to call the specified valuation model, quantitatively calculate the aggregated parameters, and generate the target vehicle's current residual value estimate. This module supports interval valuation and confidence score output, providing upper and lower limits for the residual value estimate and annotating the model score confidence level. It also supports different valuation strategies based on the scenario, such as increasing the weight of historical accidents in financial risk control scenarios and highlighting the influence of current market conditions in circulation and trading scenarios.

[0166] The assessment report generation module summarizes the residual value calculation results and components in a structured format and generates a multi-format risk assessment report suitable for system access, review by assessors, and report storage. This module supports the generation of JSON, XML structured data, and PDF graphic documents. The content includes the valuation range, key risk summary, valuation logic path, visual risk level icons, and recommended disposal plans. Preferably, the report can be pushed to financial systems, dealer platforms, or quality inspection and archiving systems through an interface.

[0167] The aforementioned residual value assessment module, through a standardized process of parameter aggregation, model invocation, and report output, enables vehicle risk information to directly inform residual value estimation, breaking down the barriers between technical judgment and valuation. This module boasts a clear structure, strong configurability, and adaptability to a wide range of scenarios. It is widely applicable to various business areas, including vehicle valuation, financial leasing assessment, non-performing asset disposal, and residual value actuarial calculations. It serves as the core calculation unit within the assessment system, connecting risk and value.

[0168] In a specific example, based on the information identified by the previous module regarding "minor deformation of the left front shock absorber seat," "paint marks on the left A-pillar," and "historical replacement of the left front shock absorber assembly," the parameter aggregation module collects and standardizes the vehicle's initial registration date as June 2019, mileage of 63,000 kilometers, regional market coefficient of 0.83, and accident impact level of "medium." The valuation rule configuration module uses a valuation logic that combines a depreciation base value with an accident risk adjustment factor, setting a moderate accident deduction coefficient of 0.12 and a component damage deduction of 0.05. Based on this, the residual value calculation module calculates an estimated residual value of 48,500 yuan and a valuation range of 46,000 to 51,000 yuan. The evaluation report generation module's final output includes the valuation range, depreciation basis, detailed risk deduction items, and a business prompt stating that the vehicle is recommended for auction, with a starting price not exceeding 50,000 yuan. It also generates a corresponding structured report for access by the financial system.

[0169] In another embodiment, a specific structure and training method of a large model applied to the present invention are provided.

[0170] This paper addresses the challenges of diverse information sources, complex semantic logic chains, and intertwined business task processes in real-time automotive risk assessment. By designing a large language model architecture, this model is capable of structured semantic adaptation, task-guided routing, causal chain reasoning, and residual value estimation. This large model not only processes natural language input but also accepts multiple input forms, including structured labels, historical maintenance data, and vehicle numerical parameters. It automatically infers fault chains and generates inspection guidance and valuation recommendations.

[0171] Its core design principles are:

[0172] Unify heterogeneous inputs into multimodal representations in a semantic space;

[0173] Control reasoning paths and information flows through semantic task routing mechanisms;

[0174] Constructing an explicit causal reasoning structure to achieve “semantic chain” generation;

[0175] The multi-task output interface meets the question answering, classification, path generation and valuation tasks in the evaluation scenario;

[0176] It supports multiple rounds of interaction and context state preservation to form a continuous reasoning process.

[0177] like Figure 2 As shown in the figure, the model structure consists of five core modules: input embedding layer, structured semantic adaptation layer, task routing layer, Transformer backbone network, and output head module. Its overall structure is shown in the figure.

[0178] like Figure 3As shown in the figure, the unified input encoding layer is used to normalize the format and semantic embedding of heterogeneous data such as speech-to-text, structured labels, historical maintenance records, and vehicle parameters.

[0179] Text input uses the WordPiece tokenizer; label input (e.g., "Left front shock absorber seat: slightly deformed") uses field tag encoding; numerical parameters (mileage, year) use positional encoding. All inputs are concatenated into a token sequence and marked with an embedding type. Optionally, this functionality can be implemented using an embedding matrix with positional information and semantic type embedding.

[0180] like Figure 4 As shown in Figure 1, SSE-Encoder is used to semantically model structured input information (labels, fields, forms) so that it can participate in the attention mechanism in the Transformer context.

[0181] Each field (such as "component name," "status," and "impact level") is mapped to a structural fragment token. Each fragment has a field type bit to identify its semantic role. All structural tokens are graph-connected through a small encoder to generate a local context tensor. The output tensor is connected to the backbone network. The implementation uses a multi-head attention structure, semantic graph convolution between fields, and residual connections.

[0182] like Figure 5 As shown in the figure, the task routing layer is used to automatically select the reasoning path based on the input semantic intent, so that different tasks (such as chain reasoning and residual value calculation) enter the dedicated sub-network for optimization.

[0183] First, a semantic classifier determines the input task state (e.g., whether it is multi-round reasoning or requires valuation). It then selects the corresponding Transformer channel or output head activation path. Furthermore, external task state embedding (e.g., the previous round output) is used as a control signal input to enhance state retention. Optionally, this can be implemented using a task attention network, a control gate network, and a state retention cache mechanism.

[0184] like Figure 6 As shown, the Transformer backbone network is used to process the contextual relevance of all input information and extract general and task-specific semantics. In this layer, the standard Transformer architecture is used, with 12 to 24 layers stacked, each consisting of a multi-head self-attention network and a feed-forward network (FFN).

[0185] Multi-tasking output headers include:

[0186] like Figure 7As shown in the figure, the natural language output head consists of a Transformer decoder structure + softmax; it outputs suggestion sentences, guiding words, etc.; and is used to generate structured language responses.

[0187] like Figure 8 As shown in the figure, the causal chain generates a header, predicting a triple (cause, relationship, result) at each time step; using Pointer Network + path generator; and generating a structured JSON graph output.

[0188] like Figure 9 As shown in the figure, the multi-label classification head and multiple independent sigmoid layers are used to distinguish risk types (structural risks, historical accidents, improper maintenance); layer normalization + dropout are used to prevent overfitting.

[0189] like Figure 10 As shown in the figure, the residual value estimation regression head uses an MLP structure consisting of a fully connected layer + ReLU + Dropout layer, with the input being the Transformer aggregate vector + numerical parameter embedding; it outputs the residual value estimate, the estimation interval, and the confidence level.

[0190] The proposed large-scale model is specifically tailored for real-world automotive risk assessment scenarios. It features five key capabilities: multi-source data fusion, structured semantic modeling, causal chain reasoning, multi-round interactive understanding, and residual value estimation output. This overcomes the technical bottlenecks of existing large-scale models, which are limited to natural language processing, lack structured output, and lack industry knowledge retention. MECR-LM not only understands heterogeneous inputs such as speech transcription, structured labels, maintenance records, and numerical parameters, but also dynamically plans risk reasoning paths and outputs structured causal chain graphs and valuation judgments. This allows language models to evolve from "text generators" to "domain knowledge-driven reasoning executors," making them particularly suitable for applications such as financial valuation, used car inspection, and residual value insurance.

[0191] In another embodiment, the present invention also provides a large model-based real-time automobile risk assessment method, which uses the aforementioned embodiment to perform hotel operation and management.

[0192] It should be noted that the explanation of the above-mentioned embodiment of the large-scale model-based automobile real-time risk assessment system is also applicable to the method of the embodiment of the present application and will not be repeated here.

[0193] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0194] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0195] In the several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0196] The above is only a specific embodiment of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of this application. For some module structures that are not particularly clear in the present invention, the content recorded in the prior art shall prevail. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be regarded as part of the present invention and is used to understand the meaning of some technical features or parameters.

Claims

1. A large-scale model-based vehicle real-time risk assessment system, characterized in that: The system comprises: A speech recognition module is used to receive speech input information from the assessor during the on-site vehicle inspection process and convert the speech input information into corresponding structured inspection item text; An inspection item parsing module, configured to extract component status information of a target vehicle based on the structured inspection item text, and map the component status information into a standard inspection tag; A multi-source information fusion module is used to obtain historical maintenance record data of the target vehicle and associate and fuse the historical maintenance record data with the component status information; The large model reasoning module, based on the trained large language model, conducts joint reasoning on the currently input inspection labels and the historical maintenance records, constructs a causal chain related to the accident risk, and generates at least one subsequent item to be inspected; A guidance prompt module, configured to generate voice guidance content based on the output result of the large model reasoning module and feed it back to the evaluator to prompt him to continue to inspect the target items of the target vehicle; A multi-round interaction module is used to repeat the speech recognition, inspection item analysis, information fusion and reasoning process after each round of inspection, gradually improving the risk inspection path of the target vehicle until the risk chain is closed; The residual value assessment module is used to call the valuation formula to calculate the estimated residual value of the target vehicle based on the final generated structured risk label, component status information, maintenance records, vehicle parameters and regional market parameters, and output a structured risk assessment report.

2. The large model-based automobile real-time risk assessment system according to claim 1, characterized in that: The speech recognition module includes: A voice acquisition module is used to collect the voice input signal of the assessor at the vehicle site and convert the voice input signal into audio format data; A speech recognition engine module, configured to transcribe the audio format data into text information; A terminology disambiguation module is used to perform standardized replacement and semantic disambiguation of professional terms, homophones, or synonyms in the transcribed text; The structured generation module is used to convert standardized text content into structured inspection item data, including part name, status description, location attribute and inspection type fields.

3. The large model-based real-time automobile risk assessment system according to claim 1, characterized in that: The inspection item parsing module includes: Semantic decomposition module, used to perform syntactic analysis on the structured inspection item text and extract the subject, predicate and description components; Field normalization module, used to perform semantic matching and normalization processing on the extracted fields and the preset vocabulary; The standard label mapping module is used to map the normalized field combination into a predefined standard inspection label, wherein the label includes the fields of component, location, status, and risk level.

4. The large model-based real-time automobile risk assessment system according to claim 1, characterized in that: The multi-source information fusion module includes: A historical data acquisition module is used to obtain the historical maintenance records of the target vehicle through the vehicle's unique identification code; Data preprocessing module, used to standardize the format and structure the fields of maintenance records; Semantic association module, used to semantically match maintenance information with the current component status, including multi-level matching methods at the component level, time level, and behavior level; The fusion construction module is used to construct a semantic fusion graph containing maintenance events and current status, and represent the potential causal path between event nodes in the form of a graph structure.

5. The large model-based real-time automobile risk assessment system according to claim 1, characterized in that: The guidance prompt module includes: The instruction generation module is used to convert the items to be checked in the large model inference results into natural language prompts; A multimodal output module, configured to present the prompts to the assessor's device terminal via voice broadcast or graphic display; The status synchronization module is used to record the response status of the prompt items and update the completion status of the check items; The priority scheduling module is used to dynamically arrange the output order of prompt content based on risk score or task urgency.

6. The large model-based vehicle real-time risk assessment system according to claim 1, characterized in that: The multi-round interaction module includes: Round control module, used to maintain inspection rounds and the input and output relationships corresponding to each round; The result confirmation module is used to confirm whether each round of voice input covers the guidance content and standardize it into a check label; Dynamic path construction module, used to call the reasoning module to update the risk path map; The closed-loop judgment module is used to determine whether the risk chain has reached the closed-loop condition and control the termination of the guidance process.

7. The large model-based real-time automobile risk assessment system according to claim 1, characterized in that: The residual value assessment module includes: Parameter aggregation module, which aggregates structured labels, vehicle parameters, and market reference data to generate valuation input vectors; Valuation rule configuration module, used to set the residual value estimation formula and the weights of influencing factors; The residual value calculation module is used to call the regression model or valuation logic to generate residual value estimates and confidence intervals; The evaluation report generation module is used to output a structured evaluation report containing valuation results, deduction reasons, risk level and recommended strategies.

8. The large model-based real-time automobile risk assessment system according to claim 1, characterized in that: The large language model used by the large model reasoning module includes: A unified input encoding layer, which converts the speech transcription text, structural labels, historical records, and numerical parameters into a unified embedding vector; The structured semantic adaptation layer is used to semantically model the relationship between structural labels using the graph attention mechanism; The task routing layer is used to dynamically schedule different task paths based on the input semantic intent; Transformer backbone network, used for context information fusion; Multi-task output head, used to generate multiple semantic outputs separately.

9. The large model-based real-time automobile risk assessment system according to claim 8, characterized in that: The multi-task output head includes: Natural language output head, used to output diagnostic suggestions, prompts, or question-and-answer results; The causal chain generation header is used to build the logical path between check items; Multi-label classification head, used to judge the dimensions of structural damage, historical accidents, and maintenance traces; The residual value estimation regression head is used to output the estimated value and credible interval.

10. A method for real-time automobile risk assessment based on a large model, characterized in that: The method uses the large model-based automobile real-time risk assessment system as described in any one of claims 1 to 9 to perform automobile real-time risk assessment.

Images