Second-hand car live broadcast car condition question answering and risk prediction system driven by large model

Abstract

The invention discloses a second-hand car live broadcast vehicle condition question answering and risk prediction system driven by a large model, and relates to the technical field of second-hand car live broadcast, and the technical scheme is characterized in that the method comprises the steps: collecting the actual static detection data and actual dynamic operation data of a vehicle in the second-hand car live broadcast process; performing risk feature analysis on the actual vehicle static detection data and the actual dynamic operation data to obtain a vehicle condition risk assessment result; the method comprises the following steps: acquiring user question information and historical question and answer data in a live question and answer process, performing question intention recognition according to the user question information to obtain a question risk level, constructing an answer strategy library in combination with the historical question and answer data, generating risk response answers, and comparing the risk response answers with a standard answer template to obtain an early warning risk answer mark; extracting historical compliance answer data and current live broadcast answer data based on the early warning risk answer mark; the effect is that live broadcast questions and answers are more professional and credible.

Description

Technical Field

[0001] This invention relates to the field of used car live streaming technology, and more specifically, to a large model-driven used car live streaming vehicle condition Q&A and risk prediction system. Background Technology

[0002] In the used car trading scenario, live streaming has gradually become an important channel for facilitating transactions due to its advantages of intuitive display and real-time interaction. However, the current used car live streaming process has many pain points, creating an urgent need for a vehicle condition Q&A and risk prediction system. Traditional methods lack intelligent recognition of user question intent during live Q&A sessions, making it difficult to promptly assess the risk level of questions; historical Q&A data is not effectively accumulated and reused, failing to provide broadcasters with standardized and accurate answering strategies; some broadcasters, due to insufficient professional knowledge and weak risk awareness, are prone to misleading statements and concealing key information in their answers; and the lack of a comparison mechanism between Q&A content and standard templates makes it difficult to provide timely warnings when risky answers occur. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide a large-model-driven live-streamed used car condition Q&A and risk prediction system.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A large-scale model-driven live-streamed used car condition Q&A and risk prediction system includes: Data Acquisition Module: Collects actual static inspection data and actual dynamic operation data of vehicles during the live broadcast of used cars. At the same time, it uses cameras and voice recognition devices to collect images of the vehicle's exterior, audio of the host's explanation, and user interactive comments in the live broadcast scene to form a multimodal raw dataset. The first analysis module analyzes the risk characteristics of actual vehicle static inspection data and actual dynamic operation data to obtain vehicle condition risk assessment results. Based on the multimodal raw dataset, it constructs a correlation mapping model between vehicle component defects and image features to identify hidden risk points in the appearance images. Through visual feature extraction algorithms, it identifies hidden risk points in the appearance images such as paint color difference and abnormal weld points. Processing module: Collects user questions and historical Q&A data during live Q&A sessions, analyzes the user questions and historical Q&A data to generate warning risk answer tags; Extraction module: Based on the warning risk answer tags, extract historical compliant answer data and current live answer data, and extract video clips, audio features and bullet screen keywords related to risk answers from the multimodal raw dataset to form a multimedia risk case library; The second analysis module analyzes the risk warning response markers to obtain the first response risk warning information, the second response risk warning information, and the response safety misjudgment information; Output module: After processing the vehicle condition risk assessment results, the first answer risk warning information, the second answer risk warning information, and the answer safety misjudgment information into risk control, output the live vehicle condition Q&A risk control results, generate a visual risk knowledge graph interactive interface, support the anchor to query risk answer cases through voice commands, and automatically push historical compliant answer voice segments that match the current vehicle condition.

[0005] Analyzing the risk warning response markers yields the first response risk warning information, the second response risk warning information, and the response safety misjudgment information, specifically including the following steps: Historical compliant answer data and current live stream answer data are extracted based on the risk warning answer markers; The comparison results are obtained by comparing historical compliant answer data with current live answer data; If the comparison results show the same situation, output the first answer risk warning information; If there are discrepancies in the comparison results, the second response risk warning information and the response safety misjudgment information will be output.

[0006] Preferably, the actual static inspection data and actual dynamic operation data of the vehicle are collected during the live broadcast of the used car business, specifically including the following steps: Initial static inspection data is obtained by inspecting and processing the vehicle's exterior, interior, and engine components. Initial static risk data is obtained by identifying risk features from the initial static detection data; Collect the vehicle's historical maintenance and repair records, and compare these records with the initial static risk data to obtain the data comparison results; the data comparison results include the difference data portion and the same data portion; Based on the difference data section, the detection dimensions are set for static and dynamic detection points respectively to obtain the static detection dimension and the dynamic detection dimension; Pre-set matching detection indicators based on initial static risk data; Based on the matching detection indicators, the actual static detection data of static detection points under static detection dimension conditions are statistically analyzed, and the actual dynamic operation data of dynamic detection points under dynamic detection dimension conditions are statistically analyzed.

[0007] Preferably, the vehicle condition risk assessment result is obtained by analyzing the risk characteristics of actual vehicle static inspection data and actual dynamic operation data, specifically including the following steps: Standard static detection data and standard dynamic operation data are obtained based on the matching detection indicators and initial static risk data; The actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance test data, and non-compliance test data; Based on the substandard detection data, the corresponding detection dimensions that need to be adjusted are selected from the static and dynamic detection dimensions. The data of the detection dimensions that need to be adjusted is adjusted to obtain the detection dimension adjustment results. The adjusted detection data is obtained by statistical analysis based on the detection dimension adjustment results. Based on the non-compliant test data, the corresponding comparison standard test data is selected from the standard static test data and standard dynamic operation data. When the adjusted test data is the same as the comparison standard test data, the comparison risk characteristics are identified based on the test dimension adjustment results, the adjusted test data and the compliant test data. The risk features and differences between the comparative risk features and the difference data are integrated to obtain the integrated risk features. The vehicle condition risk assessment result is obtained by combining the integrated risk features and the identical data.

[0008] Preferably, the actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance inspection data, and non-compliance inspection data, specifically including the following steps: When the actual static test data is the same as the standard static test data, and the actual dynamic operation data is the same as the standard dynamic operation data, the initial static risk data is determined to be the final vehicle condition basic risk data. When the actual static test data is the same as the standard static test data, but the actual dynamic operation data is different from the standard dynamic operation data, or when the actual static test data is different from the standard static test data, but the actual dynamic operation data is the same as the standard dynamic operation data, then the qualified test data and the unqualified test data will be output.

[0009] Preferably, the risk warning answer marker is obtained by analyzing user question information and historical question and answer data, specifically including the following steps: Pre-set risk assessment indicators for questions based on historical user questions and historical Q&A data; Question intent analysis data is obtained by identifying the intent of user questions based on question risk assessment indicators. The risk level of a question is determined by analyzing the question intent data. Based on the question risk level and historical question and answer data, a response strategy library is constructed and risk response responses are generated. The risk response responses are compared with standard response templates to obtain warning risk response markers.

[0010] Preferably, a response strategy library is constructed based on the risk level of the question and historical question-and-answer data, and risk response responses are generated. This specifically includes the following steps: Collect compliant answer models and answer risk case libraries from historical Q&A data; Based on the vehicle condition risk assessment results and user question information, the response effect of the compliance response model after dealing with historical risk cases is predicted to obtain a response strategy library; Based on the risk level of the question, a corresponding risk response strategy is selected from the response strategy library to generate a risk response answer.

[0011] Preferably, the risk response is compared with a standard response template to obtain a warning risk response marker, which specifically includes the following steps: After extracting the semantic features of risk response responses, we obtain the semantic features of risk response responses. The semantic features of risk response answers are compared with the semantic features of standard response templates to obtain the semantic comparison results of the answers; Extract the answer risk difference segment that exceeds the preset answer risk range value from the answer semantic comparison results; After marking the answer risk difference segment with an answer risk warning mark, an early warning risk answer mark is obtained.

[0012] Preferably, if there are discrepancies in the comparison results, a second response risk warning and a response safety misjudgment information are output, specifically including the following steps: If there are answer risk difference segments, then answer data will be collected again for the answer risk difference segments to obtain the answer dataset for the difference segments; The difference range of the answer data is obtained by comparing the answer data of the difference segment with the standard answer dataset for the corresponding segment; The second answer risk warning information is obtained by extracting the answer risk warning segment corresponding to the answer data whose difference in answer data falls within the range of the answer data difference. The answer risk warning marker segment corresponding to the answer data whose difference range does not fall within the range of answer data difference values ​​is extracted to obtain answer safety misjudgment information.

[0013] Compared with the prior art, the present invention has the following beneficial effects: This invention uses historical data to pre-set risk indicators to identify user question intent and risk levels, constructing a dynamically updated answer strategy library that makes risk response answers more relevant to real-world scenarios. Semantic feature comparison and early warning marking mechanisms standardize answer content from the source, avoiding misleading statements and identifying real risks and safety misjudgments. This provides compliance guidance for broadcasters and reduces false warnings caused by semantic bias, making live Q&A more professional and credible, thus improving the user communication experience and contributing to a healthy live Q&A ecosystem. It achieves full-process control over vehicle condition and Q&A risks. This collaborative mechanism significantly improves the operational efficiency of used car live streaming: merchants can rely on the system to quickly complete vehicle condition inspections and risk predictions, thereby optimizing their live streaming scripts. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the modules of the large-model-driven used car live-stream vehicle condition Q&A and risk prediction system proposed in this invention. Figure 2 This is a schematic diagram illustrating the steps for obtaining vehicle condition risk assessment results in the large-model-driven used car live-stream vehicle condition Q&A and risk prediction system proposed in this invention. Figure 3 This diagram illustrates the steps involved in obtaining the risk level of a problem in the large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system proposed in this invention. Detailed Implementation

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0017] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0018] Reference Figures 1-3 As shown.

[0019] The embodiments further illustrate the large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system proposed in this invention.

[0020] A large-scale model-driven live-streamed used car condition Q&A and risk prediction system includes: Data Acquisition Module: Collects actual static inspection data and actual dynamic operation data of vehicles during the live broadcast of used cars. At the same time, it uses cameras and voice recognition devices to collect images of the vehicle's exterior, audio of the host's explanation, and user interactive comments in the live broadcast scene to form a multimodal raw dataset. The first analysis module analyzes the risk characteristics of actual vehicle static inspection data and actual dynamic operation data to obtain vehicle condition risk assessment results. Based on the multimodal raw dataset, it constructs a correlation mapping model between vehicle component defects and image features to identify hidden risk points in the appearance images. Through visual feature extraction algorithms, it identifies hidden risk points in the appearance images such as paint color difference and abnormal weld points. Processing module: Collects user questions and historical Q&A data during the live Q&A session, analyzes the user questions and historical Q&A data to obtain warning risk answer tags; The second analysis module analyzes the risk warning response markers to obtain the first response risk warning information, the second response risk warning information, and the response safety misjudgment information; The system identifies the question's intent based on user question information to determine the question's risk level. It then builds a response strategy library by combining historical question and answer data and generates risk response answers. The risk response answers are compared with standard response templates to obtain warning risk response markers. Based on historical response data and vehicle condition risk assessment results, a used car knowledge network is constructed, and the system updates the relationships between entities of different fault types in real time.

[0021] The comparison module compares historical compliant answer data with current live answer data to obtain first answer risk warning information, second answer risk warning information, and answer security misjudgment information. At the same time, it compares the emotional matching degree between user bullet comments and answer content through sentiment analysis algorithm to identify user trust risks caused by semantic expression and generate sentiment deviation warning indicators. Output module: After processing the vehicle condition risk assessment results, the first answer risk warning information, the second answer risk warning information, and the answer safety misjudgment information into risk control, output the live vehicle condition Q&A risk control results, generate a visual risk knowledge graph interactive interface, support the anchor to query risk answer cases through voice commands, and automatically push historical compliant answer voice segments that match the current vehicle condition.

[0022] This application constructs a closed-loop risk management system for used car live-streaming scenarios. The data acquisition module captures static inspection (exterior, interior, etc.) and dynamic operation (driving status, etc.) data of vehicles during the live stream. The analysis module performs risk characteristic analysis on the collected static and dynamic data, identifies potential vehicle condition risks, and outputs vehicle condition risk assessment results.

[0023] This application captures detailed paint textures using an industrial-grade linear array camera (covering visible and near-infrared bands, 50 megapixels with 120fps frame rate) for static vehicle conditions (such as paint surface and parts appearance) during live streaming. Simultaneously, a 3D structured light scanner (with an accuracy of 0.05mm) reconstructs the vehicle's curved surface, accurately identifying microscopic deformations such as dents and gaps. Active polarized light sources are superimposed on dark and reflective areas (such as door handles and chrome strips) to eliminate specular reflections and ensure complete capture of paint details. For detecting implicit information such as paint thickness and putty layers, a terahertz time-domain spectrometer is used to penetrate the paint. Combined with a calibrated "optical-physical quantity" mapping model (using a micro-thickness gauge to establish a database linking different paint thicknesses to multispectral reflectivity), paint thickness and rust penetration depth are derived from the multispectral images captured during live streaming. Simultaneously, surface curvature data from 3D structured light is superimposed to correct measurement deviations caused by the vehicle's curvature.

[0024] The system collects over 150 ECU data points, including engine speed, transmission oil temperature, and fault codes, at a frequency of 10Hz via the OBD-II interface. It also records acceleration, braking, and steering attitude using an onboard inertial navigation module (IMU+GPS fusion, accuracy 0.1° / 0.5m). A high-speed camera (1000fps) is deployed to capture changes in door gaps caused by engine hood vibration during rapid acceleration and body roll during steering. Combined with an acoustic sensor array (microphone + voiceprint algorithm), the system identifies sound characteristics such as chassis noises and belt slippage, thus reconstructing the vehicle's dynamic operating status from multiple dimensions.

[0025] For scenarios requiring instrument-based testing, such as those involving paint thickness and uniformity, a "digital twin mapping" is performed on a standard vehicle with known parameters before the live stream to generate an error compensation matrix for the data acquisition equipment. This matrix corrects for deviations caused by instrument accuracy and ambient light fluctuations. During the live stream, when third-party thickness gauges or other equipment are connected, the instrument calibration parameters (including the last calibration time and error range) are recorded via blockchain and compared with the 3D point cloud and light field data collected by the proprietary equipment for multimodal consistency verification. If the deviation exceeds a threshold (e.g., 10%), manual review is triggered to ensure data reliability. To address issues of occlusion and limited viewing angles in the live stream, an 8K panoramic camera array and a multi-viewpoint stereo matching algorithm are used to synthesize images of occluded areas from different perspectives. This is combined with Neural Radiation Field (NeRF) technology to generate a complete 3D neural radiation field for the vehicle, supporting detailed viewing from any angle. The system predicts occlusion areas and triggers high-speed continuous shooting and pushes warnings to adjust the camera position, overcoming screen limitations to collect real data.

[0026] From the perspective of the entire data chain, the data acquisition end deploys a timestamp blockchain to add encrypted timestamps and device IDs to each frame of image and sensor data. Key parts are forced to be collected synchronously by multiple devices to form "evidence chain-level" data. The transmission end uses an edge computing + federated learning architecture. The on-site edge server performs homomorphic encryption preprocessing on sensitive data and only uploads feature values ​​to the cloud to ensure that the original data does not leave the local machine. The display end adds a traceability watermark to the vehicle condition data. Users can click to view the data lineage map containing blockchain hash values ​​and can also scan vehicles with AR using their mobile phones. The entire process is controlled to ensure that the live-streamed data is authentic and traceable.

[0027] The processing module collects user questions and historical Q&A data. It first identifies the user's question intent to determine the risk level, then uses historical data to build an answer strategy library to generate risk response answers. These answers are then compared with standard answer templates to mark potentially risky responses and standardize the live Q&A content. The extraction module filters historical compliant answers and current live Q&A data based on warning markers. The comparison module compares these two types of data to distinguish between different risk warning information and safety misjudgment information, identifying Q&A risks. The output module integrates vehicle condition risk assessment results and Q&A risk information, and after control processing, outputs the live vehicle condition Q&A risk control results.

[0028] Analyzing the risk warning response markers yields the first response risk warning information, the second response risk warning information, and the response safety misjudgment information, specifically including the following steps: Historical compliant answer data and current live stream answer data are extracted based on the risk warning answer markers; The comparison results are obtained by comparing historical compliant answer data with current live answer data; If the comparison results show the same situation, output the first answer risk warning information; If there are discrepancies in the comparison results, the second response risk warning information and the response safety misjudgment information will be output.

[0029] The comparison of historical compliant answer data and current live-stream answer data is a crucial step in identifying Q&A risks. Upon entering this comparison process, historical compliant answer data (verified to meet standards and accurately address vehicle condition issues) is first acquired, along with current live-stream answer data (real-time responses to user questions generated during the live stream). A detailed comparison is then performed. If the comparison results show identical situations, it means the current live-stream answer overlaps with answers previously identified as having risk warnings, and a first-level risk warning is output to alert users to these recurring risk responses. Conversely, if the comparison results show discrepancies, a second-level risk warning is output to alert users to content in the current live-stream answer that differs from historical compliant answers and may harbor new risks. Furthermore, since these discrepancies may be due to reasonable changes or misjudgments, a safety misjudgment information is also output to facilitate further investigation and determination of whether the risk is genuine or due to misjudgment caused by specific contexts or reasonable wording adjustments. This establishes a layered and precise Q&A risk identification and warning mechanism, providing strong support for risk management in used car live-stream vehicle condition Q&A.

[0030] Collecting actual static inspection data and actual dynamic operation data of vehicles during live-streaming of used cars includes the following steps: Initial static inspection data is obtained by inspecting and processing the vehicle's exterior, interior, and engine components. Initial static risk data is obtained by identifying risk features from the initial static detection data; Collect the vehicle's historical maintenance and repair records, and compare these records with the initial static risk data to obtain the data comparison results; the data comparison results include the difference data portion and the same data portion; Based on the difference data section, the detection dimensions are set for static and dynamic detection points respectively to obtain the static detection dimension and the dynamic detection dimension; Pre-set matching detection indicators based on initial static risk data; Based on the matching detection indicators, the actual static detection data of static detection points under static detection dimension conditions are statistically analyzed, and the actual dynamic operation data of dynamic detection points under dynamic detection dimension conditions are statistically analyzed.

[0031] This application first conducts inspections on the vehicle's exterior, interior, and engine components, compiling initial static inspection data based on actual observations and measurements such as paint condition, interior wear, and engine part appearance. Next, it extracts potentially risky features from this initial static inspection data, such as paint repair marks and risk indicators corresponding to large-area interior damage, transforming these into initial static risk data to mark potential vehicle condition hazards.

[0032] Collect the vehicle's past maintenance and repair records, including information such as previously repaired parts, replaced parts, and maintenance cycles. Compare these records with the initial static risk data. During the comparison, identify discrepancies (e.g., new wear on engine parts not reflected in the historical record) and similarities (e.g., wear on a certain interior part matches the historical record). Based on the discrepancies, define inspection dimensions for both static inspection points (such as exterior paint and interior seats) and dynamic inspection points (such as engine operation and vehicle driving status). Clearly define the specific angles and standards from which static inspection dimensions (e.g., paint inspection should cover thickness and color uniformity) and dynamic inspection dimensions (e.g., engine inspection should include idle stability and RPM fluctuation range) will be obtained.

[0033] Simultaneously, based on the risk points and characteristics in the initial static risk data, preset matching detection indicators are established to determine specific numerical standards for measuring whether the vehicle's condition is normal and whether any risks exist. Finally, according to the matched detection indicators, the actual static detection data of static detection points under static detection dimension conditions (such as whether the actual measured paint thickness meets the preset indicators) and the actual dynamic operation data of dynamic detection points under dynamic detection dimension conditions (such as whether the actual tested engine idle speed is within the preset stable range) are statistically analyzed. This comprehensively and accurately collects the static and dynamic data of the vehicle during the used car live broadcast, providing a foundation for subsequent vehicle condition analysis and risk prediction.

[0034] The vehicle condition risk assessment result is obtained by analyzing the risk characteristics of actual vehicle static inspection data and actual dynamic operation data, which specifically includes the following steps: Standard static detection data and standard dynamic operation data are obtained based on the matching detection indicators and initial static risk data; The actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance test data, and non-compliance test data; Based on the substandard detection data, the corresponding detection dimensions that need to be adjusted are selected from the static and dynamic detection dimensions. The data of the detection dimensions that need to be adjusted is adjusted to obtain the detection dimension adjustment results. The adjusted detection data is obtained by statistical analysis based on the detection dimension adjustment results. Based on the non-compliant test data, the corresponding comparison standard test data is selected from the standard static test data and standard dynamic operation data. When the adjusted test data is the same as the comparison standard test data, the comparison risk characteristics are identified based on the test dimension adjustment results, the adjusted test data and the compliant test data. The risk features and differences between the comparative risk features and the difference data are integrated to obtain the integrated risk features. The vehicle condition risk assessment result is obtained by combining the integrated risk features and the identical data.

[0035] In the used car live-stream vehicle condition risk assessment process, this application constructs standard static inspection data (data that the vehicle static inspection should meet under ideal conditions) and standard dynamic operation data (data that the vehicle dynamic operation should comply with) based on matching inspection indicators (pre-set vehicle condition inspection measurement standards) and initial static risk data (pre-identified potential static risk information of the vehicle).

[0036] Next, the actual vehicle static inspection data (the static data of the vehicle actually measured during the live broadcast), the actual dynamic operation data (the dynamic operation data of the vehicle actually collected), and the aforementioned standard data are compared and processed. By comparing, the final basic vehicle condition risk data (data reflecting the basic risk status of the vehicle), the compliant test data (the data portion of the actual data that meets the standard), and the non-compliant test data (the data portion of the actual data that does not meet the standard) are distinguished, thus clarifying the degree of conformity between the vehicle test data and the standard.

[0037] For the non-compliant test data, select the corresponding test dimensions that need to be adjusted (i.e. the test category to which the non-compliant data belongs) from static test dimensions (such as exterior and interior test angles) and dynamic test dimensions (such as engine operation and vehicle driving test angles). Adjust the data of these dimensions to obtain the test dimension adjustment results, and then compile the adjusted test data based on these results.

[0038] Then, based on the substandard test data, corresponding comparison standard test data (the standard data that should match the substandard data) are selected from the standard static test data and standard dynamic operation data. If the adjusted test data is consistent with the comparison standard test data, the comparison risk characteristics (the risk characteristics presented during and after the adjustment process) are identified by combining the test dimension adjustment results, the adjusted test data, and the compliant test data, and the potential vehicle condition risks behind the adjustment are explored.

[0039] Finally, the risk characteristics and difference data (data that differs from historical data and other data discovered during the initial collection and comparison process) are integrated to obtain the integrated risk characteristics. This integrated risk characteristics are then combined with the identical data (data that is consistent with actual data and standard data, etc.) to comprehensively evaluate the various risk situations of the vehicle and finally output the vehicle condition risk assessment result, providing a comprehensive and accurate basis for judging vehicle condition risks in used car live broadcasts.

[0040] The actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance inspection data, and non-compliance inspection data. This process includes the following steps: When the actual static test data is the same as the standard static test data, and the actual dynamic operation data is the same as the standard dynamic operation data, the initial static risk data is determined to be the final vehicle condition basic risk data. When the actual static test data is the same as the standard static test data, but the actual dynamic operation data is different from the standard dynamic operation data, or when the actual static test data is different from the standard static test data, but the actual dynamic operation data is the same as the standard dynamic operation data, then the qualified test data and the unqualified test data will be output.

[0041] This application compares and analyzes actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data to determine vehicle condition risk-related data. First, it acquires actual vehicle static inspection data (data on the static state of the vehicle's exterior, interior, engine components, etc.), actual dynamic operation data (data on dynamic processes such as vehicle driving and engine operation), and pre-set standard static inspection data (static reference data conforming to vehicle condition standards) and standard dynamic operation data (dynamic reference data conforming to vehicle condition standards).

[0042] If the actual static test data is completely consistent with the standard static test data, and the actual dynamic operation data is also consistent with the standard dynamic operation data, this indicates that the current static and dynamic vehicle conditions meet the standards. In this case, the initial static risk data is determined as the final basic risk data for the vehicle condition, which means that the initially identified static risk situation can be used as the basis for the current basic risk of the vehicle condition.

[0043] This means that the actual static test data is the same as the standard static test data, but the actual dynamic operating data is different from the standard dynamic operating data; or the actual static test data is different from the standard static test data, but the actual dynamic operating data is the same as the standard dynamic operating data. This indicates that the vehicle's static or dynamic condition does not meet the standard. In this case, the system will output compliant test data (the portion of the actual data that meets the standard) and non-compliant test data (the portion of the actual data that does not meet the standard). This clearly distinguishes between the compliant parts of the vehicle's condition and the non-compliant parts that pose risks and require attention. This provides accurate data support for subsequent vehicle condition risk assessment and judgment, allowing assessors to intuitively understand the degree of conformity between the vehicle's static and dynamic condition and the standard, and thus accurately identify vehicle condition risk points.

[0044] After analyzing user questions and historical Q&A data, risk warning answer tags are generated, which includes the following steps: Pre-set risk assessment indicators for questions based on historical user questions and historical Q&A data; Question intent analysis data is obtained by identifying the intent of user questions based on question risk assessment indicators. The risk level of a question is determined by analyzing the question intent data. Based on the question risk level and historical question and answer data, a response strategy library is constructed and risk response responses are generated. The risk response responses are compared with standard response templates to obtain warning risk response markers.

[0045] This application analyzes historical user questions (various questions raised by users in past live streams regarding the condition of used cars, such as "Has the vehicle been in a major accident?" or "How many times has the engine been repaired?") and historical Q&A data (corresponding to the answers to these questions and the risk assessment feedback after the answers). From this historical data, risk assessment indicators are extracted and preset. For example, questions involving keywords such as "major accident" and "engine overhaul" are set as assessment indicators for high-risk questions.

[0046] When user questions are generated during the live stream, the intent of the questions is identified based on pre-set question risk assessment indicators. This means analyzing which aspect of the vehicle the user is concerned about and what potential information they want to know, and converting this analysis into question intent analysis data. For example, if a user asks "Are there any abnormalities in the brakes?", the intent analysis data will mark this as a potential risk point related to a critical safety component of the vehicle.

[0047] Based on the analysis of question intent data, the system identifies the corresponding risk level according to the preset risk level classification rules (such as questions involving core safety components being high-risk, and inquiries about routine configurations being low-risk). This allows the system to predict the risk level of user questions in advance in live Q&A scenarios, providing a basis for generating reasonable and compliant risk response answers, and ensuring that live Q&A sessions for used cars meet user needs.

[0048] Based on the risk level of the question and historical Q&A data, a response strategy library is constructed and risk response responses are generated. The specific steps include: Collect compliant answer models and answer risk case libraries from historical Q&A data; Based on the vehicle condition risk assessment results and user question information, the response effect of the compliance response model after dealing with historical risk cases is predicted to obtain a response strategy library; Based on the risk level of the question, a corresponding risk response strategy is selected from the response strategy library to generate a risk response answer.

[0049] This application selects and collects two important parts from historical Q&A data: compliant answer models (answer patterns in past live broadcasts that comply with industry standards, effectively answer vehicle condition questions, and do not cause risks, such as standard scripts that clearly explain the true condition of the vehicle and objectively respond to fault issues) and answer risk case library (cases that record risks such as disputes and misleading users due to inappropriate answers, such as answer examples of concealing the vehicle's accident history leading to subsequent complaints).

[0050] Based on the preliminary vehicle condition risk assessment results (a comprehensive judgment of the current static and dynamic vehicle condition risks) and real-time user question information (specific questions raised by users regarding vehicle conditions during live broadcasts), the compliance response model is applied to the response scenarios of historical risk cases. The model is simulated and predicted to handle the response effect of such risk questions. For example, it is judged whether responding to questions related to "the vehicle has been in a minor accident" with compliance language can clearly explain the situation and eliminate user concerns. Through a large number of similar simulations, a response strategy library is integrated to form a library containing effective response strategies corresponding to different vehicle condition risks and different question types.

[0051] Once the risk level (e.g., high, medium, low) of a user's question is identified during the live stream, the system matches and selects the corresponding response strategy from the response strategy library. Specific risk response answers are then generated according to the strategy, ensuring that user concerns are accurately addressed during live Q&A sessions while mitigating risks arising from inappropriate responses. This makes car condition Q&A sessions in used car live streams more standardized and more risk-resistant.

[0052] The risk response responses are compared with the standard response template to obtain early warning risk response markers. This process includes the following steps: After extracting the semantic features of risk response responses, we obtain the semantic features of risk response responses. The semantic features of risk response answers are compared with the semantic features of standard response templates to obtain the semantic comparison results of the answers; Extract the answer risk difference segment that exceeds the preset answer risk range value from the answer semantic comparison results; After marking the answer risk difference segment with an answer risk warning mark, an early warning risk answer mark is obtained.

[0053] This application extracts the semantic features of the generated risk response answers, that is, it mines the key semantic information in the answers, such as whether the description of vehicle faults is accurate and whether the explanation of vehicle condition risks is clear, to form the semantic features of risk response answers.

[0054] By comparing the semantic features of these risk response answers with the semantic features of the standard response template (pre-defined semantic specifications for answers that meet compliance and accuracy requirements, such as the semantic standard that requires a truthful and comprehensive description of known vehicle issues), the semantic fit between the two is judged from a semantic perspective to obtain the answer semantic comparison results, and to understand the semantic gap between the risk response answers and the standard.

[0055] Next, based on the preset answer risk range value (a pre-set threshold for how much semantic deviation from the standard constitutes a risk range), the answer risk difference segment that exceeds this range value is selected from the answer semantic comparison results. In other words, the part of the answer that deviates significantly from the standard and may pose a risk is identified.

[0056] Finally, risk warning tags are generated for these risk-differentiated answer segments to identify answers that may have semantic issues, mislead users, or other risks during live Q&A sessions, providing clear identification for further risk management in Q&A.

[0057] If discrepancies are found in the comparison results, a risk warning message for the second answer and a safety misjudgment message for the answer will be output, specifically including the following steps: If there are answer risk difference segments, then answer data will be collected again for the answer risk difference segments to obtain the answer dataset for the difference segments; The difference range of the answer data is obtained by comparing the answer data of the difference segment with the standard answer dataset for the corresponding segment; The second answer risk warning information is obtained by extracting the answer risk warning segment corresponding to the answer data whose difference in answer data falls within the range of the answer data difference. The answer risk warning marker segment corresponding to the answer data whose difference range does not fall within the range of answer data difference values ​​is extracted to obtain answer safety misjudgment information.

[0058] When this application finds discrepancies between historical compliant answer data and current live-stream answer data, it needs to identify the second answer risk warning information and the answer security misjudgment information. If there are answer risk difference segments (i.e., parts that deviate from the standard in the answer semantic comparison), answer data for these difference segments will be collected again to form a difference segment answer dataset.

[0059] Next, the difference segment answer dataset is compared with the corresponding segments of the standard answer dataset to calculate the difference in content, semantics, etc., constructing a set of answer data difference ranges to quantify the degree of difference. Based on preset answer data difference range intervals (pre-defined numerical ranges for judging risk or misjudgment), the difference range set is classified. Data falling within the interval is marked with a corresponding answer risk warning segment, indicating a real risk, and extracted as secondary answer risk warning information, indicating that such answers have risks that require attention. Data not falling within the interval is marked with a corresponding answer risk warning segment, indicating a misjudgment due to semantic deviation, special context, etc., and extracted as answer safety misjudgment information to avoid over-warning of normal answers. This improves the accuracy of risk management in used car live Q&A sessions.

[0060] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large-scale model-driven live-streamed used car condition Q&A and risk prediction system, characterized by: include: Data Acquisition Module: Collects actual static inspection data and actual dynamic operation data of vehicles during the live broadcast of used cars, and collects vehicle exterior images, host explanation audio and user interactive bullet comments in the live broadcast scene to form a multimodal raw dataset; The first analysis module analyzes the risk characteristics of actual vehicle static inspection data and actual dynamic operation data to obtain vehicle condition risk assessment results. Based on the multimodal raw dataset, it constructs a correlation mapping model between vehicle component defects and image features to identify hidden risk points in the appearance images. Processing module: Collects user questions and historical Q&A data during the live Q&A session, analyzes the user questions and historical Q&A data to obtain warning risk answer tags; The second analysis module analyzes the risk warning response markers to obtain the first response risk warning information, the second response risk warning information, and the response safety misjudgment information; Output module: After processing the vehicle condition risk assessment results, the first response risk warning information, the second response risk warning information, and the response safety misjudgment information into risk control, output the live vehicle condition Q&A risk control results.

2. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 1, characterized in that, Analyzing the risk warning response markers yields the first response risk warning information, the second response risk warning information, and the response safety misjudgment information, specifically including the following steps: Historical compliant answer data and current live stream answer data are extracted based on the risk warning answer markers; The comparison results are obtained by comparing historical compliant answer data with current live answer data; If the comparison results show the same situation, output the first answer risk warning information; If there are discrepancies in the comparison results, the second response risk warning information and the response safety misjudgment information will be output.

3. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 2, characterized in that, Collecting actual static inspection data and actual dynamic operation data of vehicles during live-streaming of used cars includes the following steps: Initial static inspection data is obtained by inspecting and processing the vehicle's exterior, interior, and engine components. Initial static risk data is obtained by identifying risk features from the initial static detection data; Collect the vehicle's historical maintenance and repair records, and compare these records with the initial static risk data to obtain the data comparison results; the data comparison results include the difference data portion and the same data portion; Based on the difference data section, the detection dimensions are set for static and dynamic detection points respectively to obtain the static detection dimension and the dynamic detection dimension; Pre-set matching detection indicators based on initial static risk data; Based on the matching detection indicators, the actual static detection data of static detection points under static detection dimension conditions are statistically analyzed, and the actual dynamic operation data of dynamic detection points under dynamic detection dimension conditions are statistically analyzed.

4. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 3, characterized in that, The vehicle condition risk assessment result is obtained by analyzing the risk characteristics of actual vehicle static inspection data and actual dynamic operation data, which specifically includes the following steps: Standard static detection data and standard dynamic operation data are obtained based on the matching detection indicators and initial static risk data; The actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance test data, and non-compliance test data; Based on the substandard detection data, the corresponding detection dimensions that need to be adjusted are selected from the static and dynamic detection dimensions. The data of the detection dimensions that need to be adjusted is adjusted to obtain the detection dimension adjustment results. The adjusted detection data is obtained by statistical analysis based on the detection dimension adjustment results. Based on the non-compliant test data, the corresponding comparison standard test data is selected from the standard static test data and standard dynamic operation data. When the adjusted test data is the same as the comparison standard test data, the comparison risk characteristics are identified based on the test dimension adjustment results, the adjusted test data and the compliant test data. The risk features and differences between the comparative risk features and the difference data are integrated to obtain the integrated risk features. The vehicle condition risk assessment result is obtained by combining the integrated risk features and the identical data.

5. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 4, characterized in that, The actual vehicle static inspection data, actual dynamic operation data, standard static inspection data, and standard dynamic operation data are processed to obtain the final vehicle condition basic risk data, compliance inspection data, and non-compliance inspection data. This process includes the following steps: When the actual static test data is the same as the standard static test data, and the actual dynamic operation data is the same as the standard dynamic operation data, the initial static risk data is determined to be the final vehicle condition basic risk data. When the actual static test data is the same as the standard static test data, but the actual dynamic operation data is different from the standard dynamic operation data, or when the actual static test data is different from the standard static test data, but the actual dynamic operation data is the same as the standard dynamic operation data, then the qualified test data and the unqualified test data will be output.

6. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 5, characterized in that, After analyzing user questions and historical Q&A data, risk warning answer tags are generated, which includes the following steps: Pre-set risk assessment indicators for questions based on historical user questions and historical Q&A data; Question intent analysis data is obtained by identifying the intent of user questions based on question risk assessment indicators. The risk level of a question is determined by analyzing the question intent data. Based on the question risk level and historical question and answer data, a response strategy library is constructed and risk response responses are generated. The risk response responses are compared with standard response templates to obtain warning risk response markers.

7. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 6, characterized in that, Based on the risk level of the question and historical Q&A data, a response strategy library is constructed and risk response responses are generated. The specific steps include: Collect compliant answer models and answer risk case libraries from historical Q&A data; Based on the vehicle condition risk assessment results and user question information, the response effect of the compliance response model after dealing with historical risk cases is predicted to obtain a response strategy library; Based on the risk level of the question, a corresponding risk response strategy is selected from the response strategy library to generate a risk response answer.

8. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 7, characterized in that, The risk response responses are compared with the standard response template to obtain early warning risk response markers. This process includes the following steps: After extracting the semantic features of risk response responses, we obtain the semantic features of risk response responses. The semantic features of risk response answers are compared with the semantic features of standard response templates to obtain the semantic comparison results of the answers; Extract the answer risk difference segment that exceeds the preset answer risk range value from the answer semantic comparison results; After marking the answer risk difference segment with an answer risk warning mark, an early warning risk answer mark is obtained.

9. The large-model-driven used car live-streaming vehicle condition Q&A and risk prediction system according to claim 8, characterized in that, If discrepancies are found in the comparison results, a risk warning message for the second answer and a safety misjudgment message for the answer will be output, specifically including the following steps: If there are answer risk difference segments, then answer data will be collected again for the answer risk difference segments to obtain the answer dataset for the difference segments; The difference range of the answer data is obtained by comparing the answer data of the difference segment with the standard answer dataset for the corresponding segment; The second answer risk warning information is obtained by extracting the answer risk warning segment corresponding to the answer data whose difference in answer data falls within the range of the answer data difference. The answer risk warning marker segment corresponding to the answer data whose difference range does not fall within the range of answer data difference values ​​is extracted to obtain answer safety misjudgment information.

Images