Virtual automobile exhibition system based on digital twinning

Abstract

The present application belongs to the technical field of virtual car exhibition, and particularly relates to a virtual car exhibition system based on digital twinning. The system comprises a digital twinning construction module, which generates a lightweight model and supports first-person perspective dynamic update; a user behavior analysis module, which maps data such as sight line trajectory into a weighted directed graph and predicts vehicle type and configuration demand by using a multi-head graph attention network; a personalized configuration selection module, which generates a configuration scheme in combination with digital twinning parameters; a virtual test drive module, which provides immersive experience based on the configuration scheme and the scene; and a background data analysis module, which integrates data to generate a user portrait to optimize exhibition content. The system realizes dynamic virtual-real mapping, accurate demand prediction and data-driven intelligent exhibition, and improves user interaction depth and vehicle purchasing efficiency.

Description

Technical Field

[0001] This invention belongs to the field of virtual auto show technology, specifically relating to a virtual auto show system based on digital twins. Background Technology

[0002] With the digital transformation of the automotive manufacturing and exhibition industries, traditional offline auto shows are increasingly unable to meet users' personalized needs and real-time interactive experiences due to limited space, overcrowding, and high display costs. Existing online auto shows based on virtual reality (VR) or augmented reality (AR) primarily rely on static 3D models, lacking dynamic interaction with real vehicles and the showroom environment. This makes it difficult to achieve precise configuration adjustments, real-time status updates, and immersive test drive experiences. Furthermore, these systems generally lack the ability to deeply analyze user behavior, passively presenting pre-set content without proactively recommending interested models, customized configurations, or the most suitable driving scenarios based on visitor browsing patterns, configuration preferences, and test drive habits. Digital twin technology, by constructing a virtual "twin" that corresponds one-to-one with the real vehicle and showroom environment, can load sensor data and business rules into the model in real time, providing users with more accurate visual rendering and physical simulation effects. Summary of the Invention

[0003] To address the technical problems existing in the above-mentioned background art, this invention proposes a virtual car exhibition system based on digital twins.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes a digital twin construction module, a user behavior analysis module, a personalized configuration selection module, a virtual test drive module, and a background data analysis module.

[0005] The digital twin construction module is used to construct a virtual digital twin model of a car exhibition. In the model, visitors can tour from a first-person perspective, and the model status is dynamically updated based on real-time sensor data.

[0006] The user behavior analysis module is used to collect user behavior and interaction data in virtual exhibitions. By analyzing the user behavior and interaction data, it predicts the user's vehicle model and configuration requirements. Specifically, this includes:

[0007] First, user behavior interaction data is collected, including gaze trajectory, gaze duration to characterize the user's visual focus; interface clicks and parameter fine-tuning records to quantify the user's configuration exploration; and three-dimensional movement trajectory and dwell time to reflect the user's browsing path and spatial exploration.

[0008] The aforementioned behavioral interaction data is mapped to a weighted directed graph. A multi-head graph attention network is then applied to this weighted directed graph to achieve deep learning of node features: each node's latent representation integrates the spatiotemporal correlation strength of its neighboring nodes; multiple attention heads can capture user attention patterns for different types of interactions in parallel; after multi-layer attention aggregation, the weighted sum of all node representations is taken to obtain a vector H reflecting global behavioral preferences. user ;

[0009] The vector reflecting global behavioral preferences will be combined with the description vector M of each vehicle model in the digital twin system. c By using bilinear attention fusion, a joint feature representation Z for the user vehicle model is generated. c The calculation method is as follows: Among them, W Q W K W V This represents the learnable mapping matrix of queries, keys, and values, where SF represents the scaling factor.

[0010] Generate DB behavioral comparison samples online, perturb the node attributes or spatial locations in the original map, and recalculate the corresponding joint feature Z′. c,db By comparing the original Z c With the perturbation Z′ c The differences were used to obtain the sensitivity index S for each model. c The calculation method is as follows: Where c represents the vehicle index and db represents the index of the behavior comparison samples;

[0011] S c The network π serves as a strategy for predicting input demand based on immediate preference rewards. θ (M c |H user And perform a parameter learning update once: Where θ represents the parameters of the policy network, η represents the learning rate, and θ′ represents the parameters of the updated policy network;

[0012] After the update, output the model requirement and select argmax. c π θ (M c |H user The configuration requirements are obtained in the same way, and argmax is selected. p π θ (M p |H user Among them, M p Represents each configuration vector;

[0013] The personalized configuration selection module is used to generate a personalized vehicle configuration scheme based on the user's vehicle model and configuration requirements, combined with the configurable parameters of the digital twin model.

[0014] The virtual test drive module is used to generate immersive driving experience data based on personalized vehicle configuration schemes and usage scenarios;

[0015] The background data analysis module is used to integrate user behavior interaction data and driving experience data, analyze user behavior patterns in virtual auto shows, combine user profiles with user personal information, generate user profiles, and optimize virtual show content and resource allocation based on these profiles.

[0016] Preferably, the method for constructing the model using the digital twin building module includes:

[0017] Acquire the 3D scan point cloud, CAD engineering data and material spectral parameters of the target vehicle, and simultaneously collect the laser point cloud data and building BIM model of the exhibition venue;

[0018] Lightweight digital twin models of the car and the stadium were generated using an octree mesh compression algorithm, and a spatial mapping algorithm was used to unify their coordinate systems.

[0019] Preferably, the digital twin model undergoes model optimization processing. This optimization process involves adaptively optimizing the rendering parameters of the viewpoint based on different first-person perspectives and distances. The specific implementation steps are as follows:

[0020] First, let the first-person perspective head be the origin coordinate P(x) p ,y p ,z p The viewpoint direction vector is The field of view of the cone is θ; the calculation model is based on the point V(x) v ,y v ,z v The relative position vector between the viewpoint and the viewpoint

[0021] Calculate the distance d from point V to the viewing angle P and the cosine value cosα of the angle between the viewing angle direction and the direction of point V. The cosine value cosα is calculated as follows: And determine the rendering weights as follows:

[0022] The LOD level is dynamically divided based on weights, with the LOD level set as L and a preset distance threshold interval [d0, d1, ... d]. n The corresponding level is [L0, L1, ... L]. n ], where d0 < d1 < ... < d nL0 > L1 > ... > L n When the distance d is in the interval [d i ,d i+1 When [the time is specified], the level L is calculated using linear interpolation, and the dynamic detail level value L(d) is calculated based on the distance d. The calculation method is as follows: After adjusting for directional weights, the final level is: L final =L(d)·ω d (α)+L min ·(1-ω d (α)), where L final L represents the final level. min This represents the lowest level of detail.

[0023] Based on L final Dynamically subdivide or merge octree networks when L final ≥L max When, enable full-precision mesh, when L final ≤L min Enable lightweight mesh when L min <L final <L max No processing is performed at this time; next, adjust the texture pixel size T as follows: Where T max This represents the highest resolution texture size.

[0024] Preferably, the virtual test drive module, used to generate immersive driving experience data based on personalized vehicle configuration schemes and usage scenarios, includes the following specific implementation steps:

[0025] First, the virtual test drive module loads the personalized vehicle configuration scheme obtained from the personalized configuration selection module into the digital twin virtual test drive scenario in sync with the target driving scenario;

[0026] Based on the user's throttle, brake and steering input intensity, the system assesses driving style in real time and automatically adjusts the difficulty of the scenario.

[0027] Finally, the user's operation data and vehicle response data are recorded throughout the process to obtain driving experience data.

[0028] Preferably, the personalized configuration selection module, used to generate a personalized vehicle configuration scheme based on the user's vehicle model and configuration requirements, combined with the configurable parameters of the digital twin model, includes the following specific implementations:

[0029] First, the personalized configuration selection module maps the vehicle model requirements and configuration requirements obtained from the user behavior analysis module to the parameter space of the digital twin model.

[0030] For each potential configuration combination, a digital twin model is invoked for rapid simulation evaluation, which includes performance simulation and comfort simulation, to obtain performance simulation scores and comfort simulation scores.

[0031] Finally, a multi-view optimization algorithm is used to obtain the optimal personalized configuration selection.

[0032] Preferably, the background data analysis module is used to integrate user behavior interaction data and driving experience data, analyze user behavior patterns in the virtual auto show, combine them with user profiles to generate user profiles, and optimize virtual show content and resource allocation based on these profiles. Specific implementation steps include:

[0033] The backend data analysis module synchronizes and aligns user behavior interaction data and driving experience data.

[0034] All data is cleaned and missing data is filled in to generate a high-quality multi-source fusion dataset;

[0035] Based on the fused data, unsupervised clustering and sequence pattern mining were applied in three dimensions: showroom browsing, vehicle type and configuration, and test drive experience.

[0036] The clustering and pattern results are cross-mapped with user profile data, including users' age, occupation, and car purchase budget, to generate user profiles.

[0037] Based on user profiles, the order of vehicles displayed at digital twin virtual exhibitions is adjusted to optimize users' car purchasing choices. Compared with existing technologies, the advantages and positive effects of this invention are:

[0038] 1. Construct digital twin models of vehicles and venues, update their status in real time, and combine octree compression and LOD adaptive rendering to achieve full precision for close-up views and lightweight rendering for distant views from a first-person perspective. This resolves the contradiction between model rendering efficiency and realism, and ensures high consistency between virtual scenes and physical interactions at real auto shows.

[0039] 2. By mapping behavioral data into a weighted directed graph, a multi-head graph attention network is used to capture spatiotemporal features. Combined with contrastive learning and policy networks to predict demand, recommendations are dynamically updated based on gaze duration, configuration adjustments, etc., resulting in high accuracy in predicting vehicle demand and enabling proactive recommendations.

[0040] 3. Integrate multi-source data, clean and cluster it, and map it to user profiles. Dynamically adjust the order and scenarios of exhibited vehicles, optimize the display content for different user groups, increase the dwell time of target users and the number of car purchase inquiries, and form a data-driven intelligent exhibition closed loop. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the overall modular structure of a virtual car exhibition system based on digital twins. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0044] Numerous 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 than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0045] For example, existing virtual reality-based online auto shows mostly rely on static 3D models, lacking dynamic interaction with real vehicles and the showroom environment. This prevents users from experiencing real-time vehicle status updates and makes precise configuration adjustments difficult. Furthermore, they generally lack in-depth analysis capabilities of user behavior, passively presenting pre-set content without recommending models based on visitor browsing history, configuration preferences, and test-driving habits. To address these issues, this invention proposes a virtual auto show system based on digital twins. This system constructs virtual "twins" that correspond one-to-one with real vehicles and the showroom environment, loading sensor data and business rules into the model in real time. This provides users with more accurate visual rendering and physical simulation effects, while simultaneously deeply analyzing user behavior to achieve personalized recommendations and an immersive experience. Figure 1 As shown, it includes a digital twin building module, a user behavior analysis module, a personalized configuration selection module, a virtual test drive module, and a back-end data analysis module;

[0046] The digital twin construction module is used to build a virtual digital twin model of a car exhibition. In this model, a first-person perspective is used for navigation, and the model's state is dynamically updated based on real-time sensor data. The method for constructing the model includes: acquiring the 3D scanned point cloud, CAD engineering data, and material spectral parameters of the target vehicle; simultaneously collecting laser point cloud data and a building BIM model of the exhibition venue; generating lightweight digital twin models of the car and the venue respectively using an octree mesh compression algorithm; and using a spatial mapping algorithm to unify the coordinate systems of the two. Specifically, firstly, a high-precision 3D scanning device is used to acquire the original point cloud of the target vehicle, and the precise geometric structure is extracted by combining the vehicle's CAD engineering files. Simultaneously, material spectral parameters are collected to obtain diffuse reflection, specular reflection, and normal map information required for physical rendering. Simultaneously, a laser scanner is used to collect full-scene laser point clouds of the exhibition venue, which are then imported into the building BIM model to obtain spatial elements such as walls, columns, beams, and exhibition stands. Secondly, the original point clouds of vehicles and venues were preprocessed, including voxel downsampling, statistical outlier removal, and region segmentation, to ensure point cloud quality and extract target regions. Then, an octree mesh compression algorithm was used to subdivide the preprocessed point clouds from top to bottom and merge voxels: within each subdivision unit, point clouds were merged based on a distance error threshold, while retaining higher resolution for important feature regions, thus generating a lightweight mesh that balances detail and performance. Subsequently, triangular mesh models of vehicles and venues were reconstructed based on the compressed point clouds, and the previously collected spectral parameters were mapped to PBR material maps to achieve high-fidelity visual effects. Finally, a spatial mapping algorithm was used to unify the coordinate systems of the two models: using the BIM model origin as the global reference, the venue point cloud was first registered with the BIM model, and then the Iterative Closest Point (ICP) algorithm was used to accurately map the vehicle mesh to the venue coordinate system, completing the construction of a digital twin model under the same coordinate system.

[0047] To ensure a more realistic view of the car from different distances in a first-person perspective, the rendering parameters of the viewpoint are adaptively optimized. This addresses the issue of excessive detail in distant models leading to a heavy rendering burden, and insufficient detail in close models affecting realism. The specific steps are as follows: First, the origin coordinate P(x...) is set as the head in the first-person perspective. p ,y p ,z p The viewpoint direction vector is The field of view of the cone is θ; the calculation model is based on the point V(x) v ,y v ,z v The relative position vector between the viewpoint and the viewpoint Calculate the distance d from point V to the viewing angle P and the cosine value cosα of the angle between the viewing angle direction and the direction of point V. The cosine value cosα is calculated as follows: And determine the rendering weights as follows: The LOD level is dynamically divided based on weights, with the LOD level set as L and a preset distance threshold interval [d0, d1, ... d]. n The corresponding level is [L0, L1, ... L]. n ], where d0 < d1 < ... < d n L0 > L1 > ... > L n When the distance d is in the interval [d i ,d i+1 When [the time is specified], the level L is calculated using linear interpolation, and the dynamic detail level value L(d) is calculated based on the distance d. The calculation method is as follows: After adjusting for directional weights, the final level is: L final =L(d)·ω d (α)+L min ·(1-ω d (α)), where L final L represents the final level. min Represented by the lowest level of detail; based on L final Dynamically subdivide or merge octree networks when L final ≥L max When, enable full-precision mesh, when L final ≤L min Enable lightweight mesh when L min <L final <L max No processing is performed at this time; next, adjust the texture pixel size T as follows: Where T max This represents the highest resolution texture size. Through this adaptive optimization scheme, the system can dynamically adjust the rendering details of the model according to changes in the user's viewing angle. When viewed up close from the user's perspective, the model displays full-precision details with clear textures, allowing for the precise observation of the vehicle's subtle features; when viewed from a distance, the model automatically simplifies, reducing the rendering burden and ensuring smooth system operation.

[0048] The user behavior analysis module is used to collect user behavior interaction data in virtual exhibitions. By analyzing user behavior interaction data, it predicts the user's vehicle model and configuration requirements. Specifically, it includes: firstly, collecting user behavior interaction data, which includes gaze trajectory, gaze duration (user's visual focus), interface clicks and parameter fine-tuning records (to quantify user configuration exploration), and three-dimensional movement trajectory and dwell time (to reflect user's browsing path and spatial exploration); mapping the above behavior interaction data into a weighted directed graph. Specifically, gaze points in the gaze trajectory, interaction areas of interface clicks, and key positions in the three-dimensional movement trajectory are set as nodes. Node attributes include behavioral characteristics such as gaze duration, number of clicks, and parameter fine-tuning amplitude. Edges are established based on the temporal sequence of behavior and spatial transfer path. The edge weights are calculated by combining behavior intensity and spatiotemporal correlation. For example, the longer the gaze duration and the more frequent the click operation, the higher the edge weight between the corresponding nodes; the closer the spatial distance of behavior at adjacent time points, the stronger the edge weight. This quantifies the spatiotemporal continuity and attention preferences of user behavior. This mapping method transforms the behavior data into a graph structure containing spatiotemporal semantics. Then, a multi-head graph attention network is applied to the weighted directed graph to achieve deep learning of node features: the latent representation of each node integrates the spatiotemporal correlation strength of its neighboring nodes; multiple attention heads can capture user attention patterns for different types of interactions in parallel; after multi-layer attention aggregation, the weighted sum of all node representations is taken to obtain a vector H reflecting global behavioral preferences. user The vector reflecting global behavioral preferences will be compared with the description vector M of each vehicle model in the digital twin system. c By using bilinear attention fusion, a joint feature representation Z for the user vehicle model is generated. c The calculation method is as follows: Among them, W Q W K W V The learnable mapping matrix for query, key, and value is represented by SF, which represents the scaling factor. DB behavioral comparison samples are generated online, perturbing the node attributes or spatial locations in the original graph, and the corresponding joint features Z′ are recalculated. c,db By comparing the original Z c With the perturbation Z′ c The differences were used to obtain the sensitivity index S for each model. c The calculation method is as follows: Where c represents the vehicle index, and db is the index of the behavior comparison sample; S c The network π serves as a strategy for predicting input demand based on immediate preference rewards. θ (M c |H user And perform a parameter learning update once: Where θ represents the parameters of the policy network, η represents the learning rate, and θ′ represents the parameters of the updated policy network; after the update, the output model requirement is determined, and argmax is selected. c π θ (M c |H user The configuration requirements are obtained in the same way, and argmax is selected. p π θ (M p |H user Among them, M p This module represents various configuration vectors. Breaking away from the traditional passive display model, it proactively uncovers users' potential needs, providing data support for personalized configuration recommendations and test drive experiences. Simultaneously, it uses behavioral pattern analysis to optimize virtual exhibition content, enhancing the depth of user interaction.

[0049] The personalized configuration selection module is used to generate personalized vehicle configuration schemes based on the user's vehicle model and configuration requirements, combined with configurable parameters of the digital twin model. First, the personalized configuration selection module maps the vehicle model and configuration requirements obtained from the user behavior analysis module to the parameter space of the digital twin model. For each potential configuration combination, the module calls the digital twin model for rapid simulation evaluation, including performance simulation and comfort simulation, obtaining performance simulation scores and comfort simulation scores. Finally, a multi-view optimization algorithm is used to obtain the optimal personalized configuration selection. Specifically, first, the module converts the vehicle model preferences and configuration requirements output by the user behavior analysis module into parameter space coordinates recognizable by the digital twin model. For example, if the user prefers a "sporty SUV," the system will map this to the value range of parameters such as engine displacement, suspension stiffness, and tire specifications, while combining configuration requirements (such as panoramic sunroof and heated seats) to form an initial parameter combination. Second, for each potential configuration combination, the module calls the digital twin model for bidirectional simulation evaluation. Performance simulation focuses on mechanical indicators such as power response, braking distance, and cornering roll, calculating performance scores by loading virtual operating conditions (such as high-speed cruising and continuous curves). Comfort simulation quantifies experiential parameters such as seat pressure distribution, in-vehicle noise levels, and air conditioning temperature control efficiency, generating comfort scores. Both types of scores are normalized to form a multi-dimensional evaluation vector. Finally, based on a multi-objective optimization algorithm (such as NSGA-II), the module seeks a Pareto optimal solution between conflicting dimensions of performance and comfort. The algorithm uses "maximizing performance score" and "maximizing comfort score" as objective functions, combined with parameter constraints from the digital twin model (such as the negative correlation between engine displacement and fuel consumption), iteratively optimizing configuration combinations through selection, crossover, and mutation operations. The final output is a personalized solution that balances user needs and engineering feasibility, such as a configuration combination of "2.0T high-power engine + sports suspension + noise-reducing glass." The updated process execution sequence, resource allocation details, and corresponding rest arrangement strategies are output as the scheduling execution plan.

[0050] The virtual test drive module is used to generate immersive driving experience data based on personalized vehicle configuration schemes and driving scenarios. First, the virtual test drive module synchronously loads the personalized vehicle configuration scheme obtained from the personalized configuration selection module into the digital twin virtual test drive scenario, combined with the target driving scenario. Based on the user's throttle, brake, and steering input intensity, it assesses the driving style in real time and automatically adjusts the scenario difficulty. Finally, it records the user's operation data and vehicle response data throughout the process to obtain driving experience data. Specifically, the module first synchronously loads the vehicle parameters output by the personalized configuration selection module and the target driving scenario into the digital twin test drive environment. During loading, the system corrects the vehicle dynamics model in real time based on the configuration parameters. For example, the torque output curve of a 2.0T engine configuration will be adjusted synchronously, while the damper damping coefficient of a sports suspension will be modified to ensure that the physical response of the virtual vehicle is consistent with the real configuration. Simultaneously, the driving scenario dynamically generates environmental elements based on user preferences, such as automatically activating the anti-skid system simulation in rainy weather scenarios. Secondly, the module captures the user's accelerator, brake, and steering input signals in real time through sensors, converting the input intensity into quantitative indicators: accelerator depth exceeding 70% and lasting longer than 2 seconds is considered aggressive acceleration, and a steering angle change rate exceeding 150° / s is considered a sharp turn. Based on these indicators, the system constructs a driving style feature vector, matching it in real time with preset "aggressive," "mild," and "economical" driving models, and automatically adjusting the scene difficulty: for aggressive driving, it dynamically generates continuous curves and sudden obstacles; for mild driving, it switches to smooth roads and sparse traffic flow. Finally, the module records two types of data throughout the process using timestamp alignment: user operation data (accelerator / brake pedal travel, steering angle, gear shift timing) and vehicle response data (longitudinal acceleration, lateral roll force, braking distance, steering wheel return torque). This data is encapsulated into a structured log containing spatiotemporal coordinates, providing users with a basis for reviewing driving behavior and providing empirical support for the backend data analysis module to optimize virtual exhibition content.

[0051] Finally, the backend data analysis module integrates user behavior interaction data and driving experience data to analyze user behavior patterns in the virtual auto show. Combined with user profiles, it generates user personas and optimizes the virtual show content accordingly. The backend data analysis module synchronizes and aligns user behavior interaction data and driving experience data; cleans and completes all data to generate a high-quality multi-source fusion dataset; based on the fused data, it applies unsupervised clustering and sequence pattern mining across three dimensions: showroom browsing, vehicle type and configuration, and test drive experience; it cross-maps the clustering and pattern results with user profile data, including age, occupation, and purchase budget, to generate user personas; and based on these personas, it adjusts the order of vehicles displayed in the digital twin virtual show to optimize user car purchase choices.

[0052] First, the module employs a timestamp calibration mechanism to synchronize and align user behavior interaction data with driving experience data. An event association algorithm maps scattered data streams to a unified timeline, ensuring data temporal consistency and constructing a complete behavior chain. Second, a data cleaning engine performs triple processing: filtering abnormal operation data using a sliding window, filling in missing 3D movement trajectory points using multiple interpolation, and eliminating contradictory data using domain knowledge rules, generating a standardized multi-source fusion dataset. Intelligent algorithms are applied across three analytical dimensions: DBSCAN clustering identifies high-frequency browsing paths and hotspot areas in the showroom browsing dimension; hierarchical clustering is used to segment user preference groups in the vehicle type and configuration dimension; and a Hidden Markov Model is used to mine operation sequence patterns in the test drive experience dimension. The clustering results are cross-mapped with user profiles (age, occupation, budget) to construct a multi-dimensional user profile. For example, a user group aged 25-30, working in IT, with a budget of 200,000-250,000 RMB who frequently browses new energy vehicles and prefers sporty configurations will be tagged as a "young, tech-savvy car buyer." Finally, the exhibition content was optimized based on user profiles: the display priority of 7-seater models was increased for the "family user" group, and the safety configuration demonstration area was placed at the front of the vehicle display route; for "performance enthusiasts", the dynamic display of sports car series was strengthened, and a track mode was preset in the test drive scenario to upgrade the exhibition experience.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A virtual automobile exhibition system based on digital twinning, characterized in that, The digital twin construction module, the user behavior analysis module, the personalized configuration selection module, the virtual test drive module, and the background data analysis module are included. The digital twin construction module is configured to construct a virtual automobile exhibition virtual digital twin model, in which a first-person perspective is used for touring and the model state is dynamically updated based on real-time sensing data. The user behavior analysis module is configured to collect user behavior interaction data in the virtual exhibition, predict user model requirements and configuration requirements by analyzing the user behavior interaction data, and specifically implement the following steps: First, collect user behavior interaction data, including gaze trajectory, gaze time, user sketch visual focus, interface click and parameter fine-tuning record for quantifying user configuration exploration, three-dimensional movement trajectory and dwell time for reflecting user tour path and space exploration. The above behavior interaction data is mapped into a weighted directed graph, a multi-head graph attention network is used for the weighted directed graph, and deep learning of node features is realized: each node hidden representation integrates the spatio-temporal correlation strength of its neighborhood nodes; the multi-path attention head can capture the attention mode of users on different types of interactions in parallel; after multi-layer attention aggregation, the weighted sum of all node representations is taken to obtain a vector H reflecting global behavior preferences user ; The vector reflecting the global behavior preference is combined with the description vector M of each vehicle type in the digital twin system c The user-vehicle joint feature representation Z is generated through bilinear attention fusion c The calculation method is as follows: Wherein, W Q , W K , W V represent the learnable mapping matrix of query, key and value, and SF represents a scaling factor; Online generate DB behavior comparison samples, perturb the node attributes or spatial positions in the original graph, and recalculate the corresponding joint features Z' c,db , by comparing the differences between the original Z c and the perturbed Z' c , obtain the sensitivity index S c of each vehicle model, the calculation method is: where c represents the index of the vehicle, and db represents the index of the behavior comparison sample; S c As an immediate preference reward input demand prediction policy network π θ (M c |H user ) and perform a parameter learning update once: Where θ represents the parameters of the policy network, η represents the learning rate, and θ' is the updated parameter of the policy network. After updating, output the model demand, select argmax c π θ (M c |H user ); in the same way, get the configuration demand, select argmax p π θ (M p |H user ) where M p represents each configuration vector; The personalized configuration selection module is configured to generate a personalized vehicle configuration scheme based on user model requirements and configuration requirements and in combination with digital twin model configurable parameters. The virtual test drive module is configured to generate immersive driving experience data based on the personalized vehicle configuration scheme in combination with a vehicle use scenario. The background data analysis module is configured to integrate user behavior interaction data and driving experience data, analyze user behavior patterns in the virtual automobile exhibition, generate a user portrait in combination with user personal information, and optimize virtual exhibition content based on the user portrait. 2.The digital-twin-based virtual auto show system of claim 1, wherein, The method for constructing the model by the digital twin construction module includes: Obtaining three-dimensional scanning point cloud, CAD engineering data and material spectrum parameters of a target vehicle, and simultaneously collecting laser point cloud data and building BIM model of an exhibition venue. Generating lightweight digital twin models of the vehicle and the venue by an octree grid compression algorithm, and realizing the same coordinate system of the two by a space mapping algorithm. 3.The digital-twin-based virtual auto show system of claim 2, wherein, Optimizing the digital twin model, which includes adaptively optimizing rendering parameters of the perspective according to different first-person perspectives and distances. First, the first person perspective head as the original coordinate P(x p ,y p ,z p ), the perspective direction vector is The perspective cone field angle is θ; the relative position vector of any point V(x v ,y v ,z v ) in the calculation model and the perspective is The distance d of the point V to the perspective P and the cosine value cos a of the perspective direction and the direction of the point V are calculated in the following way: and the rendering weight is determined as: Dynamic division of LOD level based on weight, setting LOD level as L, preset distance threshold interval [d0, d1,... d n ] corresponds to level [L0, L1,... L n ], wherein d0 n , L0 n ; when distance d is in interval [d i , d i+1 ], level L is calculated by linear interpolation, dynamic detail level value L(d) is calculated based on distance d, and the calculation mode is as follows: The final level obtained after combining the direction weight correction is: L final =L(d)·ω d (α)+L min ·(1-ω d (α)), wherein L final represents the final level, and L min represents the lowest detail level; Based on L final Dynamic subdivision or merging of octree network is performed when L final ≥ L max Full precision grid is enabled when L final ≤ L min Lightweight grid is enabled when L min <L final No processing is performed when L max Then the texture pixel size T is adjusted as follows: Where T max represents the highest resolution texture size. 4.The digital-twin-based virtual auto show system of claim 1, wherein, The virtual test drive module is configured to generate immersive driving experience data based on the personalized vehicle configuration scheme in combination with a vehicle use scenario, and specifically implement the following steps: First, the virtual test drive module synchronously loads the personalized vehicle configuration scheme obtained by the personalized configuration selection module into a digital twin virtual test drive scene in combination with a target driving scenario. Real-time evaluation of driving style and automatic adjustment of scene difficulty according to user input intensity of accelerator, brake and steering. Finally, record user operation data and vehicle response data to obtain driving experience data. 5.The digital-twin-based virtual auto show system of claim 1, wherein, The personalized configuration selection module is configured to generate a personalized vehicle configuration scheme based on user model requirements and configuration requirements in combination with digital twin model configurable parameters, and specifically implement the following steps: First, the personalized configuration selection module maps the user model requirements and configuration requirements obtained by the user behavior analysis module to the parameter space of the digital twin model. For each potential configuration combination configuration, a digital twin model is called to perform a quick simulation evaluation including performance simulation and comfort simulation, resulting in a performance simulation score and a comfort simulation score; Finally, a multi-objective optimization algorithm is used to obtain the optimal personalized configuration selection. 6.The digital-twin-based virtual auto show system of claim 1, wherein, The background data analysis module is configured to integrate the user's behavior interaction data and driving experience data, analyze the user's behavior pattern in the virtual automobile exhibition, generate a user portrait based on the user's personal profile, and optimize the virtual exhibition content based on the user portrait. The specific implementation steps include: The background data analysis module synchronizes and aligns the user behavior interaction data and driving experience data; All data is cleaned and missing data is completed to generate a high-quality multi-source fusion data set; Based on the fused data, unsupervised clustering and sequential pattern mining are applied in three dimensions of exhibition hall browsing, vehicle type and configuration, and test drive experience; The clustering and pattern results are cross-mapped with user profile data including user age, occupation, and car purchase budget to generate a user portrait; According to the user portrait, the order of the exhibition cars in the digital twin virtual exhibition is adjusted, and the user's car purchase selection is optimized.

Images