A vehicle individualization system and method

Abstract

The application belongs to the technical field of vehicle control system, and is a vehicle individualization adaptation system and method, which comprises: a perception layer for collecting facial images and human posture images, three-dimensional point cloud data, pressure distribution data, vehicle operation behavior data and vehicle motion state data; a decision layer for generating static physiological characteristic data, dynamic behavior characteristic data and driving scene data based on the data collected by the perception layer, generating or updating a user's individualized digital file based on the static physiological characteristic data and the dynamic behavior characteristic data, and generating an individualized parameter configuration set for multiple vehicle subsystems based on the individualized digital file, the dynamic behavior characteristic data and the driving scene data; and an execution layer for sending the individualized parameter configuration set to the subsystem controllers of the vehicle, and driving physical actuators and feeding back to the decision layer. The application solves the problem that the existing vehicle safety and comfort system cannot adapt to the individual differences and real-time states of different drivers.

Description

Technical Field

[0001] This invention belongs to the field of vehicle control system technology, and in particular relates to a vehicle personalization adaptation system and method. Background Technology

[0002] With the development of the automotive industry, vehicle design has shifted from meeting basic functions to pursuing higher levels of safety, comfort, and personalized experiences.

[0003] However, traditional automotive design paradigms are generally based on standard human models, such as the 50th percentile adult male, for ergonomics and safety system development. This design reveals significant limitations when faced with real users who have different body types, sitting postures, and driving habits: In terms of safety, traditional passive safety systems, such as airbags and seat belts, have fixed trigger parameters. For petite occupants or those sitting forward, the deployment force of a standard airbag may be too great, leading to additional injury risks. Conversely, for taller occupants, their critical areas may be outside the optimal protection range of the airbag. Seat belts may also dig into the neck or abdomen due to differences in body size, increasing the risk of injury. While existing technologies have attempted to fine-tune airbags using seat position signals, the adjustment dimension is singular and cannot accurately match the user's real-time three-dimensional posture.

[0004] In terms of comfort and ergonomics, although existing vehicles are generally equipped with electric seat memory functions, users need to manually adjust and store them. The process is cumbersome and cannot guarantee that the adjusted position meets the principle of optimal safety. The memory function is usually isolated and cannot be linked to the adjustment of rearview mirrors, seat belt anchors, air conditioning and other systems. Furthermore, it is not associated with the vehicle's core safety parameters. Some systems can perform simple occupant classification, but cannot achieve refined personalized recognition and adaptation. Summary of the Invention

[0005] This invention provides a vehicle personalization adaptation system and method, which solves the problem that existing vehicle safety and comfort systems cannot intelligently and collaboratively adapt to the individual differences and real-time status of different drivers.

[0006] This invention provides a basic solution: a vehicle personalization adaptation system, comprising: The perception layer is used to collect user facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion state data. The decision layer is used to generate static physiological feature data, dynamic behavioral feature data, and driving scenario data based on data collected by the perception layer. Based on the static physiological feature data and dynamic behavioral feature data, it generates or updates the user's personalized digital profile. It is also used to generate personalized parameter configuration sets for multiple subsystems of the vehicle based on the personalized digital profile and the dynamic behavioral feature data and driving scenario data collected in real time. The execution layer is used to send personalized parameter configuration sets to the vehicle's subsystem controllers. Each subsystem controller drives the physical actuators to complete the actions and feeds back the execution status to the decision layer.

[0007] Preferably, the perception layer includes a visual perception module, a body pressure distribution perception module, and a vehicle dynamics and operation perception module; The visual perception module is used to collect user facial images, human posture images, and 3D point cloud data; Body pressure distribution sensing module, used to collect pressure distribution data of the driver's body; The vehicle dynamics and operation perception module is used to collect user vehicle operation behavior data and vehicle motion status data.

[0008] More preferably, the decision layer includes a user management module, a feature extraction network, a cross-modal fusion module, and a collaborative optimization decision module. The user management module is used to analyze the user's facial images, create new user digital profiles or load user digital profiles based on the analysis results; it is also used to optimize and update personalized digital profiles based on static physiological feature data, dynamic behavioral feature data and user feedback on vehicle adjustments. The feature extraction network includes a static physiological feature extraction subnetwork, a dynamic behavioral feature extraction subnetwork, and a driving scene recognition subnetwork. The static physiological feature extraction subnetwork adopts a hybrid deep learning architecture of convolutional neural network and three-dimensional graph neural network to generate static physiological feature data based on three-dimensional point cloud data and pressure distribution data. The static physiological feature data includes user size parameters and body shape parameters. The dynamic behavior feature extraction subnetwork adopts a hybrid architecture that combines long short-term memory network and temporal convolutional network to generate dynamic behavior feature data based on the collected facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data and vehicle motion state data. The cross-modal fusion module, based on the Transformer multi-head self-attention network, fuses static physiological feature data, dynamic behavioral feature data, and scene feature data to output a comprehensive user state vector. The collaborative optimization decision-making module receives the user's comprehensive state vector output by the cross-modal fusion module, combines it with the current user's personalized digital profile and instantaneous vehicle motion state, and generates a set of personalized parameter configurations for collaboratively controlling multiple subsystems of the vehicle through multi-objective optimization calculations.

[0009] More preferably, the personalized digital profile management strategy in the user management module is as follows: a. Based on the user's facial image, a face recognition algorithm is used to extract the feature vector of the current facial image; b. Calculate the similarity between the feature vector of the facial image and the facial feature vector of registered users in the personalized digital profile. If the highest similarity score is higher than the preset recognition threshold, the user is identified as a registered user, and the file loading process is triggered. If the highest similarity is lower than the preset rejection threshold, or the database is empty, it is determined to be a new user, triggering the new user registration process and creating a new user digital profile; If the highest similarity score is greater than the rejection threshold but less than the recognition threshold, a second comparison is performed.

[0010] More preferably, in the static physiological feature extraction sub-network, the strategy for generating static physiological feature data is as follows: 1) Preprocess the 3D point cloud data and pressure distribution data, and align and map the pressure distribution data with the 3D point cloud in space. 2) Input the preprocessed 3D point cloud into a convolutional neural network for preliminary recognition and output a feature map of local information; construct a graph structure based on the feature map according to the preset human skeleton topology, where each suspected joint point becomes a graph node and the bone connections between nodes become edges; input the graph structure into a 3D graph neural network and output a set of 3D key point coordinates through multiple iterations. 3) Using a set of 3D keypoint coordinates, linear size parameters are obtained through spatial geometric calculations; the aligned pressure distribution data is integrated to calculate the weight value; contour parameters are calculated through a classification and regression sub-network; at the same time, the weight, size parameters, and contour parameters are input into a pre-trained body type classifier to output body type classification labels. 4) Normalize and reduce the dimensionality of all the calculated size parameters and body shape parameters, and merge them into a static physiological feature vector.

[0011] More preferably, the dynamic behavioral feature data includes the user's real-time sitting posture, physiological state data, operation data, and habit data; The collaborative optimization decision-making module includes a driving style analysis unit, a multi-objective strategy network, and a parameter mapping and instruction generation unit. The driving style analysis unit extracts habitual data accumulated over a long period of time from the sub-network based on dynamic behavioral features, calibrates it in conjunction with personalized digital profiles, and generates driving style labels through cluster analysis or classification models. The multi-objective policy network takes the user's comprehensive state vector, the vehicle's instantaneous state, and the driving style condition vector as inputs, and outputs a multi-dimensional action vector. The parameter mapping and instruction generation unit will convert multi-dimensional action vectors into control instructions or parameter values ​​to form a personalized parameter configuration set according to the communication protocols and interface requirements of the vehicle's various subsystem controllers.

[0012] More preferably, the reward function formula for training the multi-objective policy network is as follows:

[0013] in, As a safety reward, For comfort rewards, For performance rewards, As a stability reward, , , , These are the weighting coefficients.

[0014] More preferably, the subsystem controller includes a human-machine interaction system controller, a constraint system controller, a chassis and power system controller, and an active safety system controller.

[0015] Another basic solution provided by this invention: a vehicle personalization adaptation method, applied to the above-mentioned vehicle personalization adaptation system, comprising the following steps: S1: Collect user's facial and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion status data; S2: Generate and update the user's personalized digital profile based on the collected data, and generate a personalized parameter configuration set for multiple vehicle subsystems based on the personalized digital profile and real-time collected dynamic behavioral feature data and driving scenario data. S3: Used to send personalized parameter configuration sets to the vehicle's subsystem controllers. Each subsystem controller drives the physical actuators to complete the actions and feeds back the execution status to the decision-making layer.

[0016] The principles and advantages of this invention are as follows: 1. By collecting multi-dimensional data such as user facial images, human posture images, and 3D point cloud data, it breaks through the limitations of existing single data collection and achieves a comprehensive and accurate depiction of users' static physiological characteristics and dynamic behavioral characteristics, providing reliable data support for personalized adaptation; it generates or updates users' personalized digital profiles by combining collected data, and generates a set of personalized parameter configurations for multiple subsystems by combining real-time dynamic behavioral characteristic data and driving scenario data, solving the problem of the lack of refined personalized recognition in existing solutions, and realizing the transformation from standard human adaptation to individual precise adaptation.

[0017] 2. Based on a personalized parameter configuration set, the system coordinates and adapts systems such as seats, rearview mirrors, seat belt anchors, airbags, and air conditioning, avoiding the shortcomings of isolated configurations of each subsystem. It deeply integrates comfort configurations with core safety parameters, balancing driving safety and comfort. By identifying dangerous states such as fatigued driving and distracted driving through multi-dimensional data, the system adjusts the sensitivity of the active safety system in advance, achieving proactive safety that prevents accidents before they occur. At the same time, the system can pre-adjust the restraint system based on the identified occupant position before a potential collision, making the best preparation for collision protection. The parameter configuration is sent to the subsystem controller through the execution layer and the execution status is fed back, forming a closed-loop control of control and feedback. This ensures that personalized configurations are accurately implemented and can be dynamically updated according to changes in user behavior and scene switching, improving the real-time performance and flexibility of adaptation.

[0018] 3. Personalized parameters can be automatically generated and updated without manual user intervention, simplifying the user operation process and adapting to users of different body types and driving habits, significantly improving the personalized experience and universality of vehicle use. Attached Figure Description

[0019] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0020] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: (See details) Figure 1 A vehicle personalization adaptation system, comprising: The perception layer is used to collect user facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion state data. The perception layer is the data input end of the system, used to collect and initially process raw sensor signals to generate standardized feature data for subsequent personalized decision-making.

[0021] The perception layer includes a visual perception module, a body pressure distribution perception module, and a vehicle dynamics and operation perception module. The visual perception module is used to collect the user's facial images, human posture images, and 3D point cloud data. The facial images and human posture images are collected by one or more high-resolution RGB cameras installed near the rearview mirror in the vehicle. The 3D point cloud data (upper body, especially the head and chest) is collected by a depth sensor, which is a time-of-flight (ToF) sensor or a structured light sensor. The body pressure distribution sensing module is used to collect pressure distribution data of the driver's body; the pressure distribution data is collected through a high-density pressure sensor array integrated inside the seat cushion and backrest of the driver's seat. The vehicle dynamics and operation perception module is used to collect user vehicle operation behavior data and vehicle motion state data. The vehicle operation behavior data includes steering wheel operation data and pedal operation data. The steering wheel operation data is collected by a steering wheel angle sensor and a torque sensor. The pedal operation data is collected by accelerator and brake pedal travel sensors. The vehicle motion state data is collected by a vehicle inertial measurement unit (IMU).

[0022] The decision layer is used to generate static physiological feature data, dynamic behavioral feature data, and driving scenario data based on data collected by the perception layer. Based on the static physiological feature data and dynamic behavioral feature data, it generates or updates the user's personalized digital profile. It is also used to generate personalized parameter configuration sets for multiple subsystems of the vehicle based on the personalized digital profile and the dynamic behavioral feature data and driving scenario data collected in real time. The decision-making layer includes a user management module, a feature extraction network, a cross-modal fusion module, and a collaborative optimization decision-making module. The user management module is used to analyze the user's facial images, create new user digital profiles or load user digital profiles based on the analysis results; it is also used to optimize and update personalized digital profiles based on static physiological feature data, dynamic behavioral feature data and user feedback on vehicle adjustments. The personalized digital profile, which associates a complete profile with each user ID, is the core data model created and maintained by each registered user, and mainly stores the following information: Identity and physiological baseline parameters include facial feature vectors, user size parameters, body shape parameters, and body pressure distribution baseline models; Human-computer interaction and comfort preference parameters include seat position and posture preference sets (preference settings or ranges for seat fore-aft position, height, seat cushion tilt angle, backrest tilt angle, lumbar support protrusion, and side wing support width for one or more different driving modes), steering wheel and pedal preferences (preference positions for steering wheel height and fore-aft extension, sensitivity curve indicators for accelerator pedal response, etc.), vision system preferences (personalized angle presets for interior / exterior rearview mirrors, projection height, brightness, and display content layout template indicators for the HUD), and comfort system preferences (default temperature, airflow, and airflow mode for air conditioning; trigger temperature thresholds and intensity preferences for seat ventilation / heating). Personalized safety restraint and protection parameters include safe position boundaries, restraint system calibration parameters, and headrest position parameters. The safe position boundaries include the upper and lower limits of the safe positions of the seat, steering wheel, and headrest, calculated based on the user's physiological parameters to ensure optimal protection. The restraint system calibration parameters include airbag strategy identifiers and seat belt control parameters. The airbag strategy identifiers include a correction coefficient for airbag deployment timing (ignition time) corresponding to the user's body type and sitting posture, deployment angle offset, and a multi-stage inflator burst intensity level mapping table. The seat belt control parameters include a personalized collision deceleration threshold for pretensioner triggering and a personalized force-limiting characteristic curve identifier for the force limiter unit. The headrest position parameters are associated with the seat back angle, showing a personalized mapping relationship between the headrest's fore-and-aft position and angle.

[0023] The chassis, powertrain, and active safety personalization parameters include chassis handling preferences, powertrain preferences, and active safety system sensitivity parameters. Chassis handling preferences include the electric power steering system's force level or custom curve label; the soft / hard mode preferences for the variable damping suspension (such as "Comfort," "Standard," and "Sport"); the powertrain preferences refer to the transmission's shift logic mode preferences; the active safety system sensitivity parameters include the triggering timing (distance / time threshold) and sensitivity level of Forward Collision Warning (FCW) and Lane Departure Warning (LDW), the personalized deceleration curve label for Automatic Emergency Braking (AEB), and the steering intervention force and correction speed parameters for Lane Keeping Assist (LKA), etc.

[0024] The personalized digital profile update strategy is as follows: a. Based on the user's facial image, a face recognition algorithm is used to extract the feature vector of the current facial image; b. Calculate the similarity between the feature vector of the facial image and the facial feature vector of registered users in the personalized digital profile. If the highest similarity score is higher than the preset recognition threshold, the user is determined to be a registered user, and the file loading process is triggered; in this embodiment, the recognition threshold is 0.95. If the highest similarity is lower than the preset rejection threshold, or the template library is empty, the user is identified as a new user, triggering the new user registration process and creating a new user digital profile. In this embodiment, the rejection threshold is 0.3. If the highest similarity score is greater than the rejection threshold but less than the recognition threshold, a second comparison is performed; the steps of the second comparison are as follows: The system guides users to maintain a natural driving posture through prompts (such as on-screen text or voice), and simultaneously collects the user's current real-time pressure distribution data through the body pressure distribution sensing module. The current pressure distribution data is then compared with the body pressure distribution baseline model stored in the profile of the top N (N=3 in this embodiment) candidate users with the highest similarity in the initial identification to obtain the body pressure distribution similarity. The overall similarity is calculated using the following formula:

[0025] in, The similarity of facial feature vectors. For similarity of body pressure distribution, , These are the weighting coefficients, and + =1, in this embodiment =0.7, =0.3; If the overall similarity is higher than the preset fusion confirmation threshold (in this embodiment, the fusion confirmation threshold is 0.85), then it is finally determined to be a registered user, triggering the profile loading process; If the overall similarity is lower than the fusion confirmation threshold, a lightweight proactive inquiry is made through the human-machine interface; for example, the most likely candidate user nickname (such as "Zhang San") is clearly displayed on the in-vehicle display screen, and a prompt is made: "Are you Zhang San? Please nod or say 'yes' to confirm." If a positive response is detected from the user, it is considered user confirmation, and the corresponding user profile is loaded. A positive response from the user can be detected by the in-vehicle camera as a nod or by the microphone recognizing a positive voice keyword (such as "yes"). If a negative response is detected from a user, the user is ultimately identified as a new user, and the new user registration process is initiated. A negative response from a user can be a clear statement of "no," no response, or a timeout (e.g., 5 seconds). Specifically, for new users, the system matches an initial configuration template that conforms to the principle of optimal safety based on their static physiological characteristics, which serves as the starting point of the file, and sends the initial personalized parameter configuration to the execution layer for control using this initial configuration template; for non-new users, the file loading process is triggered, and the file storage information is used as the initial personalized parameter configuration to send to the execution layer. In subsequent use, the system achieves dynamic optimization and updates of files through the following closed-loop logic: 1) Conflict monitoring and recording: Continuously monitor the difference between the real-time personalized parameter configuration set generated by the following collaborative optimization decision-making module and the parameter information loaded from the archive; at the same time, record the scene classification label, vehicle motion status and user physiological status when this difference occurs; 2) Conflict Analysis and Filtering: The user management module performs offline batch analysis on the recorded conflict data. If a parameter difference is highly correlated with specific non-routine driving scenarios, extreme vehicle dynamics, or abnormal user states, it is determined to be a temporary adaptive behavior and the profile will not be updated. If a parameter difference occurs repeatedly and continuously in normal driving scenarios, for example, if the user manually straightens the seat back at the same angle in 10 consecutive commuting trips, and the correlation with the scenario and dynamics is weak, it is determined that the user's habits may have changed and it is marked as an update candidate to be confirmed. 3) For candidate tags for update, the system may adopt one of the following strategies: a. Incrementally update the original parameters in the file directly using a smoothing algorithm; b. Lightly ask the user through the human-computer interface at appropriate times, and update or reject the request based on the user's clear feedback.

[0026] The feature extraction network includes a static physiological feature extraction subnetwork, a dynamic behavioral feature extraction subnetwork, and a driving scene recognition subnetwork. The static physiological feature extraction subnetwork adopts a hybrid deep learning architecture of convolutional neural network and 3D graph neural network. It generates static physiological feature data based on 3D point cloud data and pressure distribution data. The static physiological feature data includes user size parameters and body shape parameters. Specifically, the size parameters include height, sitting height, thigh length, calf length, upper arm length, forearm length, and shoulder width; the body shape parameters include overall weight and trunk contour classification. The strategy for generating static physiological characteristic data is as follows 1) Preprocess the 3D point cloud data and pressure distribution data, specifically by denoising, filtering and unifying the coordinate system of the 3D point cloud data to form a clean point cloud to be processed; and by aligning and mapping the pressure distribution data with the 3D point cloud in space to ensure data correlation. 2) Accurate human body modeling based on hybrid neural networks: Specifically, the preprocessed 3D point cloud is input into a convolutional neural network (CNN) to initially identify the shoulder bulge area, the cylindrical surface of the torso, the tubular structure of the limbs, etc., and output a set of feature maps rich in local information. Based on the feature maps output by the CNN, the network automatically or according to a preset human skeleton topology constructs a graph structure, in which each suspected joint point (such as shoulder, elbow, hip, knee) becomes a graph node, and the bone connections between nodes become edges. The graph structure is input into a 3D graph neural network, and through multiple iterations, it corrects the local misjudgments that may occur in the CNN stage, and outputs a set of high-precision, ergonomically conforming 3D key point coordinates. 3) Using a set of 3D keypoint coordinates, linear dimensional parameters are obtained through spatial geometric calculations; the aligned pressure distribution data is integrated to calculate the weight value; combining the body frame defined by the 3D keypoints with the surface shape information of the point cloud, the contour parameters of key sections such as torso thickness, chest, waist, and hips are calculated through a classification and regression sub-network; at the same time, the weight, size ratio, and contour features are input into a pre-trained body type classifier, such as a classifier based on support vector machine (SVM) or fully connected layers, to output body type classification labels. 4) All the calculated size parameters and body shape parameters are normalized and dimensionality reduced through an encoding layer, and then fused and encoded into a low-dimensional, dense static physiological feature vector.

[0027] The dynamic behavior feature extraction subnetwork adopts a hybrid architecture combining Long Short-Term Memory (LSTM) network and Temporal Convolutional Network (TCN) to generate dynamic behavior feature data based on the collected facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data and vehicle motion state data; specifically, the dynamic behavior feature data includes the user's real-time sitting posture, physiological state data, operation data and habit data. Among them, the user's real-time sitting posture is obtained by inputting the three-dimensional point cloud data into the temporal convolutional network (TCN) and outputting the three-dimensional coordinates and posture angle sequence of the user's head and chest, which is the underlying data of "real-time sitting posture". Physiological state data is generated from facial images using a cascaded architecture of CNN+LSTM. First, CNN extracts features from key regions such as the eyes and mouth frame by frame. Then, LSTM analyzes the temporal evolution of these features and learns the rhythm and abnormal patterns of physiological signals, such as calculating the number of blinks per unit time (PERCLOS), the duration of the gaze leaving the road, and the micro-expression pattern of yawning. It outputs quantified fatigue level and attention concentration index as physiological state data. The operation data is based on steering wheel operation data collected by steering wheel angle and torque sensors and pedal operation data collected by accelerator and brake pedal travel sensors. A combination of lightweight TCN and real-time feature engineering is used. Specifically, TCN directly processes the raw signals and extracts their waveform features (such as rate of change and jitter). At the same time, it combines preset rules (such as calculating steering angular velocity and pedal opening rate of change) to generate operation feature vectors. The resulting millisecond-level time sequence of steering wheel angle, torque value, and pedal opening is used as operation data. Habitual data, based on long-term recorded steering wheel operation data, pedal operation data, and vehicle motion state data collected by the vehicle's inertial measurement unit (IMU), employs a hybrid offline / online batch learning mode. Periodically (e.g., daily or every 100 kilometers), sliding window statistical analysis, unsupervised clustering (e.g., K-means), and supervised classification are performed on the accumulated time-series data to derive habitual data characterizing the user's driving style, such as the average abruptness of steering wheel operations, the smoothness index of pedal operations, and typical following distance preferences.

[0028] The driving scene recognition sub-network, based on vehicle motion state data (vehicle inertial measurement unit (IMU) signals, such as acceleration and angular velocity), employs a lightweight classification network, such as MobileNet or other efficient CNNs. It classifies and identifies the current driving environment by analyzing the vehicle's dynamic characteristics (such as acceleration / deceleration patterns and lateral sway), and outputs scene feature data. This scene feature data includes a scene classification label (such as "highway", "urban congestion", "mountain road curves", "parking") and its corresponding scene feature vector.

[0029] The cross-modal fusion module, based on the Transformer multi-head self-attention network, fuses static physiological feature data, dynamic behavioral feature data, and scene feature data to output a comprehensive user state vector.

[0030] Specifically, the integration strategy includes 1) Map the vectors corresponding to static physiological feature data, dynamic behavioral feature data, and scene feature data to the same high-dimensional semantic space; 2) A scaled dot product attention mechanism is adopted. For each head, the query, key, and value matrix is ​​calculated. The attention weight is dynamically determined by the similarity between the query and the key, as shown in the formula:

[0031] Where Q, K, and V are generated from the projected feature vectors. It is the dimension of the key vector, and the weight matrix output by the softmax function is the dynamic fusion weight across modalities; 3) The outputs of multiple attention heads are concatenated and projected again, and then processed by a feedforward neural network, residual connections, and layer normalization to finally output a fixed-dimensional, deeply fused user comprehensive state vector.

[0032] The collaborative optimization decision-making module receives the user's comprehensive state vector output by the cross-modal fusion module, combines it with the current user's personalized digital profile and instantaneous vehicle motion state, and generates a set of personalized parameter configurations for collaboratively controlling multiple subsystems of the vehicle through multi-objective optimization calculations.

[0033] The collaborative optimization decision-making module includes a driving style analysis unit, a multi-objective policy network, and a parameter mapping and instruction generation unit. The driving style analysis unit extracts habitual data accumulated over a long period of time from the dynamic behavioral feature extraction sub-network, calibrates it in conjunction with personalized digital profiles, and generates driving style labels such as "comfort" and "sport" and corresponding driving style condition vectors through clustering analysis or classification models such as support vector machines and neural network classifiers, and updates them regularly.

[0034] The multi-objective policy network adopts a deep reinforcement learning (DRL) framework, taking the user's comprehensive state vector, the vehicle's instantaneous state, and the driving style condition vector as inputs, and outputs a multi-dimensional action vector. The multi-dimensional action vector contains all parameter changes of the system to be adjusted, such as seat displacement ΔS, airbag angle ΔA, warning threshold ΔT, etc. The reward function formula for training the multi-objective policy network is as follows:

[0035] in, As a safety reward, For comfort rewards, For performance rewards, As a stability reward, , , , These are the weighting coefficients; Among them, the safety reward is based on the degree of conformity between the user's real-time sitting posture and the preset "safe position boundary" in the personalized digital profile; the comfort reward is based on the uniformity of the user's body pressure distribution, whether the body joint angles are within the ergonomic comfort range, and the vehicle's longitudinal / lateral acceleration (stability) as the main indicators; the performance reward is strongly correlated with the user's driving style; and the stability reward is calculated based on the vehicle's body posture stability at the edge of the handling limit or in emergency conditions.

[0036] The parameter mapping and instruction generation unit will convert multi-dimensional action vectors into specific, executable control instructions or parameter values ​​(such as CAN bus messages, specific motor target positions, and airbag igniter level parameters) according to the communication protocols and interface requirements of the vehicle's various subsystem controllers to form a personalized parameter configuration set.

[0037] The execution layer is used to send personalized parameter configuration sets to the vehicle's subsystem controllers via the vehicle control bus. Each subsystem controller drives physical actuators to complete actions and feeds back the execution status (such as success, failure, and current position) to the decision layer, forming a closed-loop control. The subsystem controllers include the human-machine interaction system controller, the constraint system controller, the chassis and powertrain system controller, and the active safety system controller.

[0038] Specifically, the human-machine interaction system controller controls the seat system to adjust the seat's fore-and-aft position, height, seat cushion tilt angle, backrest tilt angle, lumbar support protrusion, and side wing support width; the steering wheel system adjusts the steering wheel's height and fore-and-aft extension; the vision system adjusts the angle of the interior / exterior rearview mirrors to a preset personalized field of view and adjusts the HUD's projection height, brightness, and display content layout; the comfort system sets the air conditioner's temperature, airflow, and airflow direction, or activates the seat's ventilation, heating, and massage functions. The restraint system controller controls the airbag deployment timing (ignition time), deployment angle, and burst intensity level of the multi-stage inflator; the pretension force trigger threshold of the seat belt pretensioner and the force limiting characteristic curve parameters of the force limiter control unit; and the fore-and-aft position and angle of the headrest. The chassis and powertrain controller adjusts the force feedback of the electric power steering system of the chassis domain controller; the stiffness mode of the variable damping suspension; and the powertrain controller adjusts the responsiveness of the accelerator pedal and the shift logic of the transmission. The active safety system controller controls the triggering timing and sensitivity of the warning system; the intervention intensity (deceleration curve) of the emergency braking system; and the steering intervention force of the lane keeping assist.

[0039] Example 2 A vehicle personalization adaptation method, applied to the aforementioned vehicle personalization adaptation system, includes the following steps: S1: Collect user's facial and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion status data; S2: Generate and update the user's personalized digital profile based on the collected data, and generate a personalized parameter configuration set for multiple vehicle subsystems based on the personalized digital profile and real-time collected dynamic behavioral feature data and driving scenario data. S3: Used to send personalized parameter configuration sets to the vehicle's subsystem controllers. Each subsystem controller drives the physical actuators to complete the actions and feeds back the execution status to the decision-making layer.

[0040] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A vehicle personalization adaptation system, characterized in that, include: The perception layer is used to collect user facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion state data. The decision layer is used to generate static physiological feature data, dynamic behavioral feature data, and driving scenario data based on data collected by the perception layer. Based on the static physiological feature data and dynamic behavioral feature data, it generates or updates the user's personalized digital profile. It is also used to generate personalized parameter configuration sets for multiple subsystems of the vehicle based on the personalized digital profile and the dynamic behavioral feature data and driving scenario data collected in real time. The decision-making layer includes a user management module, a feature extraction network, a cross-modal fusion module, and a collaborative optimization decision-making module. The user management module is used to analyze the user's facial images, create new user digital profiles or load user digital profiles based on the analysis results; it is also used to optimize and update personalized digital profiles based on static physiological feature data, dynamic behavioral feature data and user feedback on vehicle adjustments. The feature extraction network includes a static physiological feature extraction subnetwork, a dynamic behavioral feature extraction subnetwork, and a driving scene recognition subnetwork. The static physiological feature extraction subnetwork adopts a hybrid deep learning architecture of convolutional neural network and three-dimensional graph neural network to generate static physiological feature data based on three-dimensional point cloud data and pressure distribution data. The static physiological feature data includes user size parameters and body shape parameters. The dynamic behavior feature extraction subnetwork adopts a hybrid architecture that combines long short-term memory network and temporal convolutional network to generate dynamic behavior feature data based on the collected facial images and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data and vehicle motion state data. The driving scene recognition sub-network classifies and identifies the current driving environment based on vehicle motion state data and outputs scene feature data. The cross-modal fusion module, based on the Transformer multi-head self-attention network, fuses static physiological feature data, dynamic behavioral feature data, and scene feature data to output a comprehensive user state vector. The collaborative optimization decision module is used to receive the user's comprehensive state vector output by the cross-modal fusion module, and combine it with the current user's personalized digital profile and instantaneous vehicle motion state to generate a set of personalized parameter configurations for collaborative control of multiple subsystems of the vehicle through multi-objective optimization calculation. The collaborative optimization decision-making module includes a driving style analysis unit, a multi-objective policy network, and a parameter mapping and instruction generation unit. The driving style analysis unit extracts habitual data accumulated over a long period of time from the sub-network based on dynamic behavioral features, calibrates it in conjunction with personalized digital profiles, and generates driving style labels through cluster analysis or classification models. The multi-objective policy network takes the user's comprehensive state vector, the vehicle's instantaneous state, and the driving style condition vector as inputs, and outputs a multi-dimensional action vector. The parameter mapping and instruction generation unit will convert multi-dimensional action vectors into control instructions or parameter values ​​to form a personalized parameter configuration set according to the communication protocols and interface requirements of the vehicle's various subsystem controllers. The execution layer is used to send personalized parameter configuration sets to the vehicle's subsystem controllers. Each subsystem controller drives the physical actuators to complete the actions and feeds back the execution status to the decision layer.

2. The vehicle personalization adaptation system according to claim 1, characterized in that: The perception layer includes a visual perception module, a body pressure distribution perception module, and a vehicle dynamics and operation perception module. The visual perception module is used to collect user facial images, human posture images, and 3D point cloud data; Body pressure distribution sensing module, used to collect pressure distribution data of the driver's body; The vehicle dynamics and operation perception module is used to collect user vehicle operation behavior data and vehicle motion status data.

3. The vehicle personalization adaptation system according to claim 2, characterized in that: The personalized digital profile management strategy in the user management module is as follows: a. Based on the user's facial image, a face recognition algorithm is used to extract the feature vector of the current facial image; b. Calculate the similarity between the feature vector of the facial image and the facial feature vector of registered users in the personalized digital profile. If the highest similarity score is higher than the preset recognition threshold, the user is identified as a registered user, and the file loading process is triggered. If the highest similarity is lower than the preset rejection threshold, or the database is empty, it is determined to be a new user, triggering the new user registration process and creating a new user digital profile; If the highest similarity score is greater than the rejection threshold but less than the recognition threshold, a second comparison is performed.

4. The vehicle personalization adaptation system according to claim 3, characterized in that: In the static physiological feature extraction subnetwork, the strategy for generating static physiological feature data is as follows: 1) Preprocess the 3D point cloud data and pressure distribution data, and align and map the pressure distribution data with the 3D point cloud in space. 2) Input the preprocessed 3D point cloud into a convolutional neural network for preliminary recognition and output a feature map of local information; construct a graph structure based on the feature map according to the preset human skeleton topology, where each suspected joint point becomes a graph node and the bone connections between nodes become edges; input the graph structure into a 3D graph neural network and output a set of 3D key point coordinates through multiple iterations. 3) Using a set of 3D keypoint coordinates, linear size parameters are obtained through spatial geometric calculations; the aligned pressure distribution data is integrated to calculate the weight value; contour parameters are calculated through a classification and regression sub-network; at the same time, the weight, size parameters, and contour parameters are input into a pre-trained body type classifier to output body type classification labels. 4) Normalize and reduce the dimensionality of all the calculated size parameters and body shape parameters, and merge them into a static physiological feature vector.

5. The vehicle personalization adaptation system according to claim 4, characterized in that: The dynamic behavioral feature data includes the user's real-time sitting posture, physiological state data, operation data, and habit data.

6. The vehicle personalization adaptation system according to claim 5, characterized in that: The reward function formula for training the multi-objective policy network is as follows: in, As a safety reward, For comfort rewards, For performance rewards, As a stability reward, , , , These are the weighting coefficients.

7. The vehicle personalization adaptation system according to claim 6, characterized in that: The subsystem controllers include a human-machine interaction system controller, a constraint system controller, a chassis and power system controller, and an active safety system controller.

8. A vehicle personalization adaptation method, applied to the vehicle personalization adaptation system as described in any one of claims 1-7, characterized in that the steps include... include: S1: Collect user's facial and human posture images, 3D point cloud data, pressure distribution data, vehicle operation behavior data, and vehicle motion status data; S2: Generate and update the user's personalized digital profile based on the collected data, and generate a personalized parameter configuration set for multiple vehicle subsystems based on the personalized digital profile and real-time collected dynamic behavioral feature data and driving scenario data. S3: Used to send personalized parameter configuration sets to the vehicle's subsystem controllers. Each subsystem controller drives the physical actuators to complete the actions and feeds back the execution status to the decision-making layer.

Images