Dynamic driving performance self-adaptive system based on driver physiological-behavior dual-mode AI

Abstract

The application provides a dynamic driving adaptability system based on a driver physiological-behavior dual-mode AI, relates to the field of automatic driving, and comprises the following steps: synchronously collecting physiological modal data and behavior modal data of a driver; pre-processing the physiological modal data and the behavior modal data; performing feature extraction, feature fusion and feature calculation on the processed physiological modal data and the processed behavior modal data respectively, determining a driver state and a driving style; based on a driving risk score and the driving style, respectively adapting control parameters of a power system, a steering system, a braking system, a chassis and damping in real time to obtain an adaptation mode; displaying the driver state, the driving style and the current adaptation mode through a central control screen, giving a voice prompt based on the driving risk score, recording data of a correction operation of the driver, and optimizing the driving style through the data of the correction operation. The application realizes personalized adaptive adjustment of system control parameters.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI. Background Technology

[0002] With the rapid development of automotive intelligent technology, the demand for personalized vehicle power, handling, and driving experience is becoming increasingly prominent. Traditional vehicles typically have preset fixed driving performance modes, which drivers can only manually switch between a limited number of preset modes, making it impossible to dynamically adapt to the driver's real-time state and operating habits. Different drivers have significant individual differences in driving style, operating habits, and physiological state, making it difficult for fixed modes to balance driving safety, comfort, and handling feel, and thus failing to meet the increasingly diverse driving needs.

[0003] The development of driver monitoring technology has provided a technological foundation for solving the above problems. Currently, the industry has various driver state monitoring solutions based on cameras or vehicle behavior signals, which provide early warnings and interventions by identifying abnormal states such as driver fatigue and distraction. However, most existing solutions rely on a single perception modality, or only passively monitor the driver's physiological state, or only perform offline analysis of driving behavior data. These solutions have significant limitations in terms of perception dimensions and fusion depth, resulting in a high degree of susceptibility to scene interference in terms of state recognition accuracy. Furthermore, they lack a systematic design that integrates driver state perception results with vehicle control parameters in a closed-loop manner.

[0004] Chinese invention patent application number 202110019230.2 discloses a system for recognizing a driver's driving style. This patent constructs a virtual driving scenario within a real-vehicle interactive environment, collects behavioral data such as the driver's accelerator, brake, and steering wheel inputs, as well as some physiological information, while operating the vehicle. It then processes the collected data and identifies the driving style using an onboard and cloud data platform, updating the identification results to a driving style database. However, its ability to dynamically perceive and integrate the driver's real-time physiological state under actual driving conditions is insufficient, failing to achieve closed-loop linkage and self-adaptation between the driver's real-time state and the control parameters of multiple vehicle systems. Summary of the Invention

[0005] In view of this, the present invention provides a dynamic driving performance self-adaptation system based on driver physiological-behavioral dual-modal AI to solve the technical problem of lack of an effective closed-loop mechanism between the driver's real-time physiological state perception and the dynamic linkage adaptation of the control parameters of multiple vehicle systems in the prior art. It realizes the real-time identification of driver state and driving style, and on this basis, performs personalized adaptive adjustment of the control parameters of systems such as power, steering, braking, chassis and damping.

[0006] The technical solution of this invention is implemented as follows: On one hand, the present invention provides a dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI, comprising: The perception layer is configured to synchronously collect the driver's physiological modality data and behavioral modality data. The physiological modality data includes visual physiological perception data and vital sign perception data, and the behavioral modality data includes the driver's longitudinal operation behavior data, lateral operation behavior data, and working environment data. The edge computing layer is configured to preprocess physiological modality data and behavioral modality data to obtain processed physiological modality data and behavioral modality data; The AI ​​fusion decision layer is configured to extract and fuse features from the processed physiological modality data and behavioral modality data respectively, calculate the driver state confidence, determine the driver state type based on the driver state confidence and calculate the driving risk score, and perform driving style clustering modeling on the structured features of the behavioral modality data, and determine the adaptation mode according to the driver state type and driving risk score. The vehicle control execution layer is configured to receive the driving style clustering results and adaptation mode output by the AI ​​fusion decision layer, determine the personalized control parameter baseline of each control system based on the driving style clustering results, and perform real-time safety correction on the baseline by the adaptation mode, and output the final control parameter command. The human-machine interaction and feedback layer is configured to display the driver's status, driving style, and current adaptation mode on the central control screen, provide voice prompts based on the driving risk score, record data of the driver's corrective actions, and optimize the driving style based on the data of corrective actions.

[0007] Based on the above technical solutions, preferably, the perception layer further includes a multi-driver identity recognition module and an automatic driver switching module. The multi-driver identity recognition module is configured to extract the driver's facial biometrics and physiological signs based on physiological modality data and perform identity verification to determine the driver's identity. The identity verification involves matching the facial biometrics and physiological signs with the locally stored feature database for similarity. The automatic driver switching module is configured to determine a new driver if the similarity threshold is lower than a set threshold, automatically create a file and use the initial default driving style model, and complete the convergence of its exclusive driving style model within the subsequent 7-15 driving cycle. If the similarity threshold is not less than the set threshold, the driver is identified as a registered driver. Then, within 100 milliseconds, the driver's exclusive driving style model and adaptation parameters are silently loaded.

[0008] Based on the above technical solutions, preferably, the AI ​​fusion decision layer includes: The early fusion unit is configured to perform first-stage feature extraction on the processed physiological modality data and behavioral modality data respectively to obtain high-dimensional physiological feature vectors and high-dimensional behavioral feature vectors, and then perform feature fusion on the high-dimensional physiological feature vectors and high-dimensional behavioral feature vectors to obtain fused features; The late fusion unit is configured to independently infer the processed physiological modality data and behavioral modality data through physiological branches and behavioral branches respectively, to obtain the physiological branch confidence and behavioral branch confidence of various driver states, and to obtain the driver state confidence through dynamic weighted fusion; The state recognition unit is configured to determine the driver state type based on the driver state confidence level, calculate the driving risk score based on the driver state confidence level and the scenario correction coefficient, and determine the adaptation mode based on the driver state type and the driving risk score. The driving style recognition unit is configured to decouple the high-dimensional feature vector of behavior into longitudinal driving features, lateral driving features, and working condition related features. Based on the longitudinal driving features, lateral driving features, and working condition related features, clustering modeling is performed to determine the driving style and output a personalized control parameter baseline.

[0009] Based on the above technical solutions, preferably, the execution process of the early fusion unit includes: firstly, extracting spatial and temporal features from the processed physiological modality data, wherein the spatial features include facial / eye features extracted by CNN, and the temporal features include physiological signal temporal evolution features extracted by GRU; and concatenating the spatial and temporal features into a high-dimensional physiological feature vector F. phy ; The processed behavioral modality data is used to extract high-dimensional behavioral feature vectors F using LSTM and Transformer. beh It includes continuous time-series features and long-period behavior-dependent features. Continuous time-series features include throttle time-series features, steering wheel time-series features, and braking time-series features. F phy With F beh After dimensional alignment, the vectors are concatenated and processed using the CBAM attention mechanism to obtain fused features; The execution process of the late fusion unit includes: The physiological branch confidence scores P were calculated based on the processed physiological modal data using physiological branching, yielding four states: fatigue, distraction, tension, and normal. phy ; The behavioral branch confidence scores P were calculated on the processed behavioral modal data by behavioral branching to obtain the behavioral branch confidence scores P for four states: fatigue, distraction, tension, and normality. beh ; Driver state confidence is calculated using the DS evidence theory and dynamic weighted voting method. , The weights of α physiological modality data and β behavioral modality data are continuously output in real time by reinforcement learning.

[0010] Based on the above technical solutions, preferably, the execution process of the state recognition unit is as follows: The driver's state type is directly determined based on the driver's state confidence level; the driver's state includes four categories: normal focus, fatigue and drowsiness, distraction and inattentiveness, and tension and anxiety. The driving risk score is calculated based on the driver's state confidence level and the scenario correction coefficient. in Assess the driver's risk level. The confidence level for the fatigue state. For the confidence level of the distracted state, Confidence level for stress and anxiety state. This is a scene correction factor; The adaptation mode is determined based on the driver's condition type and driving risk score. The adaptation modes include standard adaptation mode, smooth protection mode and robust constraint mode.

[0011] Based on the above technical solutions, preferably, the execution process of the driving style recognition unit is as follows: Semantic decoupling and feature filtering are performed on the high-dimensional feature vectors of behavior to obtain longitudinal driving features, lateral driving features and working condition related features; Driving styles are obtained by normalizing, weighting, and clustering longitudinal driving features, lateral driving features, and working condition related features. The optimal number of clusters is determined by a combination of the elbow method and the profile coefficient method.

[0012] Based on the above technical solutions, preferred driving styles include: Aggressive type: rapid acceleration, sharp turns, short following distances, frequent lane changes, vehicle handling adapted to sensitive power, light steering, and forced movement; Smooth type: linear throttle, gentle steering, constant speed following, smooth power delivery, neutral steering, and linear braking. Conservative type: slow acceleration, cautious steering, long following distance, vehicle operation adapted to gentle power, heavy steering, and gentle braking.

[0013] Based on the above technical solutions, preferably, the steps in the vehicle control execution layer include: Based on the driving style clustering results, the baselines for personalized control parameters of the powertrain, steering system, braking system, and chassis damping are determined. Among them, the response delay benchmark and torque rise gradient benchmark corresponding to different driving styles are as follows: aggressive driving style corresponds to a response delay benchmark of 80ms and a torque rise gradient benchmark of ≥80% / s; stable driving style corresponds to a response delay benchmark of 150ms and a torque rise gradient benchmark of 50%~60% / s; and conservative driving style corresponds to a response delay benchmark of 250ms and a torque rise gradient benchmark of ≤40% / s. Based on the personalized control parameter baseline, the baseline parameters are adjusted in real time for safety according to the adaptation mode; After overlaying the personalized control parameter baseline with the real-time safety correction, the final control parameter commands for the power system, steering system, braking system and chassis damping are output, driving the vehicle control parameters to adaptively adjust.

[0014] Based on the above technical solutions, preferably, the control parameters of the power system include response delay and torque rise gradient; the control parameters of the steering system include steering assist gain, steering ratio, return rate, and steering damping coefficient; the control parameters of the braking system include brake free travel, pedal force gain, and initial deceleration slope; and the control parameters of the chassis and damping include suspension damping stiffness coefficient and roll stiffness coefficient.

[0015] More preferably, the real-time security correction specifically includes: In standard adaptation mode, baseline parameters remain unchanged; In Smoothness Protection Mode, the power response is slowed down, steering and braking are softened, and handling sensitivity is reduced from the baseline. In robust constraint mode, peak power is constrained and sensitivity limits are imposed on the baseline to suppress excessive control behavior.

[0016] The present invention has the following advantages over the prior art: (1) By constructing a five-layer closed-loop architecture consisting of a perception layer, an edge computing layer, an AI fusion decision layer, a vehicle control execution layer, and a human-machine interaction and feedback layer, end-to-end closed-loop control from driver state perception to vehicle response adjustment is achieved. Compared with the existing technology that separates driver state monitoring from vehicle control, it has stronger system coordination and real-time response capabilities. (2) An early fusion and late fusion dual-layer architecture is adopted in the AI ​​fusion decision layer. The two fusion mechanisms complement each other, which improves the robustness of driver state recognition in complex perception environment and effectively reduces the probability of misjudgment when a single mode is interfered with by noise. (3) By normalizing the driver’s longitudinal driving features, lateral driving features and working condition related features, assigning feature importance weights and clustering models through the driving style recognition unit, combined with the online incremental learning mechanism, the driving style model is realized to continuously iterate and converge with the driver’s actual operating habits, so that the system can adapt to the individual differences of different drivers and realize the silent loading of millisecond-level personalized parameters in multi-driver switching scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this invention provides a dynamic driving performance adaptive system based on a driver's physiological-behavioral dual-modal AI, comprising: The perception layer is configured to synchronously collect the driver's physiological modality data and behavioral modality data. The physiological modality data includes visual physiological perception data and vital sign perception data, and the behavioral modality data includes the driver's longitudinal operation behavior data, lateral operation behavior data, and working environment data. The edge computing layer is configured to preprocess physiological modality data and behavioral modality data to obtain processed physiological modality data and behavioral modality data; The AI ​​fusion decision layer is configured to extract and fuse features from the processed physiological modality data and behavioral modality data respectively, calculate the driver state confidence, determine the driver state type based on the driver state confidence and calculate the driving risk score, and perform driving style clustering modeling on the structured features of the behavioral modality data, and determine the adaptation mode according to the driver state type and driving risk score. The vehicle control execution layer is configured to receive the driving style clustering results and adaptation mode output by the AI ​​fusion decision layer, determine the personalized control parameter baseline of each control system based on the driving style clustering results, and perform real-time safety correction on the baseline by the adaptation mode, and output the final control parameter command. The human-machine interaction and feedback layer is configured to display the driver's status, driving style, and current adaptation mode on the central control screen, provide voice prompts based on the driving risk score, record data of the driver's corrective actions, and optimize the driving style based on the data of corrective actions.

[0021] Understandably, visual physiological perception data is collected through infrared and RGB binocular cameras deployed on the A-pillars or rearview mirrors. The collected data includes eye features, facial features, and focus probability. Eye features include eyelid opening and closing angle (EAR), blink frequency and duration, gaze direction, and eye movement speed, used to identify fatigue, distraction, and focus states. Facial features include the number of yawns, mouth position, head posture (nodding / shaking / looking down angle), and micro-expressions (tension / relaxation / anger).

[0022] This invention constructs a five-layer closed-loop architecture consisting of a perception layer, an edge computing layer, an AI fusion decision layer, a vehicle control execution layer, and a human-machine interaction and feedback layer. This architecture enables end-to-end closed-loop control from driver state perception to vehicle response adjustment. Compared with the existing technology that separates driver state monitoring from vehicle control, this invention has stronger system coordination and real-time response capabilities. It enables real-time identification of driver state and driving style, and on this basis, personalized adaptive adjustments are made to system control parameters such as power, steering, braking, chassis, and damping.

[0023] In one embodiment of the present invention, facial features are detected using YOLOv8 based on a convolutional neural network (CNN) architecture combined with a CBAM attention mechanism. The facial and eye spatial features extracted by its convolutional layers are used in the early fusion unit of the AI ​​fusion decision layer, with the eye features having an F1 score of no less than 91%. Under backlight and low-light conditions, the infrared night vision channel ensures robust recognition. The spatial features are the facial / eye spatial features output by the YOLOv8 convolutional layer of the perception layer, which are accessed through a unified interface to participate in subsequent feature stitching in the early fusion unit.

[0024] In one embodiment of the present invention, the focus probability is calculated by weighted normalization of multiple factors including eye indicators, head posture, and facial micro-expressions, and outputs a continuous probability value of 0 to 1: in , , Normalized scores were assigned to the eyes, head, and face, respectively. .

[0025] In one embodiment of the present invention, vital sign perception data are collected through the fusion of contact and non-contact sensors. For example, heart rate (HR), respiratory rate (RR), and heart rate variability (HRV) are collected non-contactly by millimeter-wave radar, which mainly achieves non-sensory monitoring by capturing micro-movements of the chest cavity; electrocardiogram (ECG) and electrodermal conductance (EDA) are collected in contact by flexible electrodes (hand grip area) on the steering wheel, which reflect the driver's emotional arousal level (tension / anxiety / calm); and the distribution of sitting pressure and the frequency of body micro-movements are collected by seat pressure sensors to help identify fatigue and distraction.

[0026] In one embodiment of the present invention, longitudinal operational behavior data includes accelerator pedal opening, opening change rate, pedaling rhythm and response delay preference; brake pedal travel, pedal force, deceleration, initial braking speed, and following distance control logic. Lateral operational behavior data includes steering wheel angle, angular velocity, torque input, return-to-center speed, lane keeping accuracy, lane change timing and magnitude; lane departure amount, correction frequency, and steering intervention sensitivity requirements. Operating environment data includes vehicle speed, acceleration, jerk impact, gear / mode switching frequency, following distance, overtaking frequency, and cornering speed control.

[0027] In one embodiment of the present invention, the perception layer further includes a multi-driver identity recognition module and an automatic driver switching module. The multi-driver identity recognition module is configured to extract the driver's facial biometrics and physiological signs based on physiological modality data and perform identity verification to determine the driver's identity. The identity verification involves matching the facial biometrics and physiological signs with the locally stored feature database for similarity. The automatic driver switching module is configured to determine a new driver if the similarity threshold is lower than a set threshold, automatically create a file and use the initial default driving style model, and complete the convergence of its exclusive driving style model within the subsequent 7-15 driving cycle. If the similarity threshold is not less than the set threshold, the driver is identified as a registered driver. Then, within 100 milliseconds, the driver's exclusive driving style model and adaptation parameters are silently loaded.

[0028] Specifically, facial biometrics are based on infrared and RGB binocular camera images in visual physiological perception data to extract facial key point feature vectors and facial bone topological features, generating a unique driver facial feature code, which has robust recognition under backlight, mask, and light makeup conditions.

[0029] Physiological signs are integrated from vital sign perception data, including heart rate baseline, HRV static characteristics, hand-held ECG waveform baseline, and sitting pressure distribution baseline. The uniqueness of each person's physiological baseline is used as a secondary identity verification factor.

[0030] Identity verification specifically includes: The system performs similarity matching between facial biometrics and physiological characteristics and a locally stored feature database. This database is stored locally in an encrypted manner. Each driver is associated with a complete set of data, including facial biometrics, physiological baseline, historical driving style model, and personalized adaptation parameters. After the driver sits down and grips the steering wheel, the system simultaneously collects facial images, an initial 3-second physiological baseline, and posture pressure features, and performs similarity matching with the local feature database.

[0031] In one embodiment of the present invention, when the similarity is lower than a set threshold, the driver is identified as a new driver. The system automatically creates a file and uses an initial default driving style model. The system then converges the driver's exclusive driving style model within 7 to 15 driving cycle periods. A single driving cycle period is defined as continuous driving for 30 minutes or a driving distance of not less than 10 km. When the similarity is not lower than the set threshold, the driver is identified as a registered driver. The system silently loads the driver's exclusive driving style model and adaptation parameters within 100 milliseconds. These parameters include the adaptation parameters for the power system, steering system, braking system, chassis and damping, dual-modal fusion weights, and risk assessment thresholds. The entire process is seamless and requires no manual settings.

[0032] In one embodiment of the present invention, the edge computing layer is deployed based on an onboard GPU / MCU hardware platform and performs low-latency preprocessing on the physiological modality data and behavioral modality data collected by the perception layer. The preprocessing includes time synchronization, noise reduction, and normalization, specifically: Apply 10ms-level timestamp alignment to all sensor data to ensure consistency between physiological modality data and behavioral modality data in the time dimension and avoid modality misalignment; Wavelet transform is used to filter out baseline drift and high-frequency noise from vital sign sensing data. Kalman filtering is used to suppress sensor jitter from longitudinal operation behavior data, lateral operation behavior data, and working environment data. Outliers that are significantly outside the normal range are removed. By standardizing the various indicators in physiological modality data and behavioral modality data to the same feature space, a unified numerical basis is provided for feature extraction and fusion calculation in the AI ​​fusion decision layer.

[0033] In one embodiment of the present invention, the AI ​​fusion decision layer includes: The early fusion unit is configured to perform first-stage feature extraction on the processed physiological modality data and behavioral modality data respectively, to obtain physiological high-dimensional feature vectors and behavioral high-dimensional feature vectors, and then perform feature fusion on the physiological high-dimensional feature vectors and behavioral high-dimensional feature vectors to obtain fused features; wherein, the physiological high-dimensional feature vectors and behavioral high-dimensional feature vectors are compressed representations without physical semantic decoupling, and the physiological high-dimensional feature vectors are not directly used for downstream calculations of driving risk scoring and driving style clustering; the behavioral high-dimensional feature vectors will be further semantically decoupled by the driving style recognition unit and used as input for style clustering modeling; The late fusion unit is configured to independently infer the processed physiological modality data and behavioral modality data through physiological branches and behavioral branches respectively, to obtain the physiological branch confidence and behavioral branch confidence of various driver states, and to obtain the driver state confidence through dynamic weighted fusion; The state recognition unit is configured to determine the driver state type based on the driver state confidence level, calculate the driving risk score based on the driver state confidence level and the scenario correction coefficient, and determine the adaptation mode based on the driver state type and the driving risk score. The driving style recognition unit is configured to decouple the high-dimensional feature vector of behavior into longitudinal driving features, lateral driving features, and working condition related features. Based on the longitudinal driving features, lateral driving features, and working condition related features, clustering modeling is performed to determine the driving style and output a personalized control parameter baseline.

[0034] Among them, longitudinal driving features, lateral driving features, and working condition association features are all high-dimensional feature vectors. They are subsets after semantic decoupling and physical classification, and are used as the core input for style clustering. They form a two-level processing relationship of "implicit compression → physical decoupling" with the high-dimensional behavioral feature vectors in the early fusion stage.

[0035] This invention adopts a two-layer architecture of early fusion and late fusion in the AI ​​fusion decision layer. The two fusion mechanisms complement each other, which improves the robustness of driver state recognition in complex perception environments and effectively reduces the probability of misjudgment when a single modality is interfered with by noise.

[0036] In one embodiment of the present invention, the execution process of the early fusion unit includes: firstly, extracting spatial features and temporal features from the processed physiological modality data, wherein the spatial features include facial / eye features extracted by CNN, and the temporal features include physiological signal temporal evolution features extracted by GRU; and concatenating the spatial features and temporal features into a high-dimensional physiological feature vector F. phy ; The processed behavioral modality data is used to extract high-dimensional behavioral feature vectors F using LSTM and Transformer. behIt includes continuous time-series features and long-period behavior-dependent features. Continuous time-series features include throttle time-series features, steering wheel time-series features, and braking time-series features. F phy With F beh After dimensional alignment, the vectors are concatenated and processed using the CBAM attention mechanism to obtain fused features; The execution process of the late fusion unit includes: The physiological branch confidence scores P were calculated based on the processed physiological modal data using physiological branching, yielding four states: fatigue, distraction, tension, and normal. phy ; The behavioral branch confidence scores P were calculated on the processed behavioral modal data by behavioral branching to obtain the behavioral branch confidence scores P for four states: fatigue, distraction, tension, and normality. beh ; Driver state confidence is calculated using the DS evidence theory and dynamic weighted voting method. , , Physiological modality data weights, The weights of the behavioral modality data are continuously output in real time by reinforcement learning. They gradually change with the driver's state confidence and the degree of environmental interference. In scenarios such as strong light, occlusion, and distortion of physiological signals, the weights of the physiological branch are automatically weakened and the weights of the behavioral branch are strengthened to avoid misjudgment of a single modality. The final driver state confidence target is no less than 0.9.

[0037] Understandably, the early fusion unit and the late fusion unit form a two-layer architecture, working together as a backup: under normal operating conditions, the state recognition unit outputs the driver state confidence level through late fusion. As the primary criterion, when the quality assessment index of the early fusion feature (such as the CBAM attention score) exceeds a set threshold, the auxiliary classification result output by the fusion feature will be compared with... Weighted verification is performed to correct misjudgments of a single branch in noisy scenarios. When the confidence level of the late-fusing physiological branch is abnormally low due to sensor malfunction, the auxiliary judgment result of the early-fusing fusion feature is used as a supplementary basis. Together, they ensure that the overall recognition accuracy is not less than 95% and the false alarm rate is controlled within 3%.

[0038] In one embodiment of the present invention, the execution process of the state recognition unit is as follows: The driver's state type is directly determined based on the driver's state confidence level; the driver's state includes four categories: normal focus, fatigue and drowsiness, distraction and inattentiveness, and tension and anxiety. The driving risk score is calculated based on the driver's state confidence level and the scenario correction coefficient. in Assess the driver's risk level. The confidence level for the fatigue state. For the confidence level of the distracted state, Confidence level for stress and anxiety state. This is a scene correction factor; The adaptation mode is determined based on the driver's condition type and driving risk score. The adaptation modes include standard adaptation mode, smooth protection mode and robust constraint mode.

[0039] Specifically, the standard adaptation mode corresponds to normal focused state, maintaining the original factory basic parameters or personalized baseline parameters; Smooth Protection Mode: This mode is for those who are fatigued, drowsy, or distracted. Both of these states involve decreased attention and reduced handling precision, and the corresponding strategies are the same: slowing down the power response, softening the steering and braking, and reducing handling sensitivity. Robust constraint mode: corresponds to a state of tension and anxiety; this state is characterized by excessive manipulation and frequent corrections. The appropriate strategy is to constrain the peak power and limit the upper limit of sensitivity to suppress excessive manipulation behavior.

[0040] Among them, the driving risk score determines the parameter adjustment range within each of the above adaptation modes: slight intervention when the risk is low, and strong intervention when the risk is high, so as to achieve orthogonal separation control of state direction and intensity.

[0041] In this embodiment of the invention, the input to the state recognition unit comes from two parallel computing paths: Physiological high-dimensional feature vectors and some abnormal behavioral features serve as the core inputs to the driving risk scoring model, used to quantify the instantaneous degree of danger to the driver. Longitudinal driving features, lateral driving features, and operating condition correlation features serve as the core inputs to the style clustering model, representing the driver's long-term steady-state preferences.

[0042] The two paths compute in parallel without interfering with each other, together forming the two-dimensional output of the AI ​​fusion decision-making layer.

[0043] In one embodiment of the present invention, the scene correction coefficient The value ranges from 0.9 to 1.1, with higher coefficients for high-speed, nighttime, and curved scenes, and lower coefficients for straight daytime scenes.

[0044] In one embodiment of the present invention, the execution process of the driving style recognition unit is as follows: Semantic decoupling and feature filtering are performed on the high-dimensional feature vectors of behavior to obtain longitudinal driving features, lateral driving features, and operating condition-related features. Among them, longitudinal driving features include throttle opening change rate, frequency of rapid acceleration, average braking force, frequency of emergency braking, steady-state following distance, and average Jerk values ​​for acceleration and deceleration; lateral driving features include average steering wheel angular velocity, frequency of large-angle steering, lane change frequency, lane departure correction magnitude, and cornering rate; operating condition-related features include deviation of normal cruising speed, overtaking frequency, and frequency of driving mode switching. Driving styles are obtained by normalizing, weighting, and clustering longitudinal driving features, lateral driving features, and working condition-related features. The optimal number of clusters is determined by a combination of the elbow method and the silhouette coefficient method. Driving styles include: Aggressive type: rapid acceleration, sharp turns, short following distances, frequent lane changes, vehicle handling adapted to sensitive power, light steering, and forced movement; Smooth type: linear throttle, gentle steering, constant speed following, smooth power delivery, neutral steering, and linear braking. Conservative type: slow acceleration, cautious steering, long following distance, vehicle operation adapted to gentle power, heavy steering, and gentle braking.

[0045] Understandably, the original driving condition data for each dimension is first normalized to a score of 0 to 100, with positive indicators directly mapped and negative indicators mapped inversely. The longitudinal driving features, lateral driving features, and driving condition correlation features of the input clusters are weighted according to their importance, with higher weights given to throttle opening change rate, average steering wheel angular velocity, and following distance, while secondary features such as driving mode switching frequency are given lower weights. The elbow method and silhouette coefficient method are used together to determine the optimal number of clusters. By pre-clustering massive historical driving samples, the sum of squared errors (SSE) and silhouette coefficients under different K values ​​are calculated. When K=3, the silhouette coefficient is the highest, the intra-class distance is the smallest, and the inter-class distance is the largest. Therefore, driving styles are fixedly divided into three categories: aggressive, stable, and conservative.

[0046] LSTM combined with Transformer performs temporal modeling of multidimensional features, outputting a multidimensional feature vector as clustering input. Simultaneously, the multidimensional features are weighted and mapped to a continuous driving style score from 0 to 100. Higher scores indicate more aggressive driving; this score is only used as a visualization reference and does not directly participate in clustering calculations. The core input for clustering is a normalized and weighted multidimensional structured behavioral feature vector (including longitudinal driving features, lateral driving features, and condition-related features). After clustering using K-means or Gaussian mixture models, three discrete style labels—aggressive, stable, and conservative—are output. Furthermore, the aforementioned multidimensional features are weighted and linearly mapped to a driving style score from 0 to 100. Higher scores indicate a more aggressive driving tendency. The driving style score ranges are divided as follows: 70-100 points correspond to aggressive, 35-69 points to stable, and 0-34 points to conservative.

[0047] After each trip, the system silently initiates incremental learning in the background to update the clustering model of the driving style recognition unit. The minimum data threshold for triggering an update is that there are no fewer than 150 sets of temporal behavioral feature frames of valid driving conditions in a single trip. If the data is insufficient, it is temporarily stored and accumulated for the next trip for merging and updating. The update method adopts elastic incremental learning, where new samples are integrated into the old model with a low learning rate, preserving the historical main style and only slightly adjusting the parameters to achieve slow iteration and stable convergence. The system uses the 3σ criterion and Mahalanobis distance to detect outliers such as emergency avoidance and sudden acceleration / deceleration, and removes them from the training set. The local vehicle terminal retains snapshots of the driving style model from the last 5 trips. If a large abrupt shift in model style is detected, it automatically rolls back to the previous stable model and supports manual triggering of model reset to restore the factory initial driving style baseline.

[0048] This invention normalizes, assigns weights to, and clusters the longitudinal driving features, lateral driving features, and operating condition-related features of the driver through a driving style recognition unit. Combined with an online incremental learning mechanism, it achieves continuous iterative convergence of the driving style model according to the driver's actual operating habits, enabling the system to adapt to the individual differences of different drivers and achieve silent loading of personalized parameters at the millisecond level in multi-driver switching scenarios.

[0049] In one embodiment of the present invention, the steps in the vehicle control execution layer include: Based on the driving style clustering results, the baselines for personalized control parameters of the powertrain, steering system, braking system, and chassis damping are determined. Among them, the response delay benchmark and torque rise gradient benchmark corresponding to different driving styles are as follows: aggressive driving style corresponds to a response delay benchmark of 80ms and a torque rise gradient benchmark of ≥80% / s; stable driving style corresponds to a response delay benchmark of 150ms and a torque rise gradient benchmark of 50%~60% / s; and conservative driving style corresponds to a response delay benchmark of 250ms and a torque rise gradient benchmark of ≤40% / s. Based on the personalized control parameter baseline, the baseline parameters are adjusted in real time for safety according to the adaptation mode. In the standard adaptation mode, the baseline parameters remain unchanged. In the smooth protection mode, the power response is slowed down, the steering and braking are softened, and the handling sensitivity is reduced on the basis of the baseline. In the robust constraint mode, the power peak is constrained and the upper limit of sensitivity is limited on the basis of the baseline to suppress excessive handling behavior. The adjustment range is determined by the driving risk score: slight intervention for low risk (below 70 points), moderate intervention for medium risk (70 to 84 points), and strong intervention for high risk (not lower than 85 points). After overlaying the personalized control parameter baseline with the real-time safety correction, the final control parameter commands for the power system, steering system, braking system and chassis damping are output, driving the vehicle control parameters to adaptively adjust.

[0050] Understandably, the adaptation mode is obtained by first determining the basic control parameters based on the driving risk score, and then performing linear interpolation correction on the basic control parameters according to the driving style. All control parameters change gradually within a preset range, and parameter switching is completed within 100ms. The control impact Jerk does not exceed 2m / s³, ensuring a smooth adaptation process without any jerking sensation, and strictly adhering to the bottom line of safety constraints, without exceeding the limits of active safety systems such as ESP / ABS.

[0051] Specifically, the correction magnitude of the personalized control parameter baseline by the adaptation mode is determined by the driving risk score, according to the following rules: When the driving risk score is below 70, the standard adaptation mode is executed, and the control parameters of each subsystem maintain the personalized baseline value corresponding to the driving style clustering result. When the driving risk score is between 70 and 84, a slight intervention of the smoothness protection mode or robust constraint mode is executed, for example, the response delay is increased by 10% to 20% based on the style baseline, and the torque rise gradient is reduced by 10% to 20% based on the style baseline, in order to smooth the power response. When the driving risk score is not lower than 85, a strong intervention of the corresponding adaptation mode is executed, and the power response and steering sensitivity are both converged towards the safety constraint boundary based on the style baseline, with each parameter taking a strongly constrained value within the preset safety boundary. The above corrections start from the personalized baseline corresponding to the style clustering, and after determining the correction direction and magnitude coefficient through the adaptation mode, linear interpolation is performed to finally obtain the actual control parameter output of each subsystem.

[0052] This invention employs a dual-layer orthogonal control mechanism of "personalized baseline + real-time safety correction" at the vehicle control execution layer: the personalized baseline of each subsystem control parameter is determined by the driving style clustering results, the correction direction of the adaptation mode is determined by the driver state type, and the correction magnitude is determined by the driving risk score. The three are orthogonally separated and work together to achieve coordinated and gradual adaptation of the power system, steering system, braking system and chassis damping control parameters within the range of personalized constraints and safety constraints, avoiding the sudden change in handling caused by step switching of parameters.

[0053] In one embodiment of the present invention, the control parameters of the power system include response delay and torque rise gradient; the control parameters of the steering system include steering assist gain, steering ratio, return rate and steering damping coefficient; the control parameters of the braking system include brake free travel, pedal force gain and initial deceleration slope; and the control parameters of the chassis and damping include suspension damping stiffness coefficient and roll stiffness coefficient.

[0054] In one embodiment of the present invention, the specific process by which the vehicle control execution layer determines the final control parameters of each subsystem based on the personalized control parameter baseline and the adaptation mode is as follows: starting from the baseline value corresponding to the driving style, linear interpolation correction is performed after the adaptation mode determines the correction direction and magnitude coefficient: Control parameters for the power system: In standard adaptation mode, throttle opening is linearly mapped to torque, maintaining the style baseline value; in smooth protection mode, response delay is increased and torque rise gradient is reduced on the basis of style baseline to avoid power surge; in robust constraint mode, torque peak is constrained and power response upper limit is limited on the basis of style baseline to suppress excessive handling.

[0055] Control parameters for the steering system: Starting with the driving style baseline: the aggressive baseline features light power assist, a large steering ratio, a fast return-to-center rate, and a low steering damping coefficient; the moderate baseline features neutral power assist, a standard steering ratio, and a standard return-to-center rate; and the conservative baseline features heavy power assist, a small steering ratio, a slow return-to-center rate, and a high steering damping coefficient. Above these baselines, both the smoothness protection mode and the robust constraint mode adjust towards increasing power assist, decreasing the steering ratio, and slowing down the return-to-center rate, with the adjustment magnitude determined by the driving risk score.

[0056] Control parameters for the braking system: Starting with the driving style baseline: the aggressive baseline features short brake travel, high pedal force gain, and a large initial deceleration slope, resulting in sensitive initial braking; the smooth baseline features linear brake travel and neutral pedal force gain, providing comfortable braking; and the conservative baseline features long brake travel, low pedal force gain, and a small initial deceleration slope, resulting in gentle braking. Above these baselines, the smoothness protection mode adjusts towards increasing brake travel and reducing pedal force gain to moderate braking response and reduce occupant impact; the robust constraint mode adjusts towards moderately tightening the braking response to prevent overly aggressive braking.

[0057] The control parameters for the chassis and damping are obtained by linear interpolation using a two-dimensional lookup table based on operating conditions and driving style. For driving under fatigue or high-speed conditions, a soft suspension with a low damping coefficient and low roll stiffness coefficient is used to improve ride comfort and reduce the interference of road vibration on the driver. For driving with an aggressive style or cornering conditions, a hard suspension with a high damping coefficient and high roll stiffness coefficient is used to enhance vehicle stability and suppress roll.

[0058] In one embodiment of the present invention, the central control screen or HUD displays the current driver status (normal focus / fatigue / distraction / anxiety), current driving style (aggressive / smooth / conservative), and current adaptation mode in real time, ensuring that the adaptation process is transparent and perceptible. Based on a driving risk score, tiered voice prompts are provided: when the score is not lower than 85 points, a gentle voice prompt is triggered, such as "Fatigue detected, switched to smooth adaptation mode, recommended to rest when appropriate"; when the score is not lower than 90 points, a HUD highlight warning is superimposed on the voice prompt, and a mild power limitation is triggered.

[0059] The human-machine interaction and feedback layer records the driver's corrective actions for the current adaptation mode, including manual switching of driving modes and manual adjustment of steering assist gain. This corrective action data is used as negative feedback signals to input into the clustering model of the driving style recognition unit. This allows the model to prioritize matching the driver's explicit operational preferences during subsequent adaptations, achieving continuous self-learning and precise iterative optimization of driving style. The system also retains a manual mode entry, allowing the driver to exit self-adaptation with a single click and switch to manual control mode, recording this preference for future reference.

[0060] Local incremental learning and cloud-based federated learning employ an independent triggering and asynchronous collaboration mechanism: After each trip, local incremental learning immediately updates the local driving style model, and the updated local model takes effect instantly for use in subsequent trips; cloud-based federated learning does not affect the normal use of the local model. The global model it distributes serves as the foundation for the local model and is integrated and taken effect in the next local incremental update, rather than overwriting the currently converged local model. After every three complete driving trips, the on-board unit uploads anonymized feature data to the cloud. The cloud performs weekly federated learning (FedAvg architecture) aggregation and iteration: each vehicle only uploads model gradients or feature statistics, without uploading raw physiological modality data and behavioral modality data. The cloud aggregates the gradients of each vehicle to update the global base model, with data anonymized throughout to protect driver privacy. The global model is triggered when any of the following conditions are met: the global model recognition accuracy in the cloud improves by at least 2%, the generalization ability of newly added typical working environment scenarios is significantly optimized, or a quarterly routine iteration update is performed. By combining local incremental learning with cloud-based federated learning, the driving style model can adapt to the long-term driving habits of a single driver and continuously improve its generalization ability to different individual driver differences and diverse working conditions.

[0061] This invention records the driver's corrective operation data on the adaptation mode through human-computer interaction and feedback layer, uses it as a negative feedback signal to optimize the model of the driving style recognition unit, and combines it with cloud federated learning mechanism to realize the collaborative update of local personalized model and global generalized model, forming a complete self-learning closed loop of collection-decision-control-feedback.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI, characterized in that, include: The perception layer is configured to synchronously collect the driver's physiological modality data and behavioral modality data. The physiological modality data includes visual physiological perception data and vital sign perception data, and the behavioral modality data includes the driver's longitudinal operation behavior data, lateral operation behavior data, and working environment data. The edge computing layer is configured to preprocess physiological modality data and behavioral modality data to obtain processed physiological modality data and behavioral modality data; The AI ​​fusion decision layer is configured to extract and fuse features from the processed physiological modality data and behavioral modality data respectively, calculate the driver state confidence, determine the driver state type based on the driver state confidence and calculate the driving risk score, and perform driving style clustering modeling on the structured features of the behavioral modality data, and determine the adaptation mode according to the driver state type and driving risk score. The vehicle control execution layer is configured to receive the driving style clustering results and adaptation mode output by the AI ​​fusion decision layer, determine the personalized control parameter baseline of each control system based on the driving style clustering results, and perform real-time safety correction on the baseline by the adaptation mode, and output the final control parameter command. The human-machine interaction and feedback layer is configured to display the driver's status, driving style, and current adaptation mode on the central control screen, provide voice prompts based on the driving risk score, record data of the driver's corrective actions, and optimize the driving style based on the data of corrective actions.

2. The dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI according to claim 1, characterized in that, The perception layer also includes a multi-driver identity recognition module and an automatic driver switching module. The multi-driver identity recognition module is configured to extract the driver's facial biometrics and physiological signs based on physiological modality data and perform identity verification to determine the driver's identity. The identity verification involves matching the facial biometrics and physiological signs against a locally stored feature library for similarity. The automatic driver switching module is configured to determine a new driver if the similarity threshold is lower than a set threshold, automatically create a file and use the initial default driving style model, and complete the convergence of its exclusive driving style model within the subsequent 7-15 driving cycle. If the similarity threshold is not less than the set threshold, the driver is identified as a registered driver. Then, within 100 milliseconds, the driver's exclusive driving style model and adaptation parameters are silently loaded.

3. The dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI according to claim 1, characterized in that, The AI ​​fusion decision-making layer includes: The early fusion unit is configured to perform first-stage feature extraction on the processed physiological modality data and behavioral modality data respectively to obtain high-dimensional physiological feature vectors and high-dimensional behavioral feature vectors, and then perform feature fusion on the high-dimensional physiological feature vectors and high-dimensional behavioral feature vectors to obtain fused features; The late fusion unit is configured to independently infer the processed physiological modality data and behavioral modality data through physiological branches and behavioral branches respectively, to obtain the physiological branch confidence and behavioral branch confidence of various driver states, and to obtain the driver state confidence through dynamic weighted fusion; The state recognition unit is configured to determine the driver state type based on the driver state confidence level, calculate the driving risk score based on the driver state confidence level and the scenario correction coefficient, and determine the adaptation mode based on the driver state type and the driving risk score. The driving style recognition unit is configured to decouple the high-dimensional feature vector of behavior into longitudinal driving features, lateral driving features, and working condition related features. Based on the longitudinal driving features, lateral driving features, and working condition related features, clustering modeling is performed to determine the driving style and output a personalized control parameter baseline.

4. The dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI according to claim 3, characterized in that, The execution process of the early fusion unit includes: firstly, extracting spatial and temporal features from the processed physiological modality data, wherein the spatial features include facial / eye features extracted by CNN, and the temporal features include physiological signal temporal evolution features extracted by GRU; and concatenating the spatial and temporal features into a high-dimensional physiological feature vector F. phy ; The processed behavioral modality data is used to extract high-dimensional behavioral feature vectors F using LSTM and Transformer. beh It includes continuous time-series features and long-cycle behavior-dependent features. Continuous time-series features include throttle time-series features, steering wheel time-series features, and braking time-series features. F phy With F beh After dimensional alignment, the vectors are concatenated and processed using the CBAM attention mechanism to obtain fused features; The execution process of the late fusion unit includes: The physiological branch confidence scores P were calculated based on the processed physiological modal data using physiological branching, yielding four states: fatigue, distraction, tension, and normal. phy ; The behavioral branch confidence scores P for four states—fatigue, distraction, tension, and normal—were calculated by analyzing the processed behavioral modal data through behavioral branching. beh ; Driver state confidence is calculated using the DS evidence theory and dynamic weighted voting method. , The weights of α physiological modality data and β behavioral modality data are continuously output in real time by reinforcement learning.

5. The dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI according to claim 3, characterized in that, The execution process of the state recognition unit is as follows: The driver's status type is determined directly based on the driver's status confidence level; The driver's state is categorized into four types: normal focus, fatigue and drowsiness, distraction and inattentiveness, and tension and anxiety. The driving risk score is calculated based on the driver's state confidence level and the scenario correction coefficient. in Assess the driver's risk level. The confidence level for the fatigue state. For the confidence level of the distracted state, Confidence level for stress and anxiety state. This is a scene correction factor; The adaptation mode is determined based on the driver's condition type and driving risk score. The adaptation modes include standard adaptation mode, smooth protection mode and robust constraint mode.

6. The dynamic driving performance adaptive system based on driver physiological-behavioral dual-modal AI according to claim 5, characterized in that, The execution process of the driving style recognition unit is as follows: Semantic decoupling and feature filtering are performed on the high-dimensional feature vectors of behavior to obtain longitudinal driving features, lateral driving features and working condition related features; Driving styles are obtained by normalizing, weighting, and clustering longitudinal driving features, lateral driving features, and working condition related features. The optimal number of clusters is determined by a combination of the elbow method and the profile coefficient method.

7. A dynamic driving performance adaptive system based on driver physiological-behavioral dual-modal AI according to claim 6, characterized in that, The driving styles include: Aggressive type: rapid acceleration, sharp turns, short following distances, frequent lane changes, vehicle handling adapted to sensitive power, light steering, and forced movement; Smooth type: linear throttle, gentle steering, constant speed following, smooth power delivery, neutral steering, and linear braking. Conservative type: slow acceleration, cautious steering, long following distance, vehicle operation adapted to gentle power, heavy steering, and gentle braking.

8. A dynamic driving performance self-adaptive system based on driver physiological-behavioral dual-modal AI according to claim 7, characterized in that, The steps in the vehicle control execution layer include: Based on the driving style clustering results, the baselines for personalized control parameters of the powertrain, steering system, braking system, and chassis damping are determined. Among them, the response delay benchmark and torque rise gradient benchmark corresponding to different driving styles are as follows: aggressive driving style corresponds to a response delay benchmark of 80ms and a torque rise gradient benchmark of ≥80% / s; stable driving style corresponds to a response delay benchmark of 150ms and a torque rise gradient benchmark of 50%~60% / s; and conservative driving style corresponds to a response delay benchmark of 250ms and a torque rise gradient benchmark of ≤40% / s. Based on the personalized control parameter baseline, the baseline parameters are adjusted in real time for safety according to the adaptation mode; After overlaying the personalized control parameter baseline with the real-time safety correction, the final control parameter commands for the power system, steering system, braking system and chassis damping are output, driving the vehicle control parameters to adaptively adjust.

9. A dynamic driving performance adaptive system based on driver physiological-behavioral dual-modal AI according to claim 8, characterized in that, The control parameters of the power system include response delay and torque rise gradient; the control parameters of the steering system include steering assist gain, steering ratio, return rate, and steering damping coefficient; the control parameters of the braking system include brake free travel, pedal force gain, and initial deceleration slope; and the control parameters of the chassis and damping include suspension damping stiffness coefficient and roll stiffness coefficient.

10. A dynamic driving performance adaptive system based on driver physiological-behavioral dual-modal AI according to claim 8, characterized in that, The real-time security correction specifically includes: In standard adaptation mode, baseline parameters remain unchanged; In Smoothness Protection Mode, the power response is slowed down, steering and braking are softened, and handling sensitivity is reduced from the baseline. In robust constraint mode, peak power is constrained and sensitivity limits are imposed on the baseline to suppress excessive control behavior.

Images