Vehicle-mounted-cloud road time health multi-dimensional analysis and risk assessment system

Abstract

The invention discloses a vehicle-mounted-cloud road time health multi-dimensional analysis and risk assessment system, and belongs to a vehicle-mounted analysis and assessment technology. Capacitance, pressure and radar signals are fused, the defect of a single sensor can be overcome, a lightweight self-adaptive filtering model generated based on hardware sensing meta-neural architecture search is adapted to a vehicle-mounted edge computing power constraint and interference feature real-time sensing mechanism, so that the model can adjust a filtering strategy according to vehicle vibration and illumination changes, and the vehicle-mounted edge computing power constraint and interference feature real-time sensing mechanism is optimized. The dynamic scene self-adaptive capability is enhanced, the delay of the generation of the health state snapshot of the edge end can be effectively reduced through model quantification and hardware optimization, and the occupancy rate of vehicle-mounted computing resources can be effectively reduced by unloading a complex model training task to the cloud and only executing lightweight reasoning by the edge end. The parameter updating strategy is dynamically adjusted based on the communication quality, the data transmission amount can be effectively reduced in a weak communication environment, and the model updating failure caused by communication fluctuation is avoided.

Description

Vehicle-to-Cloud Time Health Multidimensional Analysis and Risk Assessment System Technical Field

[0001] This invention relates to vehicle-mounted analysis and evaluation technology, specifically to a vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system. Background Technology

[0002] The multidimensional health analysis and risk assessment of vehicle-to-cloud road time is a comprehensive research direction that integrates technologies from multiple fields such as intelligent transportation, vehicle-to-everything (V2X), cloud computing and health science. Its core objective is to collect time-series data during vehicle operation (i.e., vehicle-mounted) and cloud-based collaborative processing (i.e., cloud-road) to monitor and analyze the physiological, psychological and behavioral health status of drivers, passengers and even road users in real time and on this basis, and to conduct health risk assessment and early warning.

[0003] When implementing existing technical solutions, it is difficult to run complex AI models due to the limitations of vehicle computing resources in terms of size, power consumption, and cost. Health risk warnings require low-latency responses, and the communication of moving vehicles is easily affected by external environmental factors, which also affects data processing and communication quality. There is a contradiction between limited computing resources, environmental impact, and high real-time requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system to solve the technical problems of the contradiction between limited computing resources, environmental impact and high real-time requirements in existing solutions.

[0005] The objective of this invention can be achieved through the following technical solution: a vehicle-cloud-based multi-dimensional health analysis and risk assessment system, comprising: a multi-source non-contact physiological signal acquisition and processing module: acquiring multi-source non-contact physiological signals from steering wheel capacitive sensing, seat pressure distribution, and millimeter-wave radar; using a meta-neural architecture to search and generate a lightweight adaptive filtering model adapted to vehicle edge hardware; eliminating vehicle vibration and light interference in real time; and outputting a low-dimensional lightweight physiological feature vector; a cloud-edge collaborative construction and dynamic update module: based on the lightweight physiological feature vectors obtained through processing, constructing a hierarchical federated transfer learning cloud-edge collaborative framework; the edge terminal uses the deployed lightweight inference model to output a real-time health status snapshot; the cloud terminal aggregates multi-vehicle health feature knowledge through federated transfer learning; and employs a communication quality-aware parameter sparse update mechanism to address fluctuations in the vehicle communication environment; and a vehicle-mounted digital twin health risk assessment and intervention module: based on the processed edge terminal health status snapshot and cloud federated knowledge, constructing a vehicle-mounted digital twin health risk simulator; dynamically mapping health features and risk levels through reinforcement learning; and triggering graded early warning and intervention strategies.

[0006] Furthermore, during the synchronous acquisition of multi-source non-contact physiological signals, a capacitive sensor captures subtle fluctuations in heart rate, a pressure sensor senses changes in body pressure caused by respiration, and a millimeter-wave radar detects chest movement data. The time synchronization of the three signals is achieved through the vehicle's CAN bus.

[0007] Furthermore, focusing on the adaptability of in-vehicle edge hardware, a three-dimensional optimization objective is constructed, involving the following expression: Where T is the model inference delay; P is the model parameter size; and A is the accuracy of the filtered physiological features. All are weighting coefficients. A lightweight adaptive filtering model adapted to the in-vehicle environment is generated through iterative search using a near-end strategy optimization algorithm.

[0008] Furthermore, when performing real-time filtering and outputting low-dimensional features, the multi-source physiological signal matrix and interference features are input into the lightweight adaptive filtering model in parallel. The lightweight adaptive filtering model strengthens effective physiological features and suppresses interference signals through an attention mechanism. Through the global average pooling operation of the output layer of the lightweight adaptive filtering model, the high-dimensional filtering result is compressed into a 128-dimensional low-dimensional physiological feature vector. The low-dimensional physiological feature vector includes multi-dimensional fusion features of heart rate variability, respiratory depth, and body pressure distribution.

[0009] Furthermore, when generating a health status snapshot, the edge model calculates three core health indicators in real time based on the output low-dimensional physiological feature vector: heart rate variability time-domain index, respiratory rate index, and fatigue state index. Among them, the heart rate variability time-domain index includes standard deviation SDNN and HRV abnormality; the respiratory rate index includes respiratory rate RF and RF abnormality.

[0010] Furthermore, the expression for calculating the fatigue state index is as follows: ;in, This is a fatigue state index; HRV abnormality; For RF anomaly degree; This represents the rate of change in grip strength.

[0011] Furthermore, the cloud uses edge health status snapshots as input to construct a multimodal fusion model, employing transfer learning loss to optimize the generalization ability of health features across vehicles; and uses differential privacy technology to add Laplacian noise to the health features uploaded from the edge. ,in, For Laplace noise variables; The Laplace distribution scaling parameter, , For feature sensitivity, Budget for privacy.

[0012] Furthermore, when using a communication quality-aware parameter sparse update mechanism to address fluctuations in the vehicle communication environment, the on-board communication module detects communication indicators once per second. These indicators include signal strength RSRP, signal-to-noise ratio SNR, and packet loss rate PLR. The data of each communication indicator is preprocessed, and a communication quality score is calculated using the processed indicators. The communication quality score is then analyzed, and sparse updates are dynamically implemented based on the analysis results.

[0013] Furthermore, by using reinforcement learning to dynamically map health characteristics and risk levels, a state space and action space are defined, and a multi-objective reward function is constructed with early warning accuracy as the core; among them, four risk intervention actions are defined. , These correspond to no warning, mild warning, moderate warning, and severe warning, respectively.

[0014] Furthermore, when the tiered early warning and intervention strategy is triggered, the risk intervention action output by reinforcement learning triggers the corresponding level of intervention measures; among which, the corresponding risk level is determined based on the fatigue state index.

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are as follows: By fusing capacitance, pressure, and radar signals, this invention can compensate for the deficiencies of a single sensor and effectively improve the accuracy of physiological feature extraction; the lightweight adaptive filtering model generated based on hardware-aware meta-neural architecture search adapts to the constraints of vehicle edge computing power and can effectively reduce inference latency and parameter size compared to traditional CNN models; the real-time interference feature perception mechanism enables the model to adjust the filtering strategy according to vehicle vibration and lighting changes, enhancing the adaptive capability of dynamic scenes and improving the feature signal-to-noise ratio under complex road conditions.

[0016] This invention effectively reduces the latency of generating health status snapshots at the edge through model quantization and hardware optimization, further improving the efficiency of health risk warning. By offloading complex model training tasks to the cloud, the edge only performs lightweight inference, effectively reducing the occupancy of onboard computing resources. At the same time, it achieves cross-vehicle health knowledge sharing through transfer learning, effectively solving the resource conflict problem. Based on the communication quality-driven dynamic parameter update strategy, it can effectively reduce data transmission volume in weak communication environments, avoiding model update failures caused by communication fluctuations. By combining differential privacy technology with a sparse update mechanism, it can both protect the privacy of driver physiological data and achieve continuous model optimization, solving the balance problem between data security and model performance.

[0017] This invention constructs a three-dimensional twin model to achieve real-time linkage between driver physiology, vehicle status, and driving environment. Compared with traditional single-dimensional models, it can effectively reduce the prediction error of risk status. The risk mapping model based on the near-end strategy optimization algorithm can adapt to the differences in physiological characteristics of different drivers, which can effectively improve the accuracy of risk warnings for novice drivers. The warning strategy based on edge hardware acceleration can effectively reduce the triggering delay compared with existing technical solutions and cloud-triggered warnings, allowing drivers sufficient reaction time. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 is a flowchart of the operation of the vehicle-to-cloud time health multidimensional analysis and risk assessment system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] As shown in Figure 1, this invention is a vehicle-cloud-based multi-dimensional health analysis and risk assessment system, including a multi-source non-contact physiological signal acquisition and processing module, a cloud-edge collaborative construction and dynamic update module, and a vehicle-mounted digital twin health risk assessment and intervention module. The multi-source non-contact physiological signal acquisition and processing module collects multi-source non-contact physiological signals from steering wheel capacitive sensing, seat pressure distribution, and millimeter-wave radar. It uses a meta-neural architecture to search and generate a lightweight adaptive filtering model adapted to the vehicle's edge hardware, eliminating vehicle vibration and light interference in real time, and outputting a low-dimensional, lightweight physiological feature vector. Specific steps include: during synchronous acquisition of multi-source non-contact physiological signals, the capacitive sensor captures subtle heart rate fluctuations, the pressure sensor senses changes in body pressure caused by respiration, and the millimeter-wave radar detects chest movement data, achieving time synchronization of the three signals via the vehicle's CAN bus. Specifically, eight capacitive sensing electrodes are evenly arranged around the circumference of the steering wheel, covering the hand contact area; a 16×16 array thin-film pressure sensor is installed in the seat cushion area; and a 77GHz millimeter-wave radar module is deployed on the inner side of the vehicle's A-pillar, facing the driver's chest area. A fused multi-source physiological signal matrix is ​​output every 100ms. Where N is the number of sampling points per frame, and the value 3 corresponds to the capacitance dimension, pressure dimension, and radar signal dimension, respectively; and, when sensing and modeling interference features in real time, data from the vehicle-mounted three-axis accelerometer and the in-vehicle illumination sensor are collected simultaneously, corresponding to vehicle vibration interference and illumination change interference, respectively; the collected data are preprocessed, including but not limited to noise reduction and outlier processing, which are existing conventional technical solutions, and the specific implementation steps are not detailed here; the preprocessed data are calculated and combined through vibration interference feature processing to obtain the vibration interference feature vector. ;in, The root mean square acceleration reflects the average energy intensity of vibration. A higher value indicates a more bumpy driving environment, such as unpaved roads, which causes stronger and more continuous interference with physiological signals. This can be achieved through the formula... Calculated; where, The i-th acceleration sample value represents the instantaneous vibration acceleration of the vehicle at that sampling moment, including the total acceleration synthesized from the X, Y, and Z axes; i is the acceleration sampling index, i=1, 2, 3, ..., M; M is the number of sampling points of the single frame acceleration signal, determined by the sampling rate of the vehicle-mounted acceleration sensor and the data frame duration; it is worth noting that the formulas and expressions involved in the embodiments of the present invention are all standardized before calculation, including but not limited to extracting the numerical values ​​of the calculation data and normalization processing; Peak ground acceleration (PGA) reflects the instantaneous impact intensity of vibration, such as when a vehicle passes over a speed bump or manhole cover, generating extremely high PGA values. Such a sudden, strong impact can cause abrupt distortion of physiological signals; through the formula Calculated; where, It is the set of all M acceleration samples within a single frame; The absolute value of each acceleration sample is taken to eliminate the influence of the vibration direction; The dominant vibration frequency is determined by the formula. Calculated; where, A complete acceleration sampling sequence within a single frame ; Perform a fast Fourier transform on the acceleration sequence to convert the time-domain vibration signal into a frequency-domain energy distribution, and output the vibration energy values ​​corresponding to different frequencies; The frequency value corresponding to the maximum energy value in the frequency domain energy distribution reflects the core frequency characteristics of vibration. For example, the main vibration frequency of a vehicle engine at idle speed is approximately 10-20Hz; the main vibration frequency of road bumps during high-speed driving is approximately 2-5Hz. Vibrations of different frequencies interfere with physiological signals in different ways; for example, low-frequency vibrations easily interfere with respiratory signals, while high-frequency vibrations easily interfere with heart rate signals. Additionally, light intensity... With instantaneous rate of change As a constraint, when When a scene is marked as having strong light interference, the anti-interference enhancement mode of the filtering model is triggered. When generating a lightweight adaptive filtering model based on hardware-aware meta-neural architecture search (NAS), a candidate operation space is constructed, consisting of depthwise separable convolutional layers, channel attention layers, and global pooling layers. Each layer supports three parameter configurations, such as convolutional kernel size of 3×3 or 5×5, and the number of channels of 8 or 16. With vehicle edge hardware adaptability as the core, a three-dimensional optimization objective is constructed, involving the following expression: Where T is the model inference latency, the total time the model takes to process a single sample physiological signal on the target vehicle edge hardware, i.e., the complete inference time from inputting a multi-source physiological signal matrix to outputting a 128-dimensional health feature vector, measured by performance testing tools for vehicle edge chips, such as TensorRT and ONNX Runtime; P is the model parameter scale, i.e., the total number of all trainable parameters in the model, including convolutional kernel weights, fully connected layer weights, attention mechanism parameters, and interference feature branch parameters, etc., the number of trainable parameters is counted through the model.parameters() interface of deep learning frameworks, such as PyTorch and TensorFlow; A is the physiological feature accuracy after filtering, i.e., the performance ratio of the lightweight model to the original high-precision model, measuring the degree of accuracy loss during the lightweighting process and ensuring the reliability of health monitoring results; All are weighting coefficients. The default values ​​are 0.4, 0.3, and 0.3, prioritizing low latency. During the search process, an onboard MCU simulator, such as the NXP i.MX8 platform, is invoked in real-time to verify the actual runtime latency of the candidate models. A lightweight adaptive filtering model adapted to the onboard environment is generated through iterative search using the Proximal Policy Optimization (PPO) algorithm. It should be noted that Proximal Policy Optimization (PPO) is a widely used policy gradient algorithm in reinforcement learning. PPO aims to improve policy performance as much as possible while ensuring that policy updates are not too large. It introduces a trust region constraint to limit the magnitude of each policy update, thereby avoiding training instability or even crashes due to excessive policy changes. Unlike traditional policy gradient methods, PPO does not directly use the original policy gradient, but instead employs a pruning mechanism or adaptive KL divergence penalty to control the update step size. The Proximal Policy Optimization algorithm is a conventional existing technique; its specific implementation steps will not be elaborated here. When performing real-time filtering and outputting low-dimensional features, the multi-source physiological signal matrix and interference features are combined... Parallel Input Lightweight Adaptive Filtering Model The lightweight adaptive filtering model enhances effective physiological features and suppresses interference signals through an attention mechanism. The lightweight adaptive filtering model enhances existing conventional technical solutions through an attention mechanism; the specific implementation steps are not detailed here. Through global average pooling operations in the output layer of the lightweight adaptive filtering model, the high-dimensional filtering results are compressed into a 128-dimensional low-dimensional physiological feature vector. This low-dimensional physiological feature vector includes multi-dimensional fusion features such as heart rate variability, respiratory depth, and body pressure distribution. The low-dimensional physiological feature vector is cached in the vehicle edge storage unit, providing a low-bandwidth, low-latency input data foundation for subsequent cloud-edge collaborative health analysis.

[0022] In this embodiment of the invention, by fusing capacitance, pressure, and radar signals, the shortcomings of a single sensor can be compensated for, and the accuracy of physiological feature extraction can be effectively improved. The lightweight adaptive filtering model generated by hardware-aware meta-neural architecture search is adapted to the constraints of vehicle edge computing power and can effectively reduce inference latency and parameter size compared with traditional CNN models. The real-time interference feature perception mechanism enables the model to adjust the filtering strategy according to vehicle vibration and lighting changes, enhancing the adaptive capability of dynamic scenes and improving the feature signal-to-noise ratio under complex road conditions. The output low-dimensional physiological feature vector is compressed by 90% compared with the original signal data, which can provide a low-bandwidth and low-latency data foundation for cloud-edge collaborative transmission and analysis in subsequent steps.

[0023] The cloud-edge collaborative construction of a dynamic update module is based on lightweight physiological feature vectors. A hierarchical federated transfer learning cloud-edge collaborative framework is constructed. At the edge, a lightweight inference model is deployed to output a real-time health status snapshot. In the cloud, multi-vehicle health feature knowledge is aggregated through federated transfer learning, and a communication quality-aware parameter sparse update mechanism is used to address fluctuations in the vehicle communication environment. Specific steps include: integrating the lightweight adaptive filtering model generated by the meta-neural architecture search with the distilled health inference model, deploying it on an in-vehicle edge computing unit, such as the NXP i.MX8MP chip, and converting the model precision from FP32 to INT8 using TensorRT quantization to further reduce inference latency; when generating a health status snapshot, based on the output 128-dimensional low-dimensional physiological feature vector, the edge model calculates three core health indicators in real time: heart rate variability (HRV) time-domain index, respiratory rate index, and fatigue state index. The HRV time-domain index includes standard deviation (SDNN) and HRV anomaly. The expression for calculating standard deviation (SDNN) is: ;in, 1 is the standard deviation of the RR interval sequence, reflecting the long-term fluctuation of heart rate. The reference range for normal adults is 50-150 ms. K is the number of RR intervals in a single analysis window. Usually, a 1-minute window is used, that is, K is the number of heartbeats in 1 minute. The normal range for adults is about 60-100 beats per minute. is the duration of the j-th RR interval, representing the time interval between two adjacent R waves of the heartbeat; j is the RR interval index, j=1, 2, 3, ..., K; This represents the average of all RR intervals within a single window. It's important to explain that the RR interval is the time interval between two adjacent R-wave peaks in an electrocardiogram (ECG), typically measured in milliseconds or seconds, directly reflecting the heart's beat cycle. The R-wave is the highest-amplitude positive peak in the ECG waveform, representing the synchronous depolarization of the left and right ventricles. It is the most distinctive electrical signal node in the heartbeat cycle and the most easily identifiable characteristic wave in clinical ECG analysis. R-wave detection can automatically identify the R-wave peak position in each heartbeat cycle using ECG signal processing algorithms, such as the classic Pan-Tompkins algorithm, recording its timestamp, for example, the first R-wave at 100ms and the second at 1100ms. The difference in timestamps between two adjacent R-waves is a single RR interval. Continuously calculated RR intervals form an RR interval sequence. The expression for calculating HRV abnormality is: ;in, HRV abnormality; The driver's personal SDNN baseline value is calibrated using data from the first 30 minutes of driving or a general normal average value, taken as 100. This represents the standard deviation of normal fluctuations in SDNN, with a general value of 30. The anomaly score is truncated to the [0,1] interval. When the SDNN deviates from the baseline by more than one standard deviation, the anomaly score is directly set to 1, representing a severe anomaly. The respiratory rate index includes respiratory rate (RF) and RF anomaly score. The expression for calculating respiratory rate (RF) is: ;in, Respiratory rate; This is a time-domain signal sequence from the seat pressure sensor, with up to 1000 sampling points; Perform a Fast Fourier Transform on the pressure signal sequence to convert the time-domain signal into a frequency-domain energy distribution; This indicates the frequency value corresponding to the maximum energy within the frequency band corresponding to a respiratory rate of 0.1-0.5Hz; 60 indicates converting the frequency value to breaths / minute to obtain the final respiratory rate (RF), with a normal adult reference range of 12-20 breaths / minute; the expression for calculating RF abnormality is: ;in, For RF anomaly degree; The driver's individual respiratory rate baseline is 16 breaths / minute, determined through initial driving calibration or by using a general normal average. The RF anomaly score ranges from [0,1]. When the respiratory rate is below 12 breaths / minute (too slow) or above 20 breaths / minute (too fast), the RF anomaly score increases linearly with the degree of deviation, reaching 1 when it exceeds the normal range by more than twice. The fatigue state index is calculated using the following formula: ;in, This is a fatigue state index; For the rate of change of grip strength, , The grip force value detected by the steering wheel capacitive sensor at the current moment is converted into relative grip force through the capacitance value, with a range of 0-1. This represents the average grip strength over the first 10 seconds; the grip strength change rate reflects the stability of the driver's hand muscles. Grip strength fluctuations increase significantly during fatigue, with a normal range of 0-0.2, and a value of 1 when it exceeds 0.5; 0.4, 0.3, and 0.3 are preset weighting coefficients for each indicator, which can be customized according to the application needs of the actual application scenario; among them, the fatigue state index... The range is [0,1]. A larger value indicates a higher level of fatigue. For example: 0-0.3 corresponds to a normal state; 0.3-0.6 corresponds to mild fatigue, triggering a seat vibration alert; 0.6-0.9 corresponds to moderate fatigue, triggering voice and visual warnings; 0.9-1.0 corresponds to severe fatigue, triggering vehicle assistance controls such as automatic deceleration and lane keeping. A health status snapshot is output every 200ms. S=[ [Confidence], where the confidence is determined by the maximum softmax probability output of the model and is used to mark the reliability of the inference results at the edge. It should be explained that the generation of edge health status snapshots directly depends on the generated lightweight physiological feature vectors, which meet the real-time requirements through low-latency inference. When the cloud aggregates multi-vehicle health feature knowledge through federated transfer learning, it constructs a hierarchical architecture of edge-regional cloud-central cloud. The regional cloud is responsible for aggregating the health data of vehicles within a range of 50-100km, and the central cloud realizes cross-regional global knowledge fusion. The cloud uses the edge health status snapshot as input to construct a multimodal fusion model and uses transfer learning loss to optimize the generalization ability of cross-vehicle health features. The relevant expressions are: ;in, This is the total loss function; Cross-entropy loss for classifying health status. , For the first The true health status label of each sample, such as normal, mildly abnormal, and severely abnormal, is encoded using one-hot encoding. For sample index, =1, 2, 3, ... , The total number of samples; The model predicts the probability that the sample belongs to the corresponding health state. For knowledge distillation loss, the soft-label knowledge of the central cloud model is transferred to the regional cloud model, typically using KL divergence loss. , This represents the predicted probability distribution of the cloud-based model. This represents the probability distribution of the health status prediction by the local model at the edge. Let KL divergence be a metric. The classification loss weights range from 0 to 1, with a default value of 0.6. Differential privacy techniques are used to add Laplacian noise to the health features uploaded at the edge. ,in, For Laplace noise variables; The Laplace distribution scaling parameter, , For feature sensitivity, For the privacy budget, a value of 1 is used to ensure the privacy and security of driver physiological data. After aggregating health data from multiple vehicles in the cloud, a driver health-driving behavior-driving environment correlation graph is constructed and stored as a graph database for knowledge reasoning in subsequent risk assessment. It should be noted that cloud-based federated transfer learning uses edge health snapshots as local training samples, which can achieve multi-vehicle knowledge aggregation while avoiding the transmission of the entire original data. When using a communication quality-aware parameter sparse update mechanism to deal with fluctuations in the vehicle communication environment, the communication indicators are detected once per second based on the vehicle communication module. The communication indicators include signal strength RSRP, signal-to-noise ratio SNR, and packet loss rate PLR. The data in the communication indicators are preprocessed, including but not limited to noise reduction and outlier handling, which are existing conventional technical solutions. The specific implementation steps are not detailed here. The processed communication indicators are then expressed using the formula... The Communication Quality Score (CQS) is calculated, ranging from 0 to 1, with 1 representing optimal communication quality. Values ​​of 0.5, 0.3, and 0.2 represent different weights for different communication metrics and can be customized based on application specifications in specific scenarios. Data analysis of the CQS is performed, and sparse updates are dynamically implemented based on the analysis results. If CQS ≥ 0.8, it is considered high-quality communication, and a full synchronization of cloud model parameters is performed at a frequency of 5 minutes. If CQS < 0.8, it is considered medium-quality communication, and sparse updates using channel pruning are employed, synchronizing only the top 20% of channel parameters by absolute weight, achieving an 80% parameter compression rate. If CQS < 0.4, it is considered weak communication, and parameter synchronization is paused. The edge device uses a locally cached historical best model, and incremental updates are performed after communication quality recovers. When the sparsely updated parameters are sent back and the edge model is updated, the cloud device updates the sparsed parameter increments. Transmitted to the edge via the QUIC protocol, the edge model is then trained using an incremental learning algorithm. Update the weights, where, This is the updated weight matrix for the edge model; The weight matrix before the edge model is updated is the weights that were previously downloaded from the cloud or trained locally, serving as the basis for incremental learning; The parameter increment after sparsification; The learning rate is set to 0.1 by default to ensure smooth model updates without affecting real-time inference; the edge side switches to... Real-time health status inference is performed to achieve smooth iteration of model performance, while ensuring the privacy and security of driver physiological data.

[0024] It should be explained that the parameter sparse update mechanism for communication quality awareness dynamically adjusts the interaction between the cloud and the edge based on the current communication environment, ensuring that the model can continue to optimize under complex driving conditions without affecting real-time health analysis.

[0025] In this embodiment of the invention, model quantization and hardware optimization can effectively reduce the latency of generating health status snapshots at the edge, further improving the efficiency of health risk warning. By offloading complex model training tasks to the cloud, the edge only performs lightweight inference, effectively reducing the occupancy of onboard computing resources. At the same time, transfer learning enables cross-vehicle health knowledge sharing, effectively solving the resource conflict problem. Based on the communication quality-driven dynamic parameter update strategy, the amount of data transmission can be effectively reduced in weak communication environments, avoiding model update failures caused by communication fluctuations. By combining differential privacy technology with a sparse update mechanism, the privacy of driver physiological data can be guaranteed while continuous model optimization can be achieved, solving the balance between data security and model performance.

[0026] The in-vehicle digital twin health risk assessment and intervention module: Based on the processed edge-end health status snapshots and cloud-based federated knowledge, an in-vehicle digital twin health risk simulator is constructed. Through reinforcement learning, health characteristics and risk levels are dynamically mapped to trigger tiered early warning and intervention strategies. Specific steps include: constructing a three-dimensional twin model of driver physiology, vehicle state, and driving environment based on the processed edge-end health status snapshots and cloud-based graph database, including driver twin nodes, vehicle twin nodes, and environment twin nodes; specifically: the driver twin node is associated with processed HRV anomaly, RF anomaly, fatigue state index, and historical health baseline data; historical health baseline data can be provided through cloud-based knowledge aggregation; the vehicle twin node is associated with real-time vehicle speed, steering frequency, brake pedal pressure, and other in-vehicle CAN bus data; the environment twin node is associated with navigation road condition levels, real-time weather, traffic flow, and other V2X data; and the multi-source data is fused into a digital twin state vector. The expression is: ;in, Real-time vehicle speed, normalized to the [0,1] interval; Steering frequency reflects the driver's operational stability; The brake pedal pressure is normalized to the [0,1] interval. The road condition level is represented by a value of 0 or 1, where 0 indicates smooth traffic and 1 indicates congestion. This level can be obtained in real time through navigation data. The synchronization of physical state to a digital twin is achieved through memory mapping technology in the vehicle's edge computing unit. When dynamically mapping health features and risk levels using reinforcement learning, the state space S is defined as: a twin state vector... The input is 7-dimensional, with all features normalized to the [0,1] interval; the action space A is defined, which defines 4 types of risk intervention actions. , These correspond to no warning, mild warning, moderate warning, and severe warning, respectively; with warning accuracy as the core objective, a multi-objective reward function is constructed: ;in, The flag for a correct early warning is set to 1 if the warning is correctly triggered, and 0 otherwise. This is a flag for missed reports; it is set to 1 if the expected warning is not triggered, and 0 otherwise. This is a false alarm flag; it is set to 1 when an error triggers a warning, and 0 otherwise. These are the weights for correct warning rewards, missed warning penalties, and false alarm penalties, with default values ​​of 1, 0.8, and 0.5, respectively. During model training and deployment, a near-end strategy optimization algorithm is used, employing a dataset of driver health-risk correlation data aggregated in the cloud as the training set. The model is iteratively trained for 1000 rounds until its risk level prediction accuracy is ≥95%. The trained model is quantized into INT8 format and deployed on an in-vehicle edge FPGA chip. When a tiered warning intervention strategy is triggered, the corresponding level of intervention measures is triggered based on the risk intervention action output by reinforcement learning. Among these measures, the fatigue state index is used as the basis for intervention. Determine the corresponding risk level: If If the risk level is determined to be no risk, the intervention measure is no warning, and physiological data will be continuously collected; if If the risk level is determined to be mild, the intervention measure is a low-frequency vibration reminder for the seat, such as a frequency of 5Hz for 1 second; if If the risk level is determined to be moderate, the intervention measure will be a voice warning and a flashing yellow light on the dashboard; if If the risk level is determined to be severe, the intervention measures are voice emergency warning, automatic lane keeping assist, and slow deceleration to 60km / h. The warning strategy logic is embedded in the hardware state machine of the edge FPGA, which does not require calling the CPU or cloud communication. It directly triggers the intervention action based on the risk level signal output by the twin model. After the warning is triggered, the driver's reaction data, such as grip strength recovery rate and vehicle speed adjustment range, is collected in real time. The feedback data is sent back to the cloud every hour, and the federated transfer learning model is updated to simultaneously optimize the risk mapping accuracy of the digital twin.

[0027] In this embodiment of the invention, by constructing a three-dimensional twin model, real-time linkage between driver physiology, vehicle state, and driving environment is achieved, which can effectively reduce the prediction error of risk state compared with traditional single-dimensional models. The risk mapping model based on the near-end strategy optimization algorithm can adapt to the differences in physiological characteristics of different drivers, which can effectively improve the accuracy of risk warning for novice drivers. The warning strategy based on edge hardware acceleration can effectively reduce the triggering delay compared with existing technical solutions and cloud-triggered warnings, allowing drivers sufficient reaction time. The warning feedback data can continuously optimize the reinforcement learning model and the cloud-based federated knowledge graph, realizing a closed loop of the entire process of data collection, model training, risk warning, and feedback optimization. The overall system performance gradually improves with usage time, ensuring the reliability of the system's risk assessment.

[0028] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0029] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0030] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0031] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system, characterized in that, include: Multi-source non-contact physiological signal acquisition and processing module: Acquires multi-source non-contact physiological signals from steering wheel capacitive sensing, seat pressure distribution, and millimeter-wave radar. Employs a meta-neural architecture to search and generate a lightweight adaptive filtering model adapted to in-vehicle edge hardware, eliminating vehicle vibration and light interference in real time, and outputting low-dimensional lightweight physiological feature vectors. Cloud-edge collaborative construction and dynamic update module: Based on the lightweight physiological feature vectors acquired through processing, constructs a hierarchical federated transfer learning cloud-edge collaborative framework. The edge uses the deployed lightweight inference model to output a real-time health status snapshot, while the cloud aggregates health feature knowledge from multiple vehicles through federated transfer learning and adopts a communication quality-aware parameter sparse update mechanism to cope with fluctuations in the vehicle communication environment. The vehicle-mounted digital twin health risk assessment and intervention module: Based on the processed edge health status snapshot and cloud federated knowledge, a vehicle-mounted digital twin health risk simulator is constructed. Through reinforcement learning, health characteristics and risk levels are dynamically mapped to trigger graded early warning and intervention strategies.

2. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 1, characterized in that, When performing multi-source non-contact physiological signal synchronous acquisition, the system captures subtle heart rate fluctuations through a capacitive sensor, senses changes in body pressure caused by respiration through a pressure sensor, and detects chest cavity movement data through a millimeter-wave radar. The system also achieves time synchronization of the three signals through the vehicle's CAN bus.

3. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 2, characterized in that, With vehicle edge hardware adaptability as the core, a three-dimensional optimization objective is constructed, involving the following expression: Where T is the model inference delay; P is the model parameter size; and A is the accuracy of the filtered physiological features. All are weighting coefficients. A lightweight adaptive filtering model adapted to the in-vehicle environment is generated through iterative search using a near-end strategy optimization algorithm.

4. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 3, characterized in that, When performing real-time filtering and outputting low-dimensional features, the multi-source physiological signal matrix and interference features are input into the lightweight adaptive filtering model in parallel. The lightweight adaptive filtering model strengthens effective physiological features and suppresses interference signals through an attention mechanism. Through the global average pooling operation of the output layer of the lightweight adaptive filtering model, the high-dimensional filtering result is compressed into a 128-dimensional low-dimensional physiological feature vector. The low-dimensional physiological feature vector includes multi-dimensional fusion features of heart rate variability, respiratory depth, and body pressure distribution.

5. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 4, characterized in that, When generating a health status snapshot, the edge model calculates three core health indicators in real time based on the output low-dimensional physiological feature vector: heart rate variability time domain index, respiratory rate index, and fatigue state index. Among them, the heart rate variability time domain index includes standard deviation SDNN and HRV abnormality; the respiratory rate index includes respiratory rate RF and RF abnormality.

6. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 5, characterized in that, The formula for calculating the fatigue state index is: ;in, This is a fatigue state index; HRV abnormality; For RF anomaly degree; This represents the rate of change in grip strength.

7. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 5, characterized in that, The cloud platform uses edge health status snapshots as input to construct a multimodal fusion model, employing transfer learning loss to optimize the generalization ability of health features across vehicles. Differential privacy technology is used to add Laplacian noise to the health features uploaded from the edge. ,in, For Laplace noise variables; The Laplace distribution scaling parameter, ， For feature sensitivity, Budget for privacy.

8. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 7, characterized in that, When using a communication quality-aware parameter sparse update mechanism to address fluctuations in the vehicle communication environment, the on-board communication module detects communication indicators once per second. These indicators include signal strength RSRP, signal-to-noise ratio SNR, and packet loss rate PLR. The data of each communication indicator is preprocessed, and a communication quality score is calculated using the processed indicators. The communication quality score is then analyzed, and sparse updates are dynamically implemented based on the analysis results.

9. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 8, characterized in that, When dynamically mapping health characteristics and risk levels using reinforcement learning, a state space and action space are defined, and a multi-objective reward function is constructed with early warning accuracy as the core; among them, four risk intervention actions are defined. ， These correspond to no warning, mild warning, moderate warning, and severe warning, respectively.

10. The vehicle-to-cloud time-based health multi-dimensional analysis and risk assessment system according to claim 9, characterized in that, When the graded early warning and intervention strategy is triggered, the risk intervention action output by reinforcement learning will trigger the corresponding level of intervention measures; among them, the corresponding risk level is determined based on the fatigue state index.

Images