Vehicle networking driving state monitoring method and system

Abstract

The invention discloses an Internet of Vehicles driving state monitoring method and system, and the method comprises the steps: 1, collecting the driving state data in real time, and enabling the driving state data to comprise the behavior data of a driver and environment data; step 2, preprocessing the collected multi-source heterogeneous data at a vehicle-mounted edge computing node; step 3, inputting the fused data into a variational auto-encoder, and when the variational auto-encoder extracts the low-dimensional hidden variable representation of the driving state data, deducing hidden variable distribution by optimizing a variational lower bound; 4, inputting the inferred hidden variables into the fractional order differential equation model to obtain a predicted value of the driving state; step 5, federated learning is carried out on a plurality of vehicle-mounted edge nodes, and collaborative optimization of a global model is realized while data privacy is protected; according to the method, distributed collaborative decision-making of multiple agents is realized by solving Nash equilibrium, the adaptability and robustness of the system are improved, and the traffic safety and the driving experience can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to a method and system for intelligently monitoring driving status of an Internet of Vehicles, and belongs to the field of intelligent networked vehicles. Background Art

[0002] With the rapid development of connected vehicle technology, it has become particularly important to monitor drivers' behavior and physiological status in real time to improve road safety. Currently, most monitoring systems process data from drivers through central servers, resulting in insufficient response speed in emergency situations.

[0003] Most monitoring systems use traditional integer-order differential equations to model the dynamic changes of driver status, which makes it difficult to accurately reflect the long-range correlations and non-local effects of complex systems. Existing methods usually directly process high-dimensional raw data and lack effective feature extraction and dimensionality reduction steps, resulting in high computational complexity and poor model generalization. Most current monitoring systems rely on centralized data processing and decision-making, lack distributed collaborative optimization mechanisms, and are difficult to adapt to changing driving environments and individual differences. Existing driver status assistance methods are usually based on simple threshold judgments or rule reasoning, and lack systematic consideration of the dynamic interactions between drivers, vehicles, and the environment.

[0004] 6G networks can provide high speed and low latency, providing a basis for solving the above problems. The existing application technologies of 6G networks need to be further solved to reduce data transmission delays and unreasonable allocation of computing resources during monitoring. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing vehicle networking system has data processing delay when monitoring the driver's status in real time, and it is necessary to further improve traffic safety and driving experience.

[0006] In order to solve the above technical problems, the present invention provides a method for monitoring driving status of an Internet of Vehicles, comprising the following steps: Step 1: multiple sensors are installed in the vehicle to collect driving status data in real time, wherein the driving status data includes driver behavior data and environmental data; Step 2: Preprocess the collected multi-source heterogeneous data at the vehicle-mounted edge computing node. The preprocessing includes data cleaning, data normalization, and data fusion. Step 3: The fused data The input is fed into a variational autoencoder, which infers the latent variable distribution by optimizing the variational lower bound when extracting the low-dimensional latent variable representation of the driving state data. ; Step 4: Infer the latent variables Input into the fractional differential equation model, the fractional differential equation model describes the dynamic change law of the driving state, and obtains the predicted value of the driving state by solving the fractional differential equation , reflecting the driver’s physiological and behavioral state; Step 5: Federated learning is performed on multiple vehicle-mounted edge nodes. Through encrypted gradient aggregation and differential privacy algorithms, collaborative optimization of the global model is achieved while protecting data privacy. Step six, based on the game model, construct a dynamic interaction process model of driver-vehicle-environment. The dynamic interaction process model of driver-vehicle-environment adopts the HJB equation. By solving the HJB equation, collaborative optimization between participants is achieved in a dynamic environment, so that participants can obtain optimal decision-making driving data in complex driving scenarios.

[0007] The aforementioned method for monitoring driving status of an Internet of Vehicles further includes: Step seven: Based on the optimal decision-making driving data obtained in step six, the deep reinforcement learning algorithm is used to dynamically adjust the on-board computing, storage and communication resources according to the real-time road and network conditions to minimize energy consumption and task delays.

[0008] In the aforementioned method for monitoring the driving state of an Internet of Vehicles, in step 2, during the data fusion process, the weights are dynamically adjusted by solving the following multi-objective optimization problem: , , , in, represents the mean square error, For data The reconstruction value of and are respectively minimizing the reconstruction error and the L2 norm of the weight, and the adaptive weight ; The data type.

[0009] In the aforementioned method for monitoring the driving state of an Internet of Vehicles, in step three, the variational autoencoder is expressed as: , in, is the parameterized probability distribution of the decoder, is the prior distribution, are the parameters to be learned, is the divergence, which measures the difference between two probability distributions; is a hidden variable; is the reconstruction error.

[0010] In the aforementioned method for monitoring driving status of an Internet of Vehicles, in step 4, the fractional-order differential equation is: , , in, The order is The Caputo fractional derivative of is defined as: , , in, is the Gamma function, is a positive integer, when When , the Caputo derivative degenerates into the classical first-order derivative; is the integral variable, representing a point in time in the past; is a driving state variable, the driving state variable; is a nonlinear function, is the process noise.

[0011] In the aforementioned method for monitoring driving status of an Internet of Vehicles, in step 5, the optimization process of federated learning is: , , , in, For the The loss function calculated by each node based on local data, is the corresponding gradient, is the total number of nodes, in In the iteration, each node first calculates the local gradient , and then send the gradient to the central node through a secure encrypted communication protocol; the central node aggregates the received gradients and adds -Gaussian noise for differential privacy , get the global gradient ; Finally, the central node uses the global gradient to update the model parameters , and broadcast the updated model to each node to complete one iteration of federated learning.

[0012] In the aforementioned method for monitoring driving status of an Internet of Vehicles, in step 6, the game model is defined as ; in is the set of game participants, and Participants in the game The state space and action space of For participants The game participants include the driver, the vehicle, and the environment; It represents the dynamic interaction between game participants; The game model is described as the following stochastic differential equation, namely the HJB equation: , in, For participants At the moment The state variables represent the current state of the driver, vehicle or environment; For participants At the moment The action variable represents the actions that the participant can take; For other participants at the moment The action variable represents the action of the participant The actions of all other participants will affect the participants Decisions and status; Indicates participants The state transition function describes how the state changes over time given the current state and action; Indicates participants Noise influence function; is Brownian motion, which represents random disturbance.

[0013] In the aforementioned method for monitoring the driving state of an Internet of Vehicles, in step six, the solution process is: , in, For participants The value function of and time Next, participants The maximum expected return that can be obtained; For participants The state variable represents the participant at time The current status of For participants The action variable represents the actions that the participants can take; For participants The action space includes all possible action choices of the participants; is the optimal action of other participants, indicating the best strategy adopted by other participants in the current state; is the state transition function, describing the and actions and the best actions of other participants When the participants How the status of changes; is the value function for the state variable The gradient of , which indicates the impact of state changes on expected returns; is the noise influence function, which describes the random factors affecting the participants during the state transition process. The impact of Represents the trace operation of a matrix, usually used to calculate the sum of variances or covariances; For participants The utility function is expressed in the state and actions and the best actions of other participants Below, the participants’ immediate payoffs. The superscript T indicates transposition.

[0014] In the aforementioned method for monitoring driving status of an Internet of Vehicles, in step seven, the reinforcement learning model is defined as ,in: state , including the system’s queue status, network status, and data collected by sensors. The status reflects the current system operation and environmental changes; action Refers to resource scheduling decisions, including allocating onboard computing, storage, and communication resources; State transition probability : Describes the probability of transitioning to the next state after taking a certain action in the current state; Reward Function : Used to evaluate the performance of the agent in specific states and actions; Discount Factor : Used to balance the importance of current rewards and future rewards; The objective function is to maximize the expected value of the long-term cumulative reward: , in, is the scheduling strategy, Delayed task, is the maximum constraint value of delay; Reward Function for: , in, and are weight coefficient 1 and weight coefficient 2, which are used to balance the impact of energy consumption and delay.

[0015] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0016] A computer-readable storage medium stores a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0017] A computer program product comprises a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0018] The beneficial effects achieved by the present invention are as follows: the vehicle network driver status monitoring method of the present invention adopts a variational autoencoder to achieve dimensionality reduction and feature extraction of driver status data, and adaptively learns the latent space representation by optimizing the variational lower bound, thereby reducing the model complexity and thus reducing the data transmission delay; the multi-agent collaborative optimization framework based on the stochastic differential game of the present invention realizes the distributed collaborative decision-making of multiple agents by solving the Nash equilibrium, thereby improving the adaptability and robustness of the system, and can significantly improve traffic safety and driving experience.

[0019] At the same time, fractional-order differential equations are introduced to model the dynamic changes of the driver's state. Fractional-order derivatives are used to more accurately describe the memory effect and non-local correlation of complex systems, thereby improving the accuracy of state estimation. An adaptive driver state assistance method is proposed, which comprehensively considers the driver's state and environmental risks, dynamically adjusts the vehicle control parameters and human-computer interaction methods, and improves the real-time and effectiveness of auxiliary decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The present invention is a flow chart of a method for monitoring driving status of an Internet of Vehicles. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for monitoring driving status of an Internet of Vehicles, comprising the following steps: Step 1: Set up multiple sensors in the vehicle to collect driving status data in real time. The driving status data includes the driver's behavior data and environmental data. The behavior data includes eye movements, head posture, facial expressions, etc., and the environmental data includes temperature, humidity, noise, etc. in the vehicle.

[0023] Step 2: At the vehicle-mounted edge computing node, collect the multi-source heterogeneous data Preprocessing is performed, including data cleaning, data normalization and data fusion. The vehicle-mounted edge computing node can be a small computer or embedded device installed on the vehicle, located in the vehicle environment close to the data source, and can process the data nearby, reducing the time and bandwidth consumption of uploading data to the cloud.

[0024] In the data fusion process, the weights are dynamically adjusted by solving the following multi-objective optimization problem: , , , in, represents the mean square error, For data The reconstruction value of and They are respectively to minimize the reconstruction error and the L2 norm of the weight to improve the fusion accuracy and robustness, and the adaptive weight ; The data type.

[0025] Step 3: The fused data The input is fed into a variational autoencoder, which extracts a low-dimensional latent variable representation of the driving state data by optimizing the following variational lower bound (ELBO) to infer the latent variable distribution , the variational autoencoder is expressed as: , in, is the parameterized probability distribution of the decoder, is the prior distribution, usually the standard normal distribution is chosen, are the parameters to be learned, KL (Kullback-Leibler divergence) divergence is used to measure the difference between two probability distributions; is a hidden variable; In the above formula, the reconstruction error part in the variational lower bound (ELBO) is optimized by the stochastic gradient variational Bayes (SGVB) algorithm. Specifically, the latent variable is inferred by the SGVB algorithm. , where the first term is the reconstruction error, which indicates the model's ability to reconstruct the original data. The second term is the regularization term, which makes the distribution of latent variables close to the prior distribution. The SGVB algorithm optimizes the target through random sampling and reparameterization techniques, thereby improving the learning efficiency and performance of the model.

[0026] High-dimensional data has a large number of features, usually when the number of features is close to or greater than the number of samples, and may face sparsity, computational complexity, and overfitting risks when analyzing. Low-dimensional data has a relatively small number of features, which is much smaller than the number of samples. Low-dimensional data analysis is relatively simple and easy to process and model. By using variational autoencoders, the system converts high-dimensional driving state data into low-dimensional latent variable representations, thereby reducing computational complexity and improving the efficiency and accuracy of the model.

[0027] Variational lower bound It consists of two parts: reconstruction error and the regularization term , where the reconstruction error measures the latent variable For the original data The regularization term makes the distribution of latent variables close to the prior distribution, which plays a role of constraint and smoothing.

[0028] The network structure and hyperparameters of the variational autoencoder can be optimized according to the complexity of the driver's state data and the task requirements. In the optimization of the network structure and hyperparameters of the variational autoencoder, for high-dimensional, nonlinear physiological signal data, the following categories are mainly included: Electroencephalogram (EEG): used to monitor brain activity and can reflect emotional and cognitive states; Electrocardiogram (ECG): This monitors the heart's electrical activity and is often used to assess physiological states and emotional responses; Electromyography (EMG): used to record muscle activity, which can reflect the body's tension and emotional state; Galvanic Skin Response (GSR): used to monitor changes in the skin's electrical conductance, often associated with emotional states; Respiratory pattern (RSP): monitors the breathing rate and depth, which can reflect the physiological and psychological state; Pulse wave (PPG): used to monitor blood flow and can provide information on heart rate and blood oxygen saturation.

[0029] The above signals usually have complex nonlinear characteristics, so multi-layer, nonlinear encoder and decoder structures are required when processing them to better capture the intrinsic relationships and dynamic changes of the data. For low-dimensional, linear behavioral data, a shallower network structure can be used to reduce computational complexity and improve processing efficiency. At the same time, the dimension of the latent variable can also be adaptively adjusted according to the intrinsic degrees of freedom of the data to balance the accuracy of feature extraction and computational efficiency.

[0030] For low-dimensional, linear behavioral data, a shallower network structure is used, using fewer neural network layers and fewer neurons. The feature relationships of a shallower network structure are relatively simple. An overly complex model may lead to overfitting, which in turn reduces the generalization ability of the model. By using a shallower network, the basic patterns of the data can be effectively captured while improving computational efficiency and training speed.

[0031] At the same time, the dimension of the latent variable It can also be adaptively adjusted according to the intrinsic degrees of freedom of the data to balance the accuracy of feature extraction and computational efficiency.

[0032] Step 4: Infer the latent variables Input into the fractional differential equation model, the fractional differential equation model describes the dynamic change law of the driving state, and obtains the predicted value of the driving state by solving the fractional differential equation ; The fractional order differential equation is: , , in, The order is The Caputo fractional derivative of is defined as: , , in, is the Gamma function, is a positive integer, when When , the Caputo derivative degenerates into the classical first-order derivative; is the integral variable, representing a point in time in the past; is a driving state variable, and the driving state variable includes: The driver's physiological state: heart rate, blood pressure, skin galvanic response, etc.; Driver’s behavior: eye movements, head posture, facial expressions, etc. Vehicle status: speed, acceleration, steering angle, etc.; Environmental conditions: road conditions, weather conditions, traffic density, etc.

[0033] is a nonlinear function, are parameters to be learned, and the parameters to be learned include: Model parameters: such as weights and bias parameters in a variational autoencoder; State transfer function parameters: In the fractional differential equation model, function parameters used to describe the dynamic changes of driving state; Parameters of the reward function: In reinforcement learning, the parameters used to optimize the decision.

[0034] is the process noise.

[0035] The predicted value of driving state is obtained by solving fractional differential equations numerically , the driving status includes: Physiological state: fatigue level, emotional state, etc. Attention level: the degree to which attention is focused or divided; Responsiveness: speed and accuracy of response to unexpected situations.

[0036] The numerical solution may adopt a prediction-correction method.

[0037] The fractional-order differential equation can more accurately describe the memory effect and non-local correlation of a complex system by introducing fractional-order derivatives when modeling the dynamic changes of driving states, thereby improving the long-term dependence and robustness of state estimation.

[0038] Step five: Perform federated learning on multiple vehicle-mounted edge nodes, and through encrypted gradient aggregation and differential privacy algorithms, achieve collaborative optimization of the global model while protecting data privacy.

[0039] Specifically, each edge node locally calculates the gradient of the global model parameters of the driver state intelligent assistance system , rather than independent local models, the gradient is encrypted Send to the central node for aggregation: , in, is the number of nodes participating in federated learning, To satisfy -Gaussian noise for differential privacy, distributed as , the central node uses the aggregated gradient Update global model parameters , and broadcast the updated global model to each edge node, thereby achieving collaborative learning among nodes.

[0040] Number of nodes Indicates the number of vehicle-mounted edge nodes participating in the learning. Each node corresponds to a vehicle and reflects the status data of different drivers. Gaussian noise Random noise used to protect user privacy and ensure that sensitive information of individuals is not leaked during the aggregation process; Global model parameters It means that the updated model parameters after aggregation can better adapt to the driver's state changes.

[0041] Through the above method, it is possible to utilize diversified data from different nodes to improve the accuracy of driver status monitoring and the effectiveness of intelligent assisted decision-making while protecting the driver's privacy. The collaborative learning mechanism not only enhances the generalization ability of the model, but also can respond to the driver's status changes in real time, thereby improving overall driving safety.

[0042] In step 5, the federated learning collaboratively optimizes the global model parameters on multiple on-board edge nodes When , distributed learning under privacy protection is achieved through encrypted gradient aggregation and differential privacy technology. The optimization process of federated learning is: , , , in, For the The loss function calculated by each node based on local data, is the corresponding gradient, is the total number of nodes, in In the iteration, each node first calculates the local gradient , and then send the gradient to the central node through a secure encrypted communication protocol (such as homomorphic encryption); the central node aggregates the received gradients and adds -Gaussian noise for differential privacy , and get the global gradient ; Finally, the central node uses the global gradient to update the model parameters , and broadcast the updated model to each node to complete one iteration of federated learning.

[0043] The optimization algorithm and privacy protection mechanism of federated learning can be adapted according to factors such as the communication bandwidth, computing power and security requirements of the vehicle network. For example, when the network bandwidth is limited, gradient compression methods such as quantization and sparsification can be used to reduce communication overhead; when computing resources are limited, distributed optimization algorithms such as local SGD can be used to reduce the computing load of nodes; when privacy protection requirements are high, cryptographic methods such as secure multi-party computing and homomorphic encryption can be further introduced to enhance the strength of privacy protection.

[0044] Step 4 predicts the driving state through the fractional-order differential equation model, which relies on the local data of individual vehicles. However, the local data of a single vehicle may not fully reflect the global law of the driver's state. Therefore, step 5 combines the local data of multiple vehicles through federated learning and optimizes the global model under the premise of ensuring data privacy. This collaborative optimization process helps to improve the accuracy of the fractional-order differential equation model and makes the prediction of driving state more robust and generalizable. The driving state prediction in step 4 is the basis of federated learning collaborative optimization, and the global model optimization of federated learning further improves the accuracy and adaptability of fractional-order differential equation prediction.

[0045] Step 6: Based on the game model, a dynamic interaction process model of the driver, vehicle and environment is constructed. The dynamic interaction process model of the driver, vehicle and environment adopts the HJB equation. By solving the HJB equation, collaborative optimization between participants is achieved in a dynamic environment, so that the participants can obtain optimal decision-making driving data in complex driving scenarios. Define the game model as , in is the set of game participants, and Participants in the game The state space and action space of For participants The game participants include the driver, the vehicle, and the environment; It represents the dynamic interaction between game participants. It is used to refer to the interaction and decision-making process between participants (such as drivers, vehicles, and the environment) in the stochastic differential game model. The game model can systematically analyze the behaviors and decisions of participants in different situations by describing their states, actions, and utility functions. In the dynamic interaction process model of driver-vehicle-environment, the state space and action space of the participants will affect their decision results, and these decisions will feedback and affect the states of other participants, thus forming a dynamic interaction process. Therefore, The complex interdependencies and dynamics are described, reflecting the synergy between the driver, vehicle, and environment.

[0046] The game model is described as the following stochastic differential equation, namely the HJB equation: , in, For participants At the moment The state variables represent the current state of the driver, vehicle or environment, including the driver's heart rate and other physiological states, the vehicle's speed, the traffic conditions of the environment, etc. For participants At the moment The action variable represents the actions that the participant can take, including the driver's acceleration, braking operation, vehicle control strategy, etc.; For other participants at the moment The action variable represents the action of the participant The actions of all other participants will affect the participants Decisions and status; Indicates participants The state transition function describes how the state changes over time given the current state and action; Indicates participants Noise influence function, which describes the impact of random factors (such as environmental changes, driver reactions, etc.) on the system during state transition; It is Brownian motion, represents random perturbations, and is used to model randomness and uncertainty in a system.

[0047] Through the above parameters, stochastic differential equations can dynamically describe the interaction and decision-making process between the driver, vehicle and environment, and adapt to different driving environments and driver behaviors in real time, thereby improving driving safety and the overall performance of the system.

[0048] In the game model, the states and actions of the participants are interconnected through stochastic differential equations, which can effectively capture the complex dynamic interactions between the driver, the vehicle and the environment, provide a basis for subsequent decision optimization and intelligent assistance, and enable more reasonable responses in various driving scenarios.

[0049] The Hamilton-Jacobi-Bellman (HJB) equation describes the optimal decision-making problem of game participants under Nash equilibrium. The game equilibrium solution is obtained by solving the HJB equation. , that is, the optimal decision-making strategy of all participants. There is the following relationship between the game model and the previous stochastic differential equation: Stochastic differential equations are used to describe the state changes of game participants in a dynamic environment and reflect the state of the participants. and actions The HJB equation provides a way to optimize these dynamic decisions, ensuring that the actions chosen by participants in a given state maximize their benefits or utility.

[0050] The HJB equation is obtained by defining the value function , indicating that in the state and time Next, participants The maximum expected return that can be obtained, the solution of the value function is actually to find the optimal strategy, that is, how to choose the best action in the dynamic environment described by the stochastic differential equation. To cope with the changing state, the function in the stochastic differential equation and It provides a mathematical basis for the state transition and random disturbance of participants, so that the solution of the HJB equation can reflect the optimal decision in an uncertain environment.

[0051] The solution to the HJB equation is the optimal decision-making strategy of all participants. The decision-making strategies are interdependent under Nash equilibrium conditions. The optimal decision of each participant depends not only on its own state and action, but also on the actions of other participants. By solving the HJB equation, we can find out how all participants in the game described by the stochastic differential equation choose their optimal actions while considering each other's strategies, thereby improving the overall system efficiency.

[0052] The HJB equation is closely related to the stochastic differential equation. The former provides a framework for optimal decision-making for the latter. By solving the HJB equation, collaborative optimization among participants can be achieved in a dynamic environment, ensuring that optimal decisions are made in complex driving scenarios, thereby improving driving safety and the overall performance of the system. The solution process is: , in, For participants The value function of and time Next, participants The maximum expected benefit that can be obtained. The value function is the core of the optimal decision and reflects the participants' evaluation of future benefits; For participants The state variable represents the participant at time The current state of the vehicle, including the driver’s physiological state, the vehicle’s speed, etc. For participants The action variable represents the actions that the participant can take, including acceleration, deceleration, or turning; For participants The action space includes all possible action choices for the participant; is the optimal action of other participants, indicating the best strategy adopted by other participants in the current state; is the state transition function, describing the and actions and the best actions of other participants When the participants How the status of changes; is the value function for the state variable The gradient of , which indicates the impact of state changes on expected returns; is the noise influence function, which describes the random factors affecting the participants during the state transition process. The impact of Represents the trace operation of a matrix, usually used to calculate the sum of variances or covariances; For participants The utility function is expressed in the state and actions and the best actions of other participants Bottom, the participants’ immediate payoffs, the superscript T indicates transposition; The parameters and utility function of the game model can be designed according to the characteristics of the driving scenario and the optimization objectives. For example, in the highway scenario, the vehicle's state variables may include position, speed, etc., and the environment's state variables may include traffic density, weather conditions, etc. The driver's utility function can comprehensively consider factors such as driving safety, comfort, and fuel consumption; in the urban road scenario, the vehicle's state variables may also include lanes, traffic light status, etc., and the environment's state variables may also include pedestrians, non-motor vehicles, etc. The driver's utility function can further consider factors such as traffic efficiency and road right priority.

[0053] Step seven: Based on the optimal decision-making driving data obtained in step six, the deep reinforcement learning algorithm is used to dynamically adjust the on-board computing, storage and communication resources according to the real-time road and network conditions to minimize energy consumption and task delays.

[0054] Among them, the definitions of on-board computing, storage, and communication resources are as follows: On-board computing resources: refers to the hardware facilities inside the vehicle used for data processing, including the central processing unit (CPU), graphics processing unit (GPU) and memory. On-board computing resources are used to analyze sensor data in real time and execute intelligent algorithms.

[0055] Storage resources: refers to devices used to store data, including hard disks, solid-state drives, and memory. Storage devices are used to save information such as data collected by sensors, driving records, and model parameters; Communication resources: refers to the data transmission capability between the vehicle and the external environment (such as other vehicles, infrastructure, and cloud servers), including wireless communication modules (such as 5G, Wi-Fi, etc.) to achieve real-time data exchange and information sharing.

[0056] Define the reinforcement learning model as ,in: state : It includes the system’s queue status, network status, and data collected by sensors. The status reflects the current system operation status and environmental changes; action Refers to resource scheduling decisions, including how to allocate onboard computing, storage, and communication resources to optimize system performance; State transition probability : Describes the probability of the system transferring to the next state after taking a certain action in the current state, which is closely related to the changes in sensor data and the system response; Reward Function : Used to evaluate the performance of the agent in a specific state and action, and take decisions that are conducive to reducing energy consumption and delay; Discount Factor : Used to balance the importance of current rewards and future rewards to ensure that the agent makes the best decision in the long run.

[0057] In this invention, an agent refers to an entity that makes decisions in a dynamic environment. It continuously learns and optimizes resource scheduling strategies through interaction with the environment. The objective function of the agent is to maximize the expected value of long-term cumulative rewards: , , in, is the scheduling strategy, Delayed task, The maximum constraint value for the delay.

[0058] Reward Function Designed for: , in, and are weight coefficient 1 and weight coefficient 2, which are used to balance the impact of energy consumption and delay.

[0059] Deep reinforcement learning algorithms include DDPG (Deep Deterministic Policy Gradient) and SAC (Soft Actor-Critic). By continuously interacting with the environment and optimizing the policy network, the intelligent agent gradually learns the optimal resource scheduling plan.

[0060] To verify the effectiveness of the method of the present invention, experiments were conducted in three typical scenarios: highways, urban roads, and mountain roads. The method was compared with the traditional method. The experimental results are shown in Table 1: Table 1 Comparison of experimental results

[0061] It can be seen that the method of the present invention has achieved significant performance improvement in various scenarios. In the highway scenario, the average speed was increased by 13.6%, the fuel consumption was reduced by 11.1%, and the collision rate was reduced by 62.2%; in the urban road scenario, the average speed was increased by 13.2%, the number of stops was reduced by 28.0%, and the traffic efficiency was improved by 9.8%; in the mountain road scenario, the average speed was increased by 15.0%, the lateral acceleration was reduced by 28.6%, and the passenger comfort was improved by 20.0%, which shows the effectiveness and superiority of the method of the present invention in assisting the driver state.

[0062] It can be seen that the method of the present invention is superior to the traditional method in all indicators, which can illustrate the effectiveness of the method of the present invention. Especially in complex mountain road scenes, the method of the present invention has particularly obvious improvements in state estimation accuracy and auxiliary efficiency, while energy consumption is greatly reduced, demonstrating extremely strong environmental adaptability and resource utilization efficiency.

[0063] Principle of the method of the present invention: The method of the present invention adopts a variational autoencoder, and adaptively learns a low-dimensional latent space representation of the driver's state by optimizing the variational lower bound, which effectively reduces the data dimension and reduces the storage and computing costs. At the same time, the latent space representation can reveal the intrinsic structure and correlation of the data, providing more accurate and semantic features for subsequent state prediction and decision optimization.

[0064] Based on the federated learning framework, while protecting the driver's privacy, the model collaborative optimization between vehicles is achieved, so that the intelligent assistance system can quickly adapt to the individual differences and preferences of different drivers. At the same time, through the incremental learning module deployed on the vehicle edge device, the model can be updated and optimized in real time, respond quickly to changes in the driver's state, and provide more timely and targeted assistance measures.

[0065] The present invention provides a collaborative optimization mechanism for the integration of driver, vehicle and environment, introduces a stochastic differential game model, and systematically models the dynamic interaction and decision-making process among the driver, vehicle and environment from the perspective of multiple intelligent agents. By solving the game equilibrium, it achieves a balance of interests among all parties and improves the overall system efficiency, actively adapts to the uncertain changes in the environment, and enhances the system's robustness and ability to cope with complex scenarios.

[0066] Aiming at the dynamic processes of multiple time scales, such as physiology and behavior, involved in the driver status monitoring system, a distributed control framework based on singular perturbation theory is proposed to decouple the system into fast and slow subsystems, and design corresponding controllers for coordinated optimization, thus simplifying the complexity of the control problem. At the same time, multi-agent reinforcement learning is introduced, so that each subsystem can continuously optimize its own control strategy through autonomous learning, and realize adaptive collaborative control of the global system through cross-layer communication, thus realizing hierarchical decoupling and multi-time scale collaborative control of the Internet of Vehicles system.

[0067] Through driver status prediction and personalized intelligent assistance strategies, the risk of traffic accidents caused by fatigue driving and distraction is effectively reduced, and the psychological pressure and physiological burden of drivers are reduced. At the same time, intelligent adjustment of vehicle parameters and interaction methods improves the driving experience.

[0068] Example 2 A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0069] Example 3 A computer-readable storage medium stores a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0070] Example 4 A computer program product comprises a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0071] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0072] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0074] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for monitoring driving status of an Internet of Vehicles, characterized in that: The following steps are involved: Step 1: multiple sensors are installed in the vehicle to collect driving status data in real time, wherein the driving status data includes driver behavior data and environmental data; Step 2: Preprocess the collected multi-source heterogeneous data at the vehicle-mounted edge computing node. The preprocessing includes data cleaning, data normalization, and data fusion. Step 3: The fused data The input is fed into a variational autoencoder, which infers the latent variable distribution by optimizing the variational lower bound when extracting the low-dimensional latent variable representation of the driving state data. ; Step 4: Infer the latent variables Input into the fractional differential equation model, the fractional differential equation model describes the dynamic change law of the driving state, and obtains the predicted value of the driving state by solving the fractional differential equation , reflecting the driver’s physiological and behavioral state; Step 5: Federated learning is performed on multiple vehicle-mounted edge nodes. Through encrypted gradient aggregation and differential privacy algorithms, collaborative optimization of the global model is achieved while protecting data privacy. Step six, based on the game model, construct a dynamic interaction process model of driver-vehicle-environment. The dynamic interaction process model of driver-vehicle-environment adopts the HJB equation. By solving the HJB equation, collaborative optimization between participants is achieved in a dynamic environment, so that participants can obtain optimal decision-making driving data in complex driving scenarios.

2. A method for monitoring driving status of an Internet of Vehicles according to claim 1, characterized in that: Also includes: Step seven: Based on the optimal decision-making driving data obtained in step six, the deep reinforcement learning algorithm is used to dynamically adjust the on-board computing, storage and communication resources according to the real-time road and network conditions to minimize energy consumption and task delays.

3. A method for monitoring driving status of an Internet of Vehicles according to claim 1 or 2, characterized in that: In step 2, during the data fusion process, the weights are dynamically adjusted by solving the following multi-objective optimization problem: ， ， ， in, represents the mean square error, For data The reconstruction value of and are respectively minimizing the reconstruction error and the L2 norm of the weight, and the adaptive weight ; The data type.

4. The method for monitoring driving status of an Internet of Vehicles according to claim 3, characterized in that: In step three, the variational autoencoder is represented as: ， in, is the parameterized probability distribution of the decoder, is the prior distribution, are the parameters to be learned, is the divergence, which measures the difference between two probability distributions; is a hidden variable; is the reconstruction error.

5. The method for monitoring driving status of an Internet of Vehicles according to claim 4, characterized in that: In step 4, the fractional order differential equation is: ， ， in, The order is The Caputo fractional derivative of is defined as: ， ， in, is the Gamma function, is a positive integer, when When , the Caputo derivative degenerates into the classical first-order derivative; is the integral variable, representing a point in time in the past; is a driving state variable, the driving state variable; is a nonlinear function, is the process noise.

6. A method for monitoring driving status of an Internet of Vehicles according to claim 5, characterized in that: In step 5, the optimization process of federated learning is: ， ， ， in, For the The loss function calculated by each node based on local data, is the corresponding gradient, is the total number of nodes, in In the iteration, each node first calculates the local gradient , and then send the gradient to the central node through a secure encrypted communication protocol; the central node aggregates the received gradients and adds -Gaussian noise for differential privacy , get the global gradient ; Finally, the central node uses the global gradient to update the model parameters , and broadcast the updated model to each node to complete one iteration of federated learning.

7. A method for monitoring driving status of an Internet of Vehicles according to claim 6, characterized in that: In step 6, the game model is defined as ; in is the set of game participants, and Participants in the game The state space and action space of For participants The game participants include the driver, the vehicle, and the environment; It represents the dynamic interaction between game participants; The game model is described as the following stochastic differential equation, namely the HJB equation: ， in, For participants At the moment The state variables represent the current state of the driver, vehicle or environment; For participants At the moment The action variable represents the actions that the participant can take; For other participants at the moment The action variable represents the action of the participant The actions of all other participants will affect the participants Decisions and status; Indicates participants The state transition function describes how the state changes over time given the current state and action; Indicates participants Noise influence function; is Brownian motion, which represents random disturbance.

8. The method for monitoring driving status of an Internet of Vehicles according to claim 7, characterized in that: In step six, the solution process is: ， in, For participants The value function of and time Next, participants The maximum expected return that can be obtained; For participants The state variable represents the participant at time The current status of For participants The action variable represents the actions that the participants can take; For participants The action space includes all possible action choices of the participants; is the optimal action of other participants, indicating the best strategy adopted by other participants in the current state; is the state transition function, describing the and actions and the best actions of other participants When the participants How the status of changes; is the value function for the state variable The gradient of , which indicates the impact of state changes on expected returns; is the noise influence function, which describes the random factors affecting the participants during the state transition process. The impact of Represents the trace operation of a matrix, usually used to calculate the sum of variances or covariances; For participants The utility function is expressed in the state and actions and the best actions of other participants Below, the participants’ immediate payoffs. The superscript T indicates transposition.

9. The method for monitoring driving status of an Internet of Vehicles according to claim 8, characterized in that: In step 7, the reinforcement learning model is defined as ,in: state , including the system’s queue status, network status, and data collected by sensors. The status reflects the current system operation and environmental changes; action Refers to resource scheduling decisions, including allocating onboard computing, storage, and communication resources; State transition probability : Describes the probability of transitioning to the next state after taking a certain action in the current state; Reward Function : Used to evaluate the performance of the agent in specific states and actions; Discount Factor : Used to balance the importance of current rewards and future rewards; The objective function is to maximize the expected value of the long-term cumulative reward: ， in, is the scheduling strategy, Delayed task, is the maximum constraint value of delay; Reward Function for: ， in, and are weight coefficient 1 and weight coefficient 2, which are used to balance the impact of energy consumption and delay.

10. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images