A multi-parameter monitoring method and system for Internet of Vehicles based on 5G multi-access edge computing

Abstract

The present invention discloses a multi-parameter monitoring method and system of Internet of Vehicles based on 5G multi-access edge computing, including: using vehicle-mounted sensors to collect multi-source heterogeneous data in real time; using 5G network to transmit the collected multi-source heterogeneous data to the nearest edge computing node in real time; pre-processing the multi-source heterogeneous data at the edge computing node; inputting the fused data into the local model deployed on the edge node to perform real-time monitoring and real-time evaluation of the vehicle state and the driver state; uploading the obtained evaluation results to the cloud platform, and the cloud platform dynamically adjusts the driving assistance decision and control instructions according to the vehicle driving conditions and environmental data to generate the optimal decision result; and sending the optimal decision result to the vehicle-mounted execution unit through the 5G network. The method of the present invention realizes the monitoring of key parameters such as vehicle speed and driver physiological state, and improves the real-time and accuracy of the monitoring system.

Description

Technical Field

[0001] The present invention relates to a vehicle networking multi-parameter monitoring method and system based on 5G multi-access edge computing, and belongs to the technical field of vehicle networking. Background Art

[0002] With the rapid development of Internet of Vehicles and autonomous driving technologies, improving vehicle safety and driving experience is crucial for intelligent transportation systems. Traditional Internet of Vehicles systems mainly rely on centralized data processing models, which have problems such as high data transmission delays, serious information bottlenecks, and insufficient real-time response capabilities. In addition, existing data processing methods are difficult to effectively deal with massive, heterogeneous, and high-dimensional sensor data in Internet of Vehicles scenarios, resulting in limited data utilization and decision-making accuracy.

[0003] The emergence of 5G communication technology provides a basis for solving the above problems. 5G networks have the characteristics of high bandwidth, low latency, and large connections, which enable the data generated by vehicles to be quickly transmitted to the edge of the network and processed in real time. Multi-access edge computing (MEC) deploys computing, storage and other resources at the edge of the network close to users, which can significantly reduce data transmission latency and improve service quality. However, how to achieve efficient integration and processing of Internet of Vehicles data in a multi-access edge computing environment and give full play to the technical advantages of 5G and MEC is still a problem that needs to be solved. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in the process of multi-parameter monitoring of the Internet of Vehicles, how to use 5G and multi-access edge computing to achieve real-time, accurate and safe monitoring and analysis of vehicle and driver status.

[0005] In order to solve the above technical problems, the present invention provides a multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, comprising the following steps:

[0006] Step 1: Use vehicle-mounted sensors to collect multi-source heterogeneous data in real time, including vehicle operating parameters, driver physiological and behavioral data;

[0007] Step 2: Use the 5G network to transmit the collected multi-source heterogeneous data to the nearest edge computing node in real time;

[0008] Step 3: Preprocess multi-source heterogeneous data at the edge computing node, including data fusion;

[0009] Step 4: Input the fused data into the local model deployed on the edge node. The local model is a deep learning model. The deep learning model uses the spatiotemporal attention mechanism and the long short-term memory network to extract the spatiotemporal features in the input data, perform real-time monitoring and real-time evaluation of the vehicle status and the driver status, and obtain the evaluation results. At the same time, the federated learning algorithm is used to train the local models on multiple vehicle-mounted edge nodes to achieve collaborative optimization among multiple nodes.

[0010] Step 5: Upload the evaluation results obtained in step 4 to the cloud platform. The cloud platform makes global decisions and optimizations based on the evaluation results uploaded by all edge nodes, and generates driving assistance decisions and control instructions. The cloud platform then dynamically adjusts the driving assistance decisions and control instructions based on the vehicle's driving conditions and environmental data to generate the optimal decision results.

[0011] Step six: Send the optimal decision result to the on-board execution unit through the 5G network, and control the vehicle safely or issue an alarm to the driver based on the optimal decision result.

[0012] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step three, data fusion includes the following steps:

[0013] 1) Set the initial state estimate and the initial covariance matrix , define the state transfer function and the observation function , initialize the particle set ,in For the Particles, The corresponding weight is is the total number of particles;

[0014] 2) Based on the state estimate and state transfer function at the previous moment, predict the prior state estimate at the current moment and the prior covariance matrix , and update the particle collection:

[0015] ,

[0016] ,

[0017] ,

[0018] in, is the state transfer matrix, is the process noise covariance matrix; Represents the probability density function, which gives the probability distribution of the current state;

[0019] 3) According to the current observation value , calculate the Kalman gain , and update the posterior state estimate and the posterior covariance matrix :

[0020] ,

[0021] ,

[0022] ,

[0023] in, is the observation matrix, is the observation noise covariance matrix;

[0024] 4) According to the observation value and observation function at the current moment, calculate the likelihood probability of each particle and update the corresponding weight:

[0025] ,

[0026] ,

[0027] in, is the normalized particle weight; is a proportional constant, which is used to express the proportional relationship of weight update, and the specific weight value is finally determined through the normalization step;

[0028] 5) According to the particle weights, resample the particle set to eliminate the weight degradation problem;

[0029] 6) According to the updated particle set in 5), calculate the estimated state value at the current moment:

[0030] ,

[0031] 7) Adaptively adjust the process noise covariance matrix of the Kalman filter according to the residual statistical characteristics of the fusion result and the observation noise covariance matrix ;

[0032] 8) Repeat 2) to 7) until the data processing is complete.

[0033] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step 4, the process of training the local model includes:

[0034] s1) During the encrypted gradient aggregation process, each edge node calculates the gradient of the model parameters locally , and is sent to the central node for aggregation in an encrypted manner:

[0035] ,

[0036] in, is the number of nodes, To satisfy - Gaussian noise for differential privacy, and Control the strength and probability of privacy protection respectively;

[0037] s2) The central node uses the aggregated gradient to update the global model parameters , and broadcast the updated model to each node to complete a round of federated learning iteration;

[0038] Represents the aggregated gradient, which is the average of the gradients calculated by all edge nodes plus the noise term; Indicates The model parameter gradients calculated locally by edge nodes; is the number of edge nodes;

[0039] ,

[0040] in, is the learning rate; through multiple rounds of iterations, the federated learning algorithm is used to achieve collaborative optimization of local models among vehicle nodes without directly sharing the original data;

[0041] s3) Repeat s1)~s2) until all models converge or the preset number of iterations is reached.

[0042] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step 4, during the federated learning process, the quantized gradient and the sparse gradient Respectively expressed as:

[0043] ,

[0044] ,

[0045] in, is the quantization bit width, is the thinning threshold.

[0046] In the aforementioned multi-parameter monitoring method of the Internet of Vehicles based on 5G multi-access edge computing, in step 5, during the dynamic adjustment process, a stochastic differential game model is used to describe the dynamic interaction of the driver, vehicle driving conditions, and road traffic conditions, and the optimal decision result of each participant is obtained by solving the Nash equilibrium;

[0047] The game model is defined as ,in, is the set of game participants, and Participants The state space and action space of For participants The utility function represents the benefits obtained in the game process, and the game participants include the driver, the vehicle, and the environment; Represents an ordered set or tuple;

[0048] The game process is described as the following stochastic differential equation:

[0049] ,

[0050] in, and Participants exist The state and action of each moment, For the actions of other participants, For Brownian motion, and are drift function and diffusion function, respectively, which describe the changing rules of the participant’s state. For participants At the moment status changes.

[0051] The aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing includes, in step six:

[0052] ss1) The optimal decision result is transmitted to the vehicle system via the 5G network;

[0053] ss2) Use deep reinforcement learning algorithms to dynamically adjust vehicle computing resources, storage resources, and communication resources, and define the reinforcement learning model as , where: state space Including system queue status, network status; action space Represents resource scheduling decision; state transition probability Describes the probability of transitioning from one state to another; reward function Used to evaluate the effect of each action; discount factor Used to weigh the importance of future rewards;

[0054] The objective function of the deep reinforcement learning algorithm is to maximize the expectation of long-term cumulative rewards:

[0055] ,

[0056] in, Consider task delay and energy consumption The reward function is and are weight coefficient one and weight coefficient two respectively, and is the corresponding constraint threshold;

[0057] ss3) uses the optimization method based on dual gradient descent to solve the objective function in ss2) and transform the original problem into an unconstrained Lagrangian function maximization problem:

[0058] ,

[0059] in, is the Lagrangian function, through alternating optimization strategy network and the Lagrange multiplier , Lagrange multiplier 2 , and gradually converge to the optimal strategy that satisfies the constraints.

[0060] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step six, the dynamic adjustment includes:

[0061] Update policy network parameters:

[0062] ,

[0063] in, are policy network parameters, is the learning rate, Express Find the gradient;

[0064] Update the Lagrange multipliers:

[0065] ,

[0066] ,

[0067] in, and are the maximum allowed values ​​of task delay and energy consumption, respectively. is the learning rate.

[0068] A computer device / equipment / system, comprising 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-mentioned vehicle network multi-parameter monitoring method based on 5G multi-access edge computing.

[0069] A computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the above-mentioned vehicle network multi-parameter monitoring method based on 5G multi-access edge computing.

[0070] A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned vehicle network multi-parameter monitoring method based on 5G multi-access edge computing.

[0071] The beneficial effects achieved by the present invention are as follows: the multi-parameter monitoring method of the Internet of Vehicles based on 5G multi-access edge computing of the present invention utilizes the vehicle-mounted multi-source sensor data, combined with algorithms such as adaptive Kalman filtering and particle filtering, to realize the monitoring of key parameters such as vehicle speed and driver physiological state, thereby improving the real-time and accuracy of the monitoring system. The present invention utilizes the eMBB enhanced mobile broadband capability of the 5G network, reduces the transmission delay of vehicle-mounted data, and realizes localized processing of data through multi-access edge computing, thereby reducing the burden on the central cloud; at the same time, the use of federated learning, gradient quantization and sparsification methods further improves the communication and computing efficiency, thereby reducing energy consumption and costs while meeting real-time constraints.

[0072] In the method of the present invention, the improved multi-sensor fusion algorithm reduces the root mean square error of vehicle speed estimation to 0.09 m / s, which is significantly lower than the error generated by a single sensor.

[0073] The present invention utilizes a federated learning algorithm. The on-board nodes do not need to upload original data. Instead, collaborative modeling is achieved through encrypted gradient aggregation, which effectively protects the driver's privacy. At the same time, a differential privacy mechanism is introduced to add noise to the aggregated gradient to prevent the leakage of sensitive information.

[0074] Based on the stochastic differential game model, the present invention realizes the dynamic interaction optimization among the driver, the vehicle and the environment, and obtains a robust decision-making control strategy. In typical scenarios such as highways and urban roads, the safety indicators such as the collision rate and traffic violation rate of the present invention are reduced by more than 70% compared with traditional methods, and the fuel efficiency is improved by more than 20%.

[0075] In response to the stringent requirements of Internet of Vehicles applications for low latency and high reliability, this paper adopts deep reinforcement learning to minimize system energy consumption while meeting service quality constraints. Simulation results show that the proposed intelligent scheduling method reduces the average task latency by 59%, the service timeout rate by 89%, and the average energy consumption by 45%. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the multi-parameter monitoring method of the Internet of Vehicles based on 5G multi-access edge computing of the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0078] Example 1

[0079] like Figure 1 As shown, this embodiment provides a multi-parameter monitoring method for Internet of Vehicles based on 5G multi-access edge computing (MEC), including the following steps:

[0080] Step 1: Use vehicle-mounted sensors to collect multi-source heterogeneous data in real time, including vehicle operating parameters, driver physiological and behavioral data;

[0081] Various types of sensors are deployed in the vehicle, including vehicle speed sensors, acceleration sensors, steering wheel angle sensors, brake pedal pressure sensors, GPS positioning modules, heart rate sensors, breathing sensors, skin electricity sensors, infrared cameras, temperature and humidity sensors, light sensors, etc. These sensors collect relevant data during the vehicle's driving process in an all-round and multi-dimensional manner.

[0082] By speed For example, suppose at time The collected vehicle speed data is , then the acquisition frequency Down-continuous sampling The vehicle speed sequence of a data window is expressed as:

[0083] ,

[0084] in, is the initial sampling time.

[0085] Similarly, data collected by other sensors can also be represented by corresponding time series, such as acceleration series , steering wheel angle sequence , driver's heart rate sequence Etc., the above sequence data reflects the real-time status of the vehicle and the driver.

[0086] At the same time, the environmental condition sensor collects background information about the vehicle's environment, including temperature ,humidity , light intensity The dynamic changes of environmental factors have an important impact on vehicle driving and driver status, so they are included in the scope of data analysis, as shown in Table 1.

[0087] Table 1 Data collected by some sensors

[0088]

[0089] Step 2: Use the 5G network to transmit the collected multi-source heterogeneous data to the nearest edge computing node in real time;

[0090] The vehicle terminal establishes a high-speed, low-latency connection with the adjacent edge computing node through the 5G communication module. Assume that the transmission delay between the vehicle terminal and the edge node is , the packet size is , then the throughput of the 5G link satisfy:

[0091] ,

[0092] The 5G network's eMBB enhances mobile broadband capabilities, ensures real-time data transmission, and can provide Gbps-level transmission rates, significantly reducing data transmission latency. At the same time, the 5G network also supports mMTC massive machine-type communications, which can connect a large number of vehicles at the same time to meet the communication needs of Internet of Vehicles applications.

[0093] Assume that the distance between the vehicle and the edge node is , the frequency of the 5G link is , the transmission power is , the noise power spectral density is , then according to Shannon's theorem, the theoretical upper limit of the link transmission rate is for:

[0094] ,

[0095] in, is the channel bandwidth, is the path loss exponent.

[0096] Since vehicle movement can cause fast fading of wireless channels, OFDM orthogonal frequency division multiplexing, MIMO multi-antenna and other technologies are used to improve spectrum efficiency, and channel coding schemes such as Polar and LDPC are combined to ensure transmission reliability. and are the number of transmitting antennas and receiving antennas respectively, then the MIMO channel capacity for:

[0097] ,

[0098] By rationally designing the parameters of the 5G communication system, including power allocation, resource scheduling, beamforming, etc., stable, reliable, low-latency and high-speed data transmission can be achieved in the Internet of Vehicles scenario, providing data protection for edge computing.

[0099] Step three: pre-process multi-source heterogeneous data at the edge computing node, including data cleaning, feature extraction, and data fusion.

[0100] This embodiment uses an improved multi-sensor data fusion algorithm to fuse data to improve data quality and consistency. The improved multi-sensor data fusion algorithm is based on a combination of adaptive Kalman filtering and particle filtering, which combines the advantages of the two algorithms and can effectively process nonlinear, non-Gaussian complex systems. Data fusion includes the following steps:

[0101] 1) Initialization: Set the initial state estimate and the initial covariance matrix , define the state transfer function and the observation function , initialize the particle set ,in For the Particles, The corresponding weight is is the total number of particles;

[0102] 2) Prediction stage: Based on the state estimate and state transfer function at the previous moment, predict the prior state estimate at the current moment and the prior covariance matrix , and update the particle collection:

[0103] ,

[0104] ,

[0105] ,

[0106] in, is the state transfer matrix, is the process noise covariance matrix; Represents the probability density function, which gives the probability distribution of the current state.

[0107] 3) Update phase: based on the current observation value , calculate the Kalman gain , and update the posterior state estimate and the posterior covariance matrix :

[0108] ,

[0109] ,

[0110] ,

[0111] in, is the observation matrix, is the observation noise covariance matrix;

[0112] 4) Particle weight update: According to the observation value and observation function at the current moment, the likelihood probability of each particle is calculated and the corresponding weight is updated:

[0113] ,

[0114] ,

[0115] in, is the normalized particle weight; is a proportional constant, which is used to express the proportional relationship of weight update, and the specific weight value is finally determined through the normalization step;

[0116] 5) Resampling: Resample the particle set according to the particle weight to eliminate the weight degradation problem. Commonly used resampling methods include polynomial resampling, system resampling, etc.

[0117] 6) State estimation: Based on the updated particle set in 5), calculate the state estimation value at the current moment:

[0118] ,

[0119] 7) Adaptive adjustment: Adaptively adjust the process noise covariance matrix of the Kalman filter according to the residual statistical characteristics of the fusion result and the observation noise covariance matrix , to adapt to the changes of fusion system and observation noise, the estimation of noise covariance can be achieved through the expectation maximization (EM) algorithm;

[0120] 8) Iteration: Repeat 2) to 7) until the data processing is complete.

[0121] By combining adaptive Kalman filtering and particle filtering, it is possible to accurately achieve the fusion of data obtained by multiple sensors in nonlinear, non-Gaussian complex systems. The adaptive mechanism enables the algorithm to track changes in system and observation noise and dynamically adjust filter parameters to obtain more accurate and reliable fusion result data.

[0122] Table 2 shows the simulation results of vehicle speed estimation using the fusion method of the present invention. The fused estimated value is very close to the true value, and the root mean square error (RMSE) is only 0.09 m / s, which is better than the estimation performance of a single sensor, proving the effectiveness of the algorithm.

[0123] Table 2 Simulation results of vehicle speed estimation

[0124]

[0125] Step 4: Input the fused data into the local model deployed on the edge node. The local model is a deep learning model. The deep learning model uses the spatiotemporal attention mechanism and the long short-term memory network to extract the spatiotemporal features in the input data, and conduct real-time monitoring and real-time evaluation of the vehicle status and the driver status to obtain the evaluation results. At the same time, the federated learning algorithm is used to train the local models on multiple vehicle-mounted edge nodes to achieve collaborative optimization among multiple nodes while protecting data privacy. The vehicle status includes speed, acceleration, and direction, and the driver status includes attention and fatigue level.

[0126] The process of training a local model includes encrypted gradient aggregation and differential privacy steps:

[0127] s1) During the encrypted gradient aggregation process, each edge node calculates the gradient of the model parameters locally , and is sent to the central node for aggregation in an encrypted manner:

[0128] ,

[0129] in, is the number of nodes, To satisfy - Gaussian noise for differential privacy, and Control the strength and probability of privacy protection respectively;

[0130] s2) The central node uses the aggregated gradient to update the global model parameters , and broadcast the updated model to each node to complete a round of federated learning iteration;

[0131] Represents the aggregated gradient, which is the average of the gradients calculated by all edge nodes plus the noise term;

[0132] Indicates The model parameter gradients calculated locally by edge nodes;

[0133] is the number of edge nodes;

[0134] express is from a mean of 0 and a covariance of The noise is sampled from a Gaussian distribution;

[0135] is the variance of Gaussian noise, controlling the amplitude of the noise;

[0136] is the unit matrix, indicating that the noise is independent and identically distributed in each dimension;

[0137] ,

[0138] in, For the learning rate, through multiple rounds of iterations, the federated learning algorithm is used to achieve collaborative optimization of local models between on-board nodes without directly sharing the original data, thereby improving the prediction performance while protecting the driver's privacy;

[0139] s3) Repeat s1)~s2) until all models converge or the preset number of iterations is reached.

[0140] In the process of federated learning, the communication efficiency and computing performance of federated learning are improved, and the adaptive quantization method is used to reduce the bit width of the gradient, thereby reducing the communication cost; the sparsification method is used to screen out important gradient components for aggregation, thereby reducing the computational complexity.

[0141] Quantized gradient and the sparse gradient Respectively expressed as:

[0142] ,

[0143] ,

[0144] in, is the quantization bit width, The quantization and sparsification parameters can be dynamically adjusted according to the network status and computing resources to balance communication efficiency, computing performance and model accuracy.

[0145] Step 5: Upload the evaluation results obtained in step 4 to the cloud platform. The cloud platform makes global decisions and optimizations based on the evaluation results uploaded by all edge nodes, and generates driving assistance decisions and control instructions. The cloud platform then dynamically adjusts the driving assistance decisions and control instructions based on the vehicle driving conditions and environmental (road traffic conditions) data to generate the optimal decision results.

[0146] In the dynamic adjustment process, the stochastic differential game model is used to describe the dynamic interaction between drivers, vehicle driving conditions, and road traffic conditions. By solving the Nash equilibrium, the optimal decision results of each participant are obtained.

[0147] The game model is defined as ,

[0148] in, is the set of game participants, and Participants The state space and action space of For participants The utility function represents the benefits obtained in the game process, and the game participants include the driver, the vehicle, and the environment; Represents an ordered set or tuple, which here represents the structure and components of the game model;

[0149] The game process is described as the following stochastic differential equation:

[0150] ,

[0151] in, and Participants exist The state and action of each moment, For the actions of other participants, is Brownian motion, reflecting the uncertainty of the environment. and They are drift function and diffusion function respectively, which describe the changing rules of the participants' states. For participants At the moment status changes.

[0152] The above stochastic differential equations are used to describe the interactive behaviors of game participants (such as drivers, vehicles, and environments) in a dynamic environment. and the diffusion function It captures the deterministic and random changes of the system respectively, combined with Brownian motion , providing a probabilistic model for the state changes of the system.

[0153] The state space and action space in the game model are concretized through stochastic differential equations, which provide a mathematical description of the strategy selection and state changes of participants in an uncertain environment.

[0154] Game equilibrium refers to the strategy combination of all participants , so that when the strategies of other participants are fixed, no participant can improve his own utility by unilaterally changing the strategy. The mathematical definition of Nash equilibrium is:

[0155] , ,

[0156] in, For participants strategy, is the equilibrium strategy of other participants, For participants The strategy space, : Utility function, representing the participants Gains gained during the gaming process; For participants The optimal strategy (equilibrium strategy) is is the set of game participants.

[0157] In the Nash equilibrium state, any player When other participants have fixed strategies, they cannot unilaterally change their strategies. To improve your own effectiveness Nash equilibrium describes how participants achieve optimal returns through strategy selection in a dynamic environment, provides a tool for analyzing strategy interactions in multi-participant systems, and helps understand the decision-making process in complex environments.

[0158] A numerical solution method based on deep reinforcement learning is used to solve the Nash equilibrium of the stochastic differential game and obtain the optimal decision results for the driver, vehicle and environment.

[0159] Specifically, each participant uses a deep neural network (approximate the optimal Q function) and update the policy network through the policy gradient method Parameters , maximize the expected cumulative return:

[0160] ,

[0161] in, For strategy The induced state distribution, the parameters of the Q function Learning by minimizing the temporal difference error, It represents the distribution property of a variable, indicating that when calculating the expectation, the variable is random and its values ​​are obtained from a specific probability distribution.

[0162] ,

[0163] in, is the target value, For immediate returns, is the discount factor, is the state transition probability, is the target network parameter.

[0164] By alternately optimizing the strategy network and Q network of each participant and introducing methods such as experience replay and target network, the multi-agent deep reinforcement learning algorithm can gradually approach the Nash equilibrium solution of the stochastic differential game and obtain the optimal decision-making strategy for the driver, vehicle and environment.

[0165] Table 3 shows the simulation results of the method of the present invention in three typical scenarios: highways, urban roads, and rural roads. Compared with the existing rule-based decision-making methods, the present invention has achieved more than 10% performance improvement in key indicators such as vehicle speed, fuel efficiency, and safety, reflecting the superiority of game optimization.

[0166] Table 3 Simulation results under three typical scenarios

[0167]

[0168] Step six: Send the optimal decision result to the on-board execution unit through the 5G network, and control the vehicle safely or issue an alarm to the driver based on the optimal decision result.

[0169] The specific process is as follows:

[0170] ss1) The optimal decision result is transmitted to the vehicle system through the 5G network, enabling the vehicle to respond quickly based on real-time road and network conditions;

[0171] ss2) Use deep reinforcement learning algorithms to dynamically adjust vehicle computing resources, storage resources, and communication resources, and define the reinforcement learning model as , where: state space Including system queue status, network status, etc.; action space Represents resource scheduling decision; state transition probability Describes the probability of transitioning from one state to another; reward function Used to evaluate the effect of each action; discount factor Used to weigh the importance of future rewards.

[0172] The objective function of the deep reinforcement learning algorithm is to maximize the expectation of long-term cumulative rewards:

[0173] ,

[0174] in, Consider task delay and energy consumption The reward function is and are weight coefficients, and is the corresponding constraint threshold.

[0175] ss3) uses the optimization method based on dual gradient descent to solve the objective function in ss2) and transform the original problem into an unconstrained Lagrangian function maximization problem:

[0176] ,

[0177] in, is the Lagrangian function, through alternating optimization strategy network and the Lagrange multiplier , Lagrange multiplier 2 , can gradually converge to the optimal strategy that satisfies the constraints;

[0178] Dynamic adjustments include:

[0179] Update policy network parameters:

[0180] ,

[0181] in, are policy network parameters, is the learning rate, Express Find the gradient;

[0182] Update the Lagrange multipliers:

[0183] ,

[0184] ,

[0185] in, and are the maximum allowed values ​​of task delay and energy consumption, respectively. is the learning rate.

[0186] Through the above steps, it is possible to meet the task delay and reliability requirements while minimizing energy consumption, ensuring the safe operation and efficient performance of the vehicle in complex environments. Finally, the on-board execution unit performs safety control based on the received decision results, or sends corresponding warning information to the driver to ensure driving safety.

[0187] Table 4 is a performance comparison between the resource scheduling method of the present invention and the existing methods under different working conditions. It can be seen that the method of the present invention has obvious advantages in key indicators such as task delay, energy consumption, and service availability.

[0188] Table 4 Performance comparison under different working conditions

[0189]

[0190] 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 following method.

[0191] A multi-parameter monitoring method for Internet of Vehicles based on 5G multi-access edge computing, comprising the following steps:

[0192] Step 1. Use vehicle-mounted sensors to collect multi-source heterogeneous data in real time, including vehicle operating parameters, driver physiological and behavioral data;

[0193] Step 2: Use the 5G network to transmit the collected multi-source heterogeneous data to the nearest edge computing node in real time;

[0194] Step 3: Preprocess multi-source heterogeneous data at the edge computing node, including data cleaning, feature extraction, and data fusion;

[0195] Step 4. Input the fused data into the local model deployed on the edge node. The local model is a deep learning model. The deep learning model uses the spatiotemporal attention mechanism and the long short-term memory network to extract the spatiotemporal features in the input data, perform real-time monitoring and real-time evaluation of the vehicle status and the driver status, and obtain the evaluation results. At the same time, the federated learning algorithm is used to train the local models on multiple vehicle-mounted edge nodes to achieve collaborative optimization among multiple nodes.

[0196] Step 5. Upload the evaluation results obtained in step 4 to the cloud platform. The cloud platform makes global decisions and optimizations based on the evaluation results uploaded by all edge nodes, and generates driving assistance decisions and control instructions. The cloud platform then dynamically adjusts the driving assistance decisions and control instructions based on the vehicle's driving conditions and environmental data to generate the optimal decision results.

[0197] Step six: Send the optimal decision result to the on-board execution unit through the 5G network, and control the vehicle safely or issue an alarm to the driver based on the optimal decision result.

[0198] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step three, data fusion includes the following steps:

[0199] 1) Set the initial state estimate and the initial covariance matrix , define the state transfer function and the observation function , initialize the particle set ,in For the Particles, The corresponding weight is is the total number of particles;

[0200] 2) Based on the state estimate and state transfer function at the previous moment, predict the prior state estimate at the current moment and the prior covariance matrix , and update the particle collection:

[0201] ,

[0202] ,

[0203] ,

[0204] in, is the state transfer matrix, is the process noise covariance matrix; Represents the probability density function, which gives the probability distribution of the current state;

[0205] 3) According to the current observation value , calculate the Kalman gain , and update the posterior state estimate and the posterior covariance matrix :

[0206] ,

[0207] ,

[0208] ,

[0209] in, is the observation matrix, is the observation noise covariance matrix;

[0210] 4) According to the observation value and observation function at the current moment, calculate the likelihood probability of each particle and update the corresponding weight:

[0211] ,

[0212] ,

[0213] in, is the normalized particle weight; is a proportional constant, which is used to express the proportional relationship of weight update, and the specific weight value is finally determined through the normalization step;

[0214] 5) According to the particle weights, resample the particle set to eliminate the weight degradation problem;

[0215] 6) According to the updated particle set in 5), calculate the estimated state value at the current moment:

[0216] ,

[0217] 7) Adaptively adjust the process noise covariance matrix of the Kalman filter according to the residual statistical characteristics of the fusion result and the observation noise covariance matrix ;

[0218] 8) Repeat 2) to 7) until the data processing is complete.

[0219] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step 4, the process of training the local model includes:

[0220] s1) During the encrypted gradient aggregation process, each edge node calculates the gradient of the model parameters locally , and is sent to the central node for aggregation in an encrypted manner:

[0221] ,

[0222] in, is the number of nodes, To satisfy - Gaussian noise for differential privacy, and Control the strength and probability of privacy protection respectively;

[0223] s2) The central node uses the aggregated gradient to update the global model parameters , and broadcast the updated model to each node to complete a round of federated learning iteration;

[0224] Represents the aggregated gradient, which is the average of the gradients calculated by all edge nodes plus the noise term; Indicates The model parameter gradients calculated locally by edge nodes; is the number of edge nodes; express is from a mean of 0 and a covariance of The noise is sampled from a Gaussian distribution; is the variance of Gaussian noise, controlling the amplitude of the noise; is the identity matrix;

[0225] ,

[0226] in, is the learning rate; through multiple rounds of iterations, the federated learning algorithm is used to achieve collaborative optimization of local models among vehicle nodes without directly sharing the original data;

[0227] s3) Repeat s1)~s2) until all models converge or the preset number of iterations is reached.

[0228] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step 4, during the federated learning process, the quantized gradient and the sparse gradient Respectively expressed as:

[0229] ,

[0230] ,

[0231] in, is the quantization bit width, is the thinning threshold.

[0232] In the aforementioned multi-parameter monitoring method of the Internet of Vehicles based on 5G multi-access edge computing, in step 5, during the dynamic adjustment process, a stochastic differential game model is used to describe the dynamic interaction of the driver, vehicle driving conditions, and road traffic conditions, and the optimal decision result of each participant is obtained by solving the Nash equilibrium;

[0233] The game model is defined as ,in is the set of game participants, and Participants The state space and action space of For participants The utility function represents the benefits obtained in the game process, and the game participants include the driver, the vehicle, and the environment; Represents an ordered set or tuple, which here represents the structure and components of the game model;

[0234] The game process is described as the following stochastic differential equation:

[0235] ,

[0236] in, and Participants exist The state and action of each moment, For the actions of other participants, is Brownian motion, reflecting the uncertainty of the environment. and They are drift function and diffusion function respectively, which describe the changing rules of the participants' states. For participants At the moment status changes.

[0237] The aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing includes, in step six:

[0238] ss1) The optimal decision result is transmitted to the vehicle system via the 5G network;

[0239] ss2) Use deep reinforcement learning algorithms to dynamically adjust vehicle computing resources, storage resources, and communication resources, and define the reinforcement learning model as , where: state space Including system queue status, network status, etc.; action space Represents resource scheduling decision; state transition probability Describes the probability of transitioning from one state to another; reward function Used to evaluate the effect of each action; discount factor Used to weigh the importance of future rewards.

[0240] The objective function of the deep reinforcement learning algorithm is to maximize the expectation of long-term cumulative rewards:

[0241] ,

[0242] in, Consider task delay and energy consumption The reward function is and are weight coefficient one and weight coefficient two respectively, and is the corresponding constraint threshold;

[0243] ss3) uses the optimization method based on dual gradient descent to solve the objective function in ss2) and transform the original problem into an unconstrained Lagrangian function maximization problem:

[0244] ,

[0245] in, is the Lagrangian function, through alternating optimization strategy network and the Lagrange multiplier , Lagrange multiplier 2 , which can gradually converge to the optimal strategy that satisfies the constraints.

[0246] In the aforementioned multi-parameter monitoring method of Internet of Vehicles based on 5G multi-access edge computing, in step six, the dynamic adjustment includes:

[0247] Update policy network parameters:

[0248] ,

[0249] in, are policy network parameters, is the learning rate, Express Find the gradient;

[0250] Update the Lagrange multipliers:

[0251] ,

[0252] ,

[0253] in, and are the maximum allowed values ​​of task delay and energy consumption, respectively. is the learning rate.

[0254] A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that the computer program / instruction implements the steps of the above method when executed by a processor.

[0255] A computer program product comprises a computer program / instruction, wherein the computer program / instruction implements the steps of the above method when executed by a processor.

[0256] This embodiment provides a multi-parameter monitoring method for the Internet of Vehicles based on 5G multi-access edge computing. By introducing advanced data fusion algorithms, deep learning models and dynamic resource management methods, it effectively solves the deficiencies of traditional Internet of Vehicles systems in data processing, state prediction, privacy protection and resource utilization. It can not only significantly improve the real-time, accuracy and safety of vehicle and driver state monitoring, but also provide new ideas and key technical support for the application of intelligent connected vehicles, and has broad application prospects.

[0257] 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.

[0258] 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.

[0259] 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 A step that specifies a function in one or more boxes.

[0260] 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 multi-parameter monitoring method for Internet of Vehicles based on 5G multi-access edge computing, characterized in that: The following steps are involved: Step 1: Use vehicle-mounted sensors to collect multi-source heterogeneous data in real time, including vehicle operating parameters, driver physiological and behavioral data; Step 2: Use the 5G network to transmit the collected multi-source heterogeneous data to the nearest edge computing node in real time; Step 3: Preprocess multi-source heterogeneous data at the edge computing node, including data fusion; Data fusion includes the following steps: 1) Set the initial state estimate And the initial covariance matrix P0, define the state transfer function f(·) and observation function h(·), initialize the particle set in is the ith particle, is its corresponding weight, N is the total number of particles; 2) Based on the state estimate and state transfer function at the previous moment, predict the prior state estimate at the current moment and the prior covariance matrix P k|k-1 , and update the particle collection: Among them, F k is the state transfer matrix, Q k is the process noise covariance matrix; p(·) represents the probability density function, which gives the probability distribution of the current state; 3) According to the current observation value z k , calculate the Kalman gain K k , and update the posterior state estimate and the posterior covariance matrix P k|k : P k|k =(I-K k H k )P k|k-1 Among them, H k is the observation matrix, R k is the observation noise covariance matrix; 4) According to the observation value and observation function at the current moment, calculate the likelihood probability of each particle and update the corresponding weight: in, is the normalized particle weight; ∝ is a proportional constant, which is used to express the proportional relationship of weight update. The specific weight value is finally determined through the normalization step; 5) According to the particle weights, resample the particle set to eliminate the weight degradation problem; 6) According to the updated particle set in 5), calculate the estimated state value at the current moment: 7) According to the residual statistical characteristics of the fusion result, the process noise covariance matrix Q of the Kalman filter is adaptively adjusted k and the observation noise covariance matrix R k ; 8) Repeat 2) to 7) until the data processing is completed; Step 4: Input the fused data into the local model deployed on the edge node. The local model is a deep learning model. The deep learning model uses the spatiotemporal attention mechanism and the long short-term memory network to extract the spatiotemporal features in the input data, perform real-time monitoring and real-time evaluation of the vehicle status and the driver status, and obtain the evaluation results. At the same time, the federated learning algorithm is used to train the local models on multiple vehicle-mounted edge nodes to achieve collaborative optimization among multiple nodes. The process of training a local model includes encrypted gradient aggregation and differential privacy steps: s1) During the encrypted gradient aggregation process, each edge node calculates the gradient of the model parameters locally And send it to the central node for aggregation in an encrypted manner: Where N is the number of nodes, n k To satisfy the Gaussian noise of (∈, δ)-differential privacy, ∈ and δ control the strength and probability of privacy protection respectively; s2) The central node uses the aggregated gradient to update the global model parameters Θ k , and broadcast the updated model to each node to complete a round of federated learning iteration; g k Represents the aggregated gradient, which is the average of the gradients calculated by all edge nodes plus the noise term; represents the model parameter gradient calculated locally by the i-th edge node; N is the number of edge nodes; Indicates n k is from a mean of 0 and a covariance of The noise is sampled from a Gaussian distribution; is the variance of Gaussian noise, controlling the amplitude of the noise; I is the unit matrix, indicating that the noise is independent and identically distributed in each dimension; I k+1 =Θ k -or k g k Among them, η k is the learning rate; s3) Repeat s1)-s2) until all models converge or the preset number of iterations is reached; In the process of federated learning, the sparse method is used to screen out important gradient components for aggregation. Quantized gradient and the sparse gradient Respectively expressed as: Among them, b k is the quantization bit width, ρ k is the thinning threshold; Step 5: Upload the evaluation results obtained in step 4 to the cloud platform. The cloud platform makes global decisions and optimizations based on the evaluation results uploaded by all edge nodes, and generates driving assistance decisions and control instructions. The cloud platform then dynamically adjusts the driving assistance decisions and control instructions based on the vehicle's driving conditions and environmental data to generate the optimal decision results. In the dynamic adjustment process, the stochastic differential game model is used to describe the dynamic interaction between drivers, vehicle driving conditions, and road traffic conditions. By solving the Nash equilibrium, the optimal decision results of each participant are obtained. The game model is defined as in, is the set of game participants, S i and A i are the state space and action space of participant i, J i is the utility function of participant i, which indicates the benefits obtained during the game. The game participants include the driver, the vehicle, and the environment; < > indicates an ordered set or tuple, which is the structure and component of the game model; The game process is described as the following stochastic differential equation: ds i (t)=f i (s i (t),a i (t),a -i (t))dt+g i (s i (t),a i (t),a -i (t))dw i (t) Among them, s i (t) and a i (t) are the state and action of participant i at time t, a -i (t) is the action of other participants, w i (t) is Brownian motion, reflecting the uncertainty of the environment, f i (·) and g i (·) are drift function and diffusion function, respectively, describing the changing law of the participant state, ds i (t) is the state change of participant i at time t; Step 6: Send the optimal decision result to the vehicle execution unit through the 5G network, and control the vehicle safely or issue a warning to the driver based on the optimal decision result. The specific process is as follows: ss1) The optimal decision result is transmitted to the vehicle system through the 5G network, enabling the vehicle to respond quickly based on real-time road and network conditions; ss2) Use deep reinforcement learning algorithms to dynamically adjust vehicle computing resources, storage resources, and communication resources, and define the reinforcement learning model as Among them: the state space S includes the system queue state and network state; the action space A represents the resource scheduling decision; the state transition probability P describes the possibility of transitioning from one state to another; the reward function Used to evaluate the effect of each action; the discount factor γ is used to weigh the importance of future rewards; The objective function of the deep reinforcement learning algorithm is to maximize the expectation of long-term cumulative rewards: Among them, r(s t , a t )=-β1D(s t , a t )-β2E(s t , a t ) is the reward function considering task delay D(·) and energy consumption E(·), β1 and β2 are weight coefficients, D * and E * is the corresponding constraint threshold; ss3) uses the dual gradient descent optimization method to solve the objective function in ss2) and transforms the original problem into an unconstrained Lagrangian function maximization problem: Among them, L(π, λ1, λ2) is the Lagrangian function. By alternately optimizing the strategy network π and the Lagrangian multiplier 1 λ1 and the Lagrangian multiplier 2 λ2, it can gradually converge to the optimal strategy that satisfies the constraints. Dynamic adjustments include: Update policy network parameters: Among them, θ is the policy network parameter, α is the learning rate, V θ It means to find the gradient of θ; Update the Lagrange multipliers: λ1←max(0,λ1-β(D * -D max )) λ2←max(0,λ2-β(E * -E max )) Among them, D max and E max are the maximum allowed values ​​of task delay and energy consumption respectively, and β is the learning rate.

2. A computer 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 vehicle network multi-parameter monitoring method based on 5G multi-access edge computing as described in claim 1.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the vehicle network multi-parameter monitoring method based on 5G multi-access edge computing described in claim 1 are implemented.

4. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the vehicle network multi-parameter monitoring method based on 5G multi-access edge computing described in claim 1 are implemented.

Images