Intelligent network connection automobile active safety teaching control method based on digital twinning

Abstract

The invention discloses an intelligent network connection automobile active safety teaching control method based on digital twinning. The method comprises the steps that S1, multi-source sensor data are collected and preprocessed to generate a driving state data set; s2, inputting the state data into a digital twin model to drive a traffic scene and outputting a synchronous state; s3, based on Dueling-DDQN, executing strategy learning to generate an active safety control instruction; s4, the instruction response effect is simulated and verified in the virtual environment; s5, collecting driver operation behaviors, inputting the improved MHA-BiLSTM model to extract time sequence features, and outputting behavior features; and S6, comparing the driving behavior with a standard instruction item by item, calculating an operation deviation and a response difference, and generating a personalized active safety teaching task. According to the invention, efficient comparison and teaching feedback of the driving behavior and the active control strategy can be realized, and the intelligent level of driving training is improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin modeling and reinforcement learning, and in particular to a method for active safety teaching and control of intelligent connected vehicles based on digital twins. Background Technology

[0002] In recent years, with the rapid development of autonomous driving and intelligent connected vehicle technologies, active safety teaching and control methods based on driving simulation and behavior analysis have gradually become an important direction for the construction of intelligent vehicle safety systems. However, existing teaching and control systems still have significant shortcomings in data acquisition, virtual-real interaction, behavior analysis, and personalized feedback, making it difficult to meet the training needs of complex road environments and individual differences.

[0003] Currently, most teaching control methods use fixed scenario scripts or static rule models to evaluate driving behavior, lacking dynamic evolution and interactive feedback in real road scenarios, resulting in a disconnect between teaching effectiveness and actual vehicle performance. Although some systems support sensor data input, they are often limited to single-modal information and fail to integrate multi-source data such as lidar, millimeter-wave radar, cameras, and environmental monitoring, affecting the comprehensiveness and timeliness of driving state modeling.

[0004] Furthermore, existing methods for modeling driving behavior still primarily rely on unidirectional RNNs or simple feature statistics, which struggle to fully capture the temporal dependencies and abrupt changes in the sequence of actions. This leads to unstable judgments of complex driving maneuvers, particularly in rapidly changing scenarios where they fail to accurately reflect differences in driver reaction time and stability. Simultaneously, teaching feedback is often presented with standardized scoring, lacking personalized task adjustment mechanisms based on behavioral biases and scenario risk differences, thus limiting the precision and sustainable optimization capabilities of teaching strategies.

[0005] Therefore, how to provide a digital twin-based active safety teaching control method for intelligent connected vehicles is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a digital twin-based active safety teaching and control method for intelligent connected vehicles. This invention utilizes an improved MHA-BiLSTM model to perform multi-scale attention weighting and bidirectional temporal dependency modeling on driving behavior time series, precisely extracting features related to changes in driving actions, operational delays, and stability, thereby enhancing the sensitivity to behavioral deviations in abrupt scenarios. By introducing a Dueling-DDQN network for reinforcement learning optimization of the teaching strategy, it distinguishes between state value and action advantage, significantly improving the generalization ability and safety adaptability of the recommended teaching tasks. This invention possesses advantages such as high accuracy in behavioral modeling, strong adjustability in teaching control, and high efficiency in feedback optimization. It effectively solves problems in existing teaching methods such as insufficient granularity of behavioral perception, rigid training task templates, and poor adaptability to individual differences, comprehensively improving the intelligence and practicality of active safety teaching.

[0007] According to an embodiment of the present invention, an active safety teaching control method for intelligent connected vehicles based on digital twins includes the following steps: S1. Collect multi-source sensor data from intelligent connected vehicles and preprocess it to form a driving status dataset; S2. Map the driving status dataset to the trained digital twin model, load the road and traffic object positions in sequence and drive scene changes, and output the synchronization status in real time. S3. Input the synchronization state output by the digital twin model into Dueling-DDQN to establish a multi-dimensional reward function, perform policy training in different scenarios, and output an active safety control instruction sequence; S4. Perform dynamic simulation based on the digital twin model, apply the active safety control command sequence to the virtual vehicle model, and verify the adaptability and stability of the active safety control command sequence to maintaining safe distance, obstacle avoidance response and braking timing. S5. Collect the driver's operation data in simulated driving, construct an improved MHA-BiLSTM model for driving behavior time series input, perform bidirectional temporal feature extraction and multi-head attention weighted calculation on changes in driving actions, reaction delay and operation stability, and output driving behavior features. S6. Compare driving behavior characteristics with active safety control command sequences item by item along the same time axis, calculate the driving operation deviation value and the difference in response to dangerous scenarios, and generate personalized active safety teaching tasks based on the deviation value and the difference in dangerous response.

[0008] Optionally, S1 specifically includes: S11. Simultaneously activate multiple source sensors through synchronous trigger signals to collect vehicle operating status data, traffic participant information and road environment parameters respectively, read the raw data streams output by each sensor, and record them in the order of collection time labels; S12. Compare each time tag, interpolate the misaligned data points according to the nearest time principle, synchronize the multi-source data of lidar point cloud frame, millimeter wave radar echo frame, camera image frame and environmental monitoring sensor detection frame to a unified time axis, and perform outlier removal processing on the synchronized data. S13. Perform noise suppression processing on the multi-source data after anomaly removal. Use a filtering algorithm to remove random noise from the lidar point cloud frame and millimeter-wave radar echo frame. Perform smoothing processing on the camera image frame and median filtering on the environmental monitoring sensor detection frame. Combine the multi-source data after time synchronization, anomaly removal and noise suppression processing into structured data entries in chronological order, and output the driving status dataset.

[0009] Optionally, S2 specifically includes: S21. Read the vehicle speed, acceleration, steering wheel angle and body posture values ​​recorded in the driving state dataset, arrange the values ​​at each moment in chronological order to form a vehicle dynamics numerical sequence, and write it into the trained digital twin model. S22. Divide the road area into square grid cells according to a preset size, write the corresponding road, lane lines, traffic signs or obstacle markings into each grid cell, and write the grid cells into the road coordinate matrix of the digital twin model. S23. Read the vehicle position from the traffic participant information, record the coordinate value, direction of movement angle and speed value of each object in time sequence to form a traffic participant time series table, and write it into the traffic object data structure of the digital twin model. S24. Write the velocity values ​​in the vehicle dynamics numerical sequence into the longitudinal velocity control field of the virtual vehicle in the digital twin model, write the steering wheel angle values ​​into the steering control field of the virtual vehicle in the digital twin model, write the acceleration values ​​into the acceleration control field of the virtual vehicle in the digital twin model, and write the vehicle body attitude values ​​into the attitude angle field of the virtual vehicle in the digital twin model. S25. Write the road signs recorded in the road coordinate matrix into the road plane of the digital twin model, draw the lane line signs as continuous lines according to the coordinate positions, draw the obstacles as three-dimensional blocks according to their positions, and place the traffic signs as sign primitives according to their coordinates to generate the road scene structure of the digital twin model. S26. Write the traffic participant time series table into the traffic scene of the digital twin model one moment at a time. At each moment, place the virtual vehicle at the corresponding coordinate position and move the object position to the coordinate position of the next moment to generate a frame-by-frame traffic motion picture of the digital twin model. S27. Drive the movement of virtual vehicles in the digital twin model according to the vehicle dynamics numerical sequence, drive the environmental changes of the digital twin model according to the road scene structure and traffic scene data, output each frame of the digital twin image in chronological order, so that the vehicles, roads and traffic objects in the digital twin model are synchronized with the real scene, and output the synchronization status in real time.

[0010] Optionally, S3 specifically includes: S31. Read the synchronization status of the current moment from the digital twin model, and sequentially obtain the collision risk value, vehicle distance change rate value, vehicle body attitude stability value and path offset value. Scale the four values ​​linearly according to the set maximum allowable value and minimum safety value, and concatenate them into a state vector in a preset order. S32. Establish an action list based on driving control requirements, assign a unique action number to each combination of braking level, steering level and throttle level, and form a discrete active control action library. S33. After each vehicle action is executed, read the collision risk value, distance change rate value, vehicle posture stability value and path offset value from the digital twin model. Combine the four values ​​into a reward value according to the preset weight. Store the current state vector, current action number, reward value, next state vector and state termination marker into the experience playback buffer in sequence. S34. Randomly select a batch of samples from the experience replay buffer, input the state vector of the batch of samples into the shared feature processing layer of Dueling-DDQN, and generate intermediate feature vectors through linear weighting and non-linear activation. S35. Input the intermediate feature vector into the value branch and the advantage branch respectively. In the value branch, the current state value score is output by weighting the preset weight vector through a fully connected structure and adding a bias term. In the advantage branch, the action advantage value is output for each action by weighting the preset weight vector through another fully connected structure and adding a bias term. S36. Add and combine the state value score and the action advantage value, and average all action advantage values ​​to obtain a reference average value. Subtract the reference average value from each action advantage value, and add the adjusted action advantage value to the state value score to obtain the comprehensive decision score for each action. S37. Use Dueling-DDQN to determine the target score. In the next time step, find the action number with the highest comprehensive decision score, read the corresponding comprehensive decision score, and sum the current reward value and the comprehensive decision score according to a preset ratio to obtain the target value. If the current state is a terminated state, only the reward value is retained as the target value. S38. Calculate the difference between the comprehensive decision score and the target value item by item to obtain the error signal, and multiply it by the learning rate coefficient of the corresponding layer in turn to obtain the adjustment amount of each parameter of the corresponding layer and perform the update operation a set number of times. S39. When controlling the output, an exploration-utilization balance strategy is adopted. Any action in the action list is selected as the exploration action with a set probability, and the action with the highest current comprehensive decision score is selected as the execution action with the remaining probability. The corresponding braking command, steering command and throttle command are written into the active safety control command sequence and sent for execution.

[0011] Optionally, S4 specifically includes: S41. Read the current position, speed, acceleration and body posture angle of the virtual vehicle from the digital twin model, and write them into the braking, steering and driving fields respectively in combination with the active safety control commands. Substitute them into the preset longitudinal and lateral motion equations respectively to calculate the vehicle speed, acceleration, driving direction angle and posture angle at the next moment. S42. Update the virtual vehicle status and position, read the position of surrounding traffic objects, calculate the relative distance one by one, if it is less than the safe distance threshold, record the safe distance trigger event calculation and the shortest distance to the obstacle, if it is less than the obstacle avoidance threshold, record the obstacle avoidance failure, read the braking response deceleration and delay time, if it exceeds the range, record the response delay. S43. Organize all types of events by time and compare them with the control command sequence. If the preset safety indicators are met, mark them as valid control commands; otherwise, mark them as invalid.

[0012] Optionally, S5 specifically includes: S51. During the simulated driving process, the driver's brake pedal travel, steering wheel angle change and throttle opening data are collected sequentially. Each data point is written with a sampling time label and the original value is recorded. S52. Arrange the braking, steering and throttle data in chronological order according to the sampling time label, delete duplicate time labels and perform interpolation to fill in missing time points, and generate a driving behavior time series composed of the three channels of braking, steering and throttle. S53. The driving behavior time series is submitted as input to the improved MHA-BiLSTM model, and the output is a sequence of driving behavior feature vectors.

[0013] Optionally, the improved MHA-BiLSTM model specifically includes: Before the driving behavior time series is entered into the improved MHA-BiLSTM model, a sensitive gating mechanism for action mutation is introduced to calculate the rate of change of throttle opening, steering wheel angular velocity and braking force step of adjacent time tags, and the absolute values ​​of the three are taken to form the original mutation degree. The baseline fluctuation amplitude of the three channels is calculated using a sliding window method. The net fluctuation amplitude is obtained by subtracting the baseline fluctuation amplitude from the original fluctuation amplitude. The net fluctuation amplitude is compared with the preset throttle change rate threshold, steering wheel angular velocity threshold, and braking force step threshold, respectively. Initial values ​​of the gating coefficient are generated for time tags that exceed the threshold, and a gating coefficient of 1 is generated for time tags that do not exceed the threshold. The initial gating coefficient values ​​are linearly mapped to a preset upper limit range according to the amplitude. Values ​​exceeding the upper limit are truncated. The gating coefficients are smoothed by a one-sided moving average for continuous abrupt segments. The sampling points of the reaction window are expanded to form the final gating coefficient sequence. The final gating coefficient sequence is then multiplied with the three-channel input features one by one to obtain the input vector. The input vector is fed into the forward layer of the improved MHA-BiLSTM model in chronological order. At the first time label, it is multiplied with the initial forward hidden state vector item by item, and the weighted sum is fed into the preset update function to output the first forward hidden state vector. At subsequent time labels, the current input vector is multiplied with the forward hidden state vector of the previous time step item by item, the weighted sum is fed into the update function to output the current forward hidden state vector. The input vector is input into the reverse layer in reverse chronological order. At the last time tag, it is multiplied with the initial reverse hidden state vector item by item and added item by item according to the preset weight. The result is fed into the preset update function to obtain the reverse hidden state vector. At the second to last and the preceding time tags, the input vector is multiplied with the reverse hidden state vector of the previous time tag item by item, weighted, and then fed into the update function to output the reverse hidden state vector of the current time tag. At each time label, the corresponding forward hidden state vector and backward hidden state vector are read and concatenated item by item to generate a bidirectional hidden state vector, which is then arranged in chronological order to form a bidirectional operational feature sequence. The bidirectional operation feature sequence is copied into multiple copies and used as input to multiple attention heads. For each attention head, the bidirectional hidden state vector of any time label is multiplied element by element with the bidirectional hidden state vectors of all time labels and the results are summed to obtain the similarity score of the current time label relative to all time labels. The similarity score sequence of the current attention head is generated by traversing all time labels. Divide the score of each time tag in the similarity score sequence by the sum of all scores in the sequence to obtain the attention weight value of each time tag. Multiply the gating coefficient with the attention weight of the corresponding time tag one by one to obtain the gated weighted attention weight. Arrange them in chronological order to form the gated attention weight sequence. The weight value of each time tag in the gated attention weight sequence is multiplied by the bidirectional hidden state vector of the corresponding time tag and summed along the time axis to obtain the attention head weighted feature vector. All attention head weighted feature vectors are concatenated in order to form a multi-head attention fusion vector. The multi-head attention fusion vector is input into a fully connected structure. Weighted summation and nonlinear activation are performed sequentially in the fully connected structure to obtain driving action change scores, reaction delay scores, and operational stability scores. The results are then output as a driving behavior feature sequence in time label order.

[0014] Optionally, S6 specifically includes: S61. Read the driving behavior feature sequence sequentially according to the sampling time label, and form a driving behavior feature vector by the driving action change score, reaction delay score and operation stability score corresponding to each time label. Read the active safety control command sequence sequentially according to the same time label, and form a command feature vector by the braking command amount, steering command angle and throttle command opening corresponding to each time label. S62. At each time tag, subtract the driving action change score in the driving behavior feature vector from the braking command amount, direction command angle change and throttle command change in the command feature vector one by one, calculate the driving action deviation value, and sort them in ascending order of absolute value. S63. Subtract the reaction delay score in the driving behavior feature vector from the instruction change time at the corresponding time tag in the instruction feature vector one by one to obtain the reaction delay deviation value. Subtract the operation stability score in the driving behavior feature vector from the stable execution time interval corresponding to the instruction feature vector one by one to obtain the operation stability deviation value. S64. The action deviation value, reaction delay deviation value and operation stability deviation value are weighted and summed according to preset weights to obtain the comprehensive deviation score of the current time tag. The teaching task level is divided from low to high according to the comprehensive deviation score, and the current time tag with the largest comprehensive deviation is set as the key practice content.

[0015] The beneficial effects of this invention are: This invention constructs an active safety teaching control process for intelligent connected vehicles by integrating driving behavior data collection, digital twin scenario construction, and deep reinforcement learning strategy reasoning, significantly improving the personalization level of teaching tasks and the stability of response decisions. In the strategy decision-making stage, a state value function and action advantage function are constructed based on the Dueling-DDQN structure to achieve branch updates and joint decision-making, effectively enhancing the model's ability to select the optimal teaching strategy in complex scenarios. An improved MHA-BiLSTM model is used to extract features from the driver's operation time series, introducing multi-head attention weighting and action mutation sensitive gating mechanisms to strengthen the perception of key driving deviations and improve the accuracy of driving behavior modeling and the timeliness of risk response. Teaching tasks and training difficulty are dynamically generated by comparing the deviation between driving behavior and active safety control commands, supporting a task adaptation mechanism tailored to individual needs and significantly enhancing the pertinence and effectiveness of teaching interaction. This invention can dynamically adapt to individual differences in changing traffic environments, improving the effectiveness and intelligence of active safety education. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of an active safety teaching control method for intelligent connected vehicles based on digital twins proposed in this invention; Figure 2 This invention presents the active safety control strategy learning and instruction generation process based on the DuelingDDQN structure. Figure 3 This is a flowchart of the improved MHABiLSTM driving behavior temporal feature extraction and driving deviation recognition proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A digital twin-based active safety teaching control method for intelligent connected vehicles includes the following steps: S1. Collect multi-source sensor data from intelligent connected vehicles, including LiDAR, millimeter-wave radar, cameras and environmental monitoring sensors, to obtain vehicle operating status, traffic participant information and road environment data, and perform time synchronization, outlier removal and noise suppression processing on the multi-source data to form a driving status dataset. S2. Map the driving state dataset to the trained digital twin model, write the vehicle speed, acceleration, steering and attitude into the virtual vehicle control field, load the road and traffic object positions in sequence and drive scene changes, and output the synchronization status in real time. S3. Input the synchronous state output by the digital twin model into Dueling-DDQN to establish a multi-dimensional reward function for collision risk, vehicle distance change rate, vehicle attitude stability and path deviation. Execute strategy training in scenarios of normal driving, emergency braking, obstacle avoidance and collision risk, and output active safety control command sequence. S4. Based on the digital twin model, perform dynamic simulation, apply the active safety control command sequence to the virtual vehicle model, simulate vehicle operation behavior, and verify the adaptability and stability of the active safety control command sequence to safety distance maintenance, obstacle avoidance response and braking timing. S5. Collect the driver's braking, steering and throttle operation data during the simulation process, and construct an improved MHA-BiLSTM model of driving behavior time series input in chronological order. Perform bidirectional temporal feature extraction and multi-head attention weighted calculation on driving action changes, reaction time delay and operation stability, and output driving behavior features. S6. Compare driving behavior characteristics with active safety control command sequences item by item along the same time axis, calculate the driving operation deviation value and the difference in response to dangerous scenarios, and generate personalized active safety teaching tasks based on the deviation value and the difference in dangerous response.

[0020] This implementation significantly enhances the personalization and effectiveness of driving behavior instruction for intelligent connected vehicles by constructing a closed-loop active safety assessment system that integrates simulation and teaching feedback. First, multi-source sensor data is collected and simultaneously input into a trained digital twin model to simulate the vehicle's real-world operating state and traffic environment, ensuring the verifiability of control strategies in complex dynamic scenarios. The Dueling-DDQN approach is then used to optimize the active safety control command sequence and assess collision risks, achieving sensitive responses to emergency braking, obstacle avoidance, and lane departure. Furthermore, driver actions are collected to construct a time series of behaviors, which are then input into an improved MHA-BiLSTM model to extract temporal features and deviation indicators. By comparing driving behavior with safety control strategies, individual operational deviations and scenario response differences are quantified, intelligently generating differentiated teaching tasks, strengthening learning behavior feedback, and effectively improving safety awareness and operational consistency during the teaching process.

[0021] In this embodiment, S1 specifically includes: S11. Simultaneously activate multi-source sensors including lidar, millimeter-wave radar, camera, and environmental monitoring through synchronous trigger signal, collect vehicle operation status data, traffic participant information, and road environment parameters respectively, read the raw data streams output by each sensor, and record lidar point cloud, millimeter-wave radar distance and velocity echo, camera image frames, and environmental monitoring sensor detection values ​​in the order of collection time label. S12. Compare each time tag, interpolate misaligned data points according to the nearest time principle, synchronize the multi-source data of lidar point cloud frame, millimeter-wave radar echo frame, camera image frame and environmental monitoring sensor detection frame to a unified time axis, perform outlier removal processing on the synchronized data, delete distance jump anomaly points, velocity change anomaly values ​​and image distortion frames, and eliminate erroneous readings of environmental monitoring sensors. S13. Perform noise suppression processing on the multi-source data after anomaly removal. Use a filtering algorithm to remove random noise from the lidar point cloud frame and millimeter-wave radar echo frame. Perform smoothing processing on the camera image frame and median filtering on the environmental monitoring sensor detection frame. Combine the multi-source data after time synchronization, anomaly removal and noise suppression processing into structured data entries in chronological order, and output the driving status dataset.

[0022] In this embodiment, S2 specifically includes: S21. Read the vehicle speed, acceleration, steering wheel angle and body posture values ​​recorded in the driving state dataset, arrange the values ​​at each moment in chronological order to form a vehicle dynamics numerical sequence, and write it into the trained digital twin model. S22. Divide the road area into square grid cells according to a preset size, write the corresponding road, lane lines, traffic signs or obstacle markings into each grid cell, and write the grid cells into the road coordinate matrix of the digital twin model. S23. Read the vehicle position from the traffic participant information, record the coordinate value, direction of movement angle and speed value of each object in time sequence to form a traffic participant time series table, and write it into the traffic object data structure of the digital twin model. S24. Write the velocity values ​​in the vehicle dynamics numerical sequence into the longitudinal velocity control field of the virtual vehicle in the digital twin model, write the steering wheel angle values ​​into the steering control field of the virtual vehicle in the digital twin model, write the acceleration values ​​into the acceleration control field of the virtual vehicle in the digital twin model, and write the vehicle body attitude values ​​into the attitude angle field of the virtual vehicle in the digital twin model. S25. Write the road signs recorded in the road coordinate matrix into the road plane of the digital twin model, draw the lane line signs as continuous lines according to the coordinate positions, draw the obstacles as three-dimensional blocks according to their positions, and place the traffic signs as sign primitives according to their coordinates to generate the road scene structure of the digital twin model. S26. Write the traffic participant time series table into the traffic scene of the digital twin model one moment at a time. At each moment, place the virtual vehicle at the corresponding coordinate position and move the object position to the coordinate position of the next moment to generate a frame-by-frame traffic motion picture of the digital twin model. S27. Drive the movement of virtual vehicles in the digital twin model according to the vehicle dynamics numerical sequence, drive the environmental changes of the digital twin model according to the road scene structure and traffic scene data, output each frame of the digital twin image in chronological order, so that the vehicles, roads and traffic objects in the digital twin model are synchronized with the real scene, and output the synchronization status in real time.

[0023] In this embodiment, S3 specifically includes: S31. Read the synchronization status of the current moment from the digital twin model, and sequentially obtain the collision risk value, vehicle distance change rate value, vehicle body attitude stability value and path offset value. Scale the four values ​​linearly according to the set maximum allowable value and minimum safety value, and concatenate them into a state vector in a preset order. S32. Establish an action list based on driving control requirements, assign a unique action number to each combination of braking level, steering level and throttle level, and form a discrete active control action library. S33. After each vehicle action is executed, read the collision risk value, distance change rate value, vehicle posture stability value and path offset value from the digital twin model. Combine the four values ​​into a reward value according to the preset weight. Store the current state vector, current action number, reward value, next state vector and state termination marker into the experience playback buffer in sequence. S34. Randomly select a batch of samples from the experience replay buffer, input the state vector of the batch of samples into the shared feature processing layer of Dueling-DDQN, and generate intermediate feature vectors through linear weighting and non-linear activation. S35. Input the intermediate feature vector into the value branch and the advantage branch respectively. In the value branch, the current state value score is output by weighting the preset weight vector through a fully connected structure and adding a bias term. In the advantage branch, the action advantage value is output for each action by weighting the preset weight vector through another fully connected structure and adding a bias term. S36. Add and combine the state value score and the action advantage value, and average all action advantage values ​​to obtain a reference average value. Subtract the reference average value from each action advantage value, and add the adjusted action advantage value to the state value score to obtain the comprehensive decision score for each action. S37. Use Dueling-DDQN to determine the target score. In the next time step, find the action number with the highest comprehensive decision score, read the corresponding comprehensive decision score, and sum the current reward value and the comprehensive decision score according to a preset ratio to obtain the target value. If the current state is a terminated state, only the reward value is retained as the target value. S38. Calculate the difference between the comprehensive decision score and the target value item by item to obtain the error signal, and multiply it by the learning rate coefficient of the corresponding layer in turn to obtain the adjustment amount of each parameter of the corresponding layer and perform the update operation a set number of times. S39. When controlling the output, an exploration-utilization balance strategy is adopted. Any action in the action list is selected as the exploration action with a set probability, and the action with the highest current comprehensive decision score is selected as the execution action with the remaining probability. The corresponding braking command, steering command and throttle command are written into the active safety control command sequence and sent for execution.

[0024] This implementation introduces Dueling-DDQN to construct an active safety control decision network, significantly improving the risk response capability and training efficiency of the control strategy in complex driving scenarios. First, collision risk, distance change rate, attitude stability, and path deviation are read from a digital twin model and linearly scaled to construct a state vector. Simultaneously, an active control action library with numbered actions is built. After each control execution, the state, action, and reward value are recorded and stored in an experience replay buffer for batch sampling training. Through a shared feature extraction layer and a value-advantage dual-branch structure, the state value and the advantage of each action are calculated separately, and mean normalization is performed to fuse them and generate a comprehensive decision score. In the Q-value update stage, a target network is introduced, combining the current reward and the target action value to generate a target value, and the network parameters are updated through gradient error backpropagation. An exploration-exploitation strategy is adopted in the control output to achieve a dynamic balance between prioritizing high-value actions and strategy diversity, thereby generating a stable and reliable active safety control command sequence.

[0025] In this embodiment, S4 specifically includes: S41. Read the current position, speed, acceleration and body posture angle of the virtual vehicle from the digital twin model, and write them into the braking, steering and driving fields respectively in combination with the active safety control commands. Substitute them into the preset longitudinal and lateral motion equations respectively to calculate the vehicle speed, acceleration, driving direction angle and posture angle at the next moment. S42. Update the virtual vehicle status and position, read the position of surrounding traffic objects, calculate the relative distance one by one, if it is less than the safe distance threshold, record the safe distance trigger event calculation and the shortest distance to the obstacle, if it is less than the obstacle avoidance threshold, record the obstacle avoidance failure, read the braking response deceleration and delay time, if it exceeds the range, record the response delay. S43. Organize all types of events by time and compare them with the control command sequence. If the preset safety indicators are met, mark them as valid control commands; otherwise, mark them as invalid.

[0026] In this embodiment, S5 specifically includes: S51. During the simulated driving process, the driver's brake pedal travel, steering wheel angle change and throttle opening data are collected sequentially. Each data point is written with a sampling time label and the original value is recorded. S52. Arrange the braking, steering and throttle data in chronological order according to the sampling time label, delete duplicate time labels and perform interpolation to fill in missing time points, and generate a driving behavior time series composed of the three channels of braking, steering and throttle. S53. The driving behavior time series is submitted as input to the improved MHA-BiLSTM model, and the output is a sequence of driving behavior feature vectors.

[0027] In this embodiment, the improved MHA-BiLSTM model specifically includes: Before the driving behavior time series is entered into the improved MHA-BiLSTM model, a sensitive gating mechanism for action mutation is introduced to calculate the rate of change of throttle opening, steering wheel angular velocity and braking force step of adjacent time tags, and the absolute values ​​of the three are taken to form the original mutation degree. The baseline fluctuation amplitude of the three channels is calculated using a sliding window method. The net fluctuation amplitude is obtained by subtracting the baseline fluctuation amplitude from the original fluctuation amplitude. The net fluctuation amplitude is compared with the preset throttle change rate threshold, steering wheel angular velocity threshold, and braking force step threshold, respectively. Initial values ​​of the gating coefficient are generated for time tags that exceed the threshold, and a gating coefficient of 1 is generated for time tags that do not exceed the threshold. The initial gating coefficient values ​​are linearly mapped to a preset upper limit range according to the amplitude. Values ​​exceeding the upper limit are truncated. The gating coefficients are smoothed by a one-sided moving average for continuous abrupt segments. The sampling points of the reaction window are expanded to form the final gating coefficient sequence. The final gating coefficient sequence is then multiplied with the three-channel input features one by one to obtain the input vector. The input vector is fed into the forward layer of the improved MHA-BiLSTM model in chronological order. At the first time label, it is multiplied with the initial forward hidden state vector item by item, and the weighted sum is fed into the preset update function to output the first forward hidden state vector. At subsequent time labels, the current input vector is multiplied with the forward hidden state vector of the previous time step item by item, the weighted sum is fed into the update function to output the current forward hidden state vector. The input vector is input into the reverse layer in reverse chronological order. At the last time tag, it is multiplied with the initial reverse hidden state vector item by item and added item by item according to the preset weight. The result is fed into the preset update function to obtain the reverse hidden state vector. At the second to last and the preceding time tags, the input vector is multiplied with the reverse hidden state vector of the previous time tag item by item, weighted, and then fed into the update function to output the reverse hidden state vector of the current time tag. At each time label, the corresponding forward hidden state vector and backward hidden state vector are read and concatenated item by item to generate a bidirectional hidden state vector, which is then arranged in chronological order to form a bidirectional operational feature sequence. The bidirectional operation feature sequence is copied into multiple copies and used as input to multiple attention heads. For each attention head, the bidirectional hidden state vector of any time label is multiplied element by element with the bidirectional hidden state vectors of all time labels and the results are summed to obtain the similarity score of the current time label relative to all time labels. The similarity score sequence of the current attention head is generated by traversing all time labels. Divide the score of each time tag in the similarity score sequence by the sum of all scores in the sequence to obtain the attention weight value of each time tag. Multiply the gating coefficient with the attention weight of the corresponding time tag one by one to obtain the gated weighted attention weight. Arrange them in chronological order to form the gated attention weight sequence. The weight value of each time tag in the gated attention weight sequence is multiplied by the bidirectional hidden state vector of the corresponding time tag and summed along the time axis to obtain the attention head weighted feature vector. All attention head weighted feature vectors are concatenated in order to form a multi-head attention fusion vector. The multi-head attention fusion vector is input into a fully connected structure. Weighted summation and nonlinear activation are performed sequentially in the fully connected structure to obtain driving action change scores, reaction delay scores, and operational stability scores. The results are then output as a driving behavior feature sequence in time label order.

[0028] This implementation achieves fine-grained perception and accurate feature extraction of driving behavior by introducing a change-sensitive gating mechanism and an improved MHA-BiLSTM model. First, a three-channel driving behavior time series is constructed for braking, steering, and throttle operations during the simulation. The model's sensitivity to operational changes is enhanced through change-sensitivity calculation and gating coefficient adjustment. Forward and reverse temporal features are simultaneously extracted in the MHA-BiLSTM to construct a bidirectional hidden state sequence. A multi-head attention mechanism is introduced to assign dynamic weights to operational features at different time labels, achieving accurate modeling of operational changes, reaction delays, and stability. The final output driving behavior feature sequence can be used for personalized driving deviation identification and safety teaching task generation, significantly improving the model's ability to capture dynamic driving behavior and the accuracy of behavioral deviation discrimination.

[0029] In this embodiment, S6 specifically includes: S61. Read the driving behavior feature sequence sequentially according to the sampling time label, and form a driving behavior feature vector by the driving action change score, reaction delay score and operation stability score corresponding to each time label. Read the active safety control command sequence sequentially according to the same time label, and form a command feature vector by the braking command amount, steering command angle and throttle command opening corresponding to each time label. S62. At each time tag, subtract the driving action change score in the driving behavior feature vector from the braking command amount, direction command angle change and throttle command change in the command feature vector one by one, calculate the driving action deviation value, and sort them in ascending order of absolute value. S63. Subtract the reaction delay score in the driving behavior feature vector from the instruction change time at the corresponding time tag in the instruction feature vector one by one to obtain the reaction delay deviation value. Subtract the operation stability score in the driving behavior feature vector from the stable execution time interval corresponding to the instruction feature vector one by one to obtain the operation stability deviation value. S64. The action deviation value, reaction delay deviation value and operation stability deviation value are weighted and summed according to preset weights to obtain the comprehensive deviation score of the current time tag. The teaching task level is divided from low to high according to the comprehensive deviation score, and the current time tag with the largest comprehensive deviation is set as the key practice content.

[0030] Example 1: To verify the feasibility and superiority of this invention in the active safety teaching control scenario of intelligent connected vehicles, it was deployed in the intelligent driving teaching experimental platform of a key transportation vocational college. This platform includes 12 intelligent connected passenger vehicles used for teaching, equipped with onboard environmental perception equipment, onboard control units, V2X communication terminals, and a digital twin platform, covering six types of teaching and testing scenarios, including typical urban roads, roundabouts, and intersections. Taking the teaching experimental period from March to May as an example, a total of 87 training sessions were conducted, involving 46 driving students, forming a complete closed-loop data process of driving task-execution behavior-feedback correction.

[0031] In traditional teaching, the assessment of active vehicle safety operations mainly relies on manual scoring by teachers and video review, which suffers from delayed feedback, strong subjectivity, and difficulty in covering all risk states. This invention, by constructing an improved MHA-BiLSTM driving behavior temporal feature extraction model and a Dueling-DDQN reinforcement learning strategy optimization model, achieves real-time risk identification and dynamic generation of control commands for driving behavior in teaching scenarios. The teaching vehicle is bidirectionally connected to a digital twin platform via a CAN bus, collecting state data such as vehicle speed, acceleration, steering angle, distance to other vehicles, and lane departure at every moment during the student's driving process, while simultaneously monitoring indicators such as collision risk and path deviation trends in the virtual environment. This data is linearly normalized and concatenated into a state vector input model. The MHA-BiLSTM model extracts the temporal dependency features of driving behavior, identifying behavioral segments that deviate from the normal paradigm. Subsequently, the state vector is input into the Dueling-DDQN model, which generates a comprehensive score based on action advantages and state value, outputting action commands including active braking, path correction, and deceleration warning, which are automatically transmitted back to the teaching platform for real-time intervention or scoring recording.

[0032] The following table presents comparative data between the method of this invention and traditional methods on several key teaching indicators during actual deployment: Table 1. Comparison data between the method of the present invention and the traditional teaching model in proactive safety teaching and control tasks.

[0033] As can be seen from the comparison of teaching experiment data shown in Table 1, the improved MHA-BiLSTM model and Dueling-DDQN policy learning mechanism proposed in this invention are significantly better than traditional teaching models in the active safety teaching and control scenario of intelligent connected vehicles, and have achieved comprehensive improvement in multiple key dimensions such as response timeliness, behavior recognition accuracy, path correction and risk intervention.

[0034] In terms of response time, this invention uses a reinforcement learning strategy to dynamically optimize the control command generation chain, which significantly reduces the average response time from the traditional 6.3 seconds to 0.86 seconds, a reduction of 86.4%. This significantly improves the efficiency of intervention in sudden driving deviations during training and effectively reduces the safety risks caused by delays.

[0035] Regarding the accuracy of driving behavior recognition and path deviation detection, thanks to the improved multi-head attention bidirectional LSTM model's ability to model the temporal sequence features of state sequence features, this invention achieves accuracy rates of 94.2% and 92.7% for dangerous driving behavior recognition and lane departure detection, respectively, representing improvements of 22.4% and 24.2% compared to traditional methods relying on human experience and rule matching. Especially in dynamic scenarios such as complex curves and congested traffic, this invention demonstrates stronger stability and robustness.

[0036] Furthermore, by introducing an automatic correction mechanism into the teaching scenario, this invention can output precise braking, steering, and throttle control commands based on real-time status, achieving an active correction coverage rate of 89.5%, breaking away from the lagging model of traditional teaching that relies on teacher post-grading and video playback. Combined with an automatic grading mechanism, this invention improves the consistency of teaching grading by 91.3%, ensuring the objectivity and repeatability of teaching results.

[0037] In terms of error identification and intervention, this invention increases the correction rate of traditional error scoring from 56.4% to 93.8%, significantly reducing the proportion of missed and incorrect judgments. The average number of missed judgments is reduced from 2.4 times per class hour to 0.3 times, saving a lot of manual review time and costs.

[0038] In summary, this invention constructs an efficient, accurate, and deployable active safety teaching control method in intelligent driving teaching scenarios by integrating digital twin modeling, temporal feature extraction, and reinforcement learning control strategies. It effectively overcomes the shortcomings of traditional methods, such as delayed feedback, strong subjectivity in judgment, and untimely response, and has good practical value and prospects for promotion.

[0039] In summary, this invention significantly optimizes the processes of active safety behavior recognition, control command generation, and teaching intervention in the teaching control scenario of intelligent connected vehicles. It features high timeliness, high accuracy, and good model interpretability, making it particularly suitable for teaching and training deployment in dynamic environments. It has significant application value and promotion potential.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital-twin-based intelligent networked vehicle active safety teaching control method, characterized in that, The method comprises the following steps: S1, collecting multi-source sensor data of the intelligent connected vehicle, and preprocessing to form a driving state data set; S2, mapping the driving state data set to the trained digital twin model, loading and driving the scene change in time sequence according to the positions of the road and traffic objects, and outputting the synchronous state in real time; S3, inputting the synchronous state output by the digital twin model into the Dueling-DDQN to establish a multi-dimensional reward function, performing policy training in different scenes, and outputting an active safety control instruction sequence; S4, based on the digital twin model, performing dynamic simulation, applying the active safety control instruction sequence to the virtual vehicle model, and verifying the adaptability and stability of the active safety control instruction sequence to the safety distance maintenance, obstacle avoidance response and braking time sequence; S5, collecting the operation data of the driver in the simulation driving, constructing a driving behavior time sequence, inputting the improved MHA-BiLSTM model, performing bidirectional time sequence feature extraction and multi-head attention weighted calculation on the driving action change, reaction time delay and operation stability, and outputting the driving behavior features; S6, comparing the driving behavior features and the active safety control instruction sequence item by item according to the same time axis, calculating the driving operation deviation value and the dangerous scene response difference, and generating a personalized active safety teaching task according to the deviation value and the dangerous response difference.

2. The digital-twin-based intelligent connected vehicle active safety teaching control method according to claim 1, characterized in that, The S1 specifically comprises: S11, simultaneously starting the multi-source sensors through a synchronous trigger signal, respectively collecting vehicle operating state data, traffic participant information and road environment parameters, reading the original data stream output by each sensor, and recording in order according to the collection time tag; S12, comparing each time tag, interpolating the unaligned data points according to the nearest time principle, synchronizing the multi-source data of the laser radar point cloud frame, the millimeter wave radar echo frame, the camera image frame and the environmental monitoring sensor detection frame to the unified time axis, and performing abnormal point elimination processing on the synchronized data of each type; S13, performing noise suppression processing on the multi-source data after abnormality elimination, removing random noise from the laser radar point cloud frame and the millimeter wave radar echo frame, performing smoothing processing on the camera image frame, and using median filtering on the environmental monitoring sensor detection frame, grouping the multi-source data after time synchronization, abnormal point elimination and noise suppression processing into structured data entries in time sequence, and outputting the driving state data set. 3.The digital-twin-based intelligent networked vehicle active safety teaching control method according to claim 1, characterized in that, The S2 specifically comprises: S21, reading the vehicle speed value, acceleration value, steering wheel angle value and vehicle body attitude value recorded in the driving state data set, arranging the values at each time in time sequence to form a vehicle dynamics value sequence, and writing into the trained digital twin model; S22, dividing the road range into square grid units according to the preset size, writing the corresponding road, lane line, traffic sign or obstacle identifier in each grid unit, and writing the grid unit into the road coordinate matrix of the digital twin model; S23, reading the vehicle position from the traffic participant information, recording the coordinate value, motion direction angle and motion speed value of each object in time sequence to form a traffic participant time sequence table, and writing into the traffic object data structure of the digital twin model; S24, write the speed value in the vehicle dynamics numerical sequence into the longitudinal speed control field of the virtual vehicle in the digital twin model, write the steering wheel angle value into the steering control field of the virtual vehicle in the digital twin model, write the acceleration value into the acceleration control field of the virtual vehicle in the digital twin model, and write the vehicle body attitude value into the attitude angle field of the virtual vehicle in the digital twin model; S25, write the road identification recorded in the road coordinate matrix into the road plane of the digital twin model, draw the lane line identification as a continuous line according to the coordinate position, draw the obstacle as a solid block according to the position, and place the traffic sign as a sign primitive according to the coordinate to generate the road scene structure of the digital twin model; S26, write the traffic participant time sequence table into the traffic scene of the digital twin model at each time, place the virtual vehicle at the corresponding coordinate position at each time, and move the object position according to the next time coordinate position to generate the frame-by-frame traffic motion picture of the digital twin model; S27, drive the virtual vehicle in the digital twin model according to the vehicle dynamics numerical sequence, drive the environment change of the digital twin model according to the road scene structure and traffic scene data, output each frame of digital twin picture in time sequence, make the vehicle, road and traffic object in the digital twin model keep synchronization with the real scene, and output the synchronization state in real time.

4. The digital-twin-based intelligent networked vehicle active safety teaching control method according to claim 1, characterized in that, The S3 specifically comprises: S31, read the synchronization state of the current time from the digital twin model, sequentially obtain the collision risk value, the vehicle distance change rate value, the vehicle body attitude stability value and the path offset value, linearly scale the four values according to the preset maximum allowed value and the minimum safety value, and splice them into a state vector according to the preset order; S32, establish an action list according to the driving control demand, assign a unique action number to each combination of brake level, steering level and throttle level to form a discrete active control action library; S33, after each vehicle action is executed, read the collision risk value, the vehicle distance change rate value, the vehicle body attitude stability value and the path offset value in the digital twin model this time, combine the four values into a reward value according to the preset weight, store the current state vector, this time action number, reward value, next state vector and state termination mark in the experience replay buffer in sequence; S34, randomly extract a batch of samples from the experience replay buffer, input the state vector of the batch of samples into the shared feature processing layer of the Dueling-DDQN, and generate an intermediate feature vector through linear weighting and nonlinear activation; S35, input the intermediate feature vector into the value branch and the advantage branch respectively, output the state value score of the current state in the value branch through a layer of fully connected structure and a preset weight vector after weighting and adding a bias term, and output the action advantage value of each action in the advantage branch through another fully connected structure and a preset weight vector after weighting and adding a bias term. S36, add the state value score and the action advantage value, average all the action advantage values to obtain a reference average value, subtract the reference average value from each action advantage value, add the adjusted action advantage value to the state value score to obtain a comprehensive decision score corresponding to each action; S37, determine the target score by using the Dueling-DDQN, find the action number with the highest comprehensive decision score in the next state, read the corresponding comprehensive decision score, add the current reward value and the comprehensive decision score according to a preset proportion to obtain a target value, if the current state is a termination state, only the reward value is reserved as the target value; S38, calculate the difference between the comprehensive decision score and the target value item by item to obtain an error signal, and multiply the error signal by the learning rate coefficient of the corresponding layer in turn to obtain the adjustment amount of each parameter of the corresponding layer and perform the update operation for a set number of times; S39, adopt an exploration-exploitation balance strategy when outputting control, select any one action in the action list as an exploration action with a set probability, select the action with the highest current comprehensive decision score as an execution action with the remaining probability, and write the corresponding brake instruction, steering instruction and throttle instruction into the active safety control instruction sequence and send it for execution.

5. The digital-twin-based intelligent connected vehicle active safety teaching control method according to claim 1, characterized in that, The S4 specifically comprises: S41, read the current position, speed, acceleration and body attitude angle of the virtual vehicle from the digital twin model, combine the active safety control instruction to write into the brake, steering and drive fields respectively, and substitute into the preset longitudinal and lateral motion equations to calculate the vehicle speed, acceleration, driving direction angle and attitude angle at the next moment; S42, update the virtual vehicle state and position, read the positions of surrounding traffic objects, calculate the relative distance one by one, if it is less than the safety distance threshold, record the safety distance trigger event calculation and the shortest distance to the obstacle, if it is less than the obstacle avoidance threshold, record the obstacle avoidance failure, read the brake response deceleration and delay time, if it is out of range, record the response delay; S43, sort all kinds of events by time and compare with the control instruction sequence, if the preset safety indicators are met, mark as valid control instruction, otherwise mark as invalid.

6. The digital-twin-based intelligent networked vehicle active safety teaching control method according to claim 1, characterized in that, The S5 specifically comprises: S51, collect the brake pedal stroke, steering wheel angle change and throttle opening data of the driver in sequence during the simulation driving process, write the sampling time label for each data and record the original value; S52, arrange the brake, steering and throttle data in time sequence according to the sampling time label, delete the repeated time label and perform interpolation to fill in the missing time points, generate a driving behavior time sequence composed of three channels of brake, steering and throttle; S53, submit the driving behavior time sequence as input to the improved MHA-BiLSTM model, and output a driving behavior feature vector sequence.

7. The digital-twin-based intelligent networked vehicle active safety teaching control method according to claim 6, characterized in that, The improved MHA-BiLSTM model specifically comprises: Before the driving behavior time sequence enters the improved MHA-BiLSTM model, an action mutation sensitive gating mechanism is introduced, the throttle opening rate of change, the steering wheel angular velocity and the brake force step amount of adjacent time labels are calculated, and the absolute values of the three are taken to form the original mutation degree; The baseline fluctuation amplitude of the three channels is calculated in a sliding window manner, the original mutation degree is subtracted by the baseline fluctuation amplitude to obtain a net mutation degree, the net mutation degree is compared with preset throttle change rate threshold, steering wheel angle speed threshold and brake force step threshold respectively, the time tags exceeding the threshold are generated with initial gating coefficient, and the time tags not exceeding the threshold are generated with gating coefficient being 1; The initial gating coefficient is linearly mapped to a preset upper limit range according to the amplitude, the values exceeding the upper limit are truncated, the gating coefficient is smoothed by using a one-sided moving average for a continuous mutation segment, and the sampling points of the reaction window are expanded to form a final gating coefficient sequence, and the final gating coefficient sequence is multiplied with the three-channel input features in sequence to obtain an input vector; The input vector is input into the improved MHA-BiLSTM model forward layer in time sequence, the first time tag is multiplied with an initial forward hidden state vector in sequence, the weighted sum is obtained according to a preset weight, and then the weighted sum is sent to a preset update function to output a first forward hidden state vector, in subsequent time tags, the current input vector is multiplied with a previous forward hidden state vector in sequence, the weighted sum is obtained, and then the weighted sum is sent to the update function to output a current forward hidden state vector; The input vector is input into the reverse layer in reverse time sequence, the last time tag is multiplied with an initial reverse hidden state vector in sequence and the weighted sum is obtained according to a preset weight, and then the weighted sum is sent to a preset update function to obtain a reverse hidden state vector, in the second last and subsequent time tags, the input vector is multiplied with a previous reverse hidden state vector in sequence, the weighted sum is obtained, and then the weighted sum is sent to the update function to output a reverse hidden state vector of the current time tag; The corresponding forward hidden state vector and reverse hidden state vector are read at each time tag, and the bidirectional hidden state vector is generated by splicing the forward hidden state vector and the reverse hidden state vector in sequence, and the bidirectional hidden state vector is arranged in time sequence as a bidirectional operation feature sequence; The bidirectional operation feature sequence is copied as multiple copies, and is used as the input of multiple attention heads, the bidirectional hidden state vector of any time tag in each attention head is multiplied with the bidirectional hidden state vector of all time tags to obtain a similarity score of the current time tag relative to all time tags, and the similarity score sequence of the attention head is generated by traversing all time tags; The similarity score of each time tag in the similarity score sequence is divided by the sum of all scores in the sequence to obtain the attention weight value of each time tag, and the gating coefficient is multiplied with the attention weight of the corresponding time tag to obtain a gated and weighted attention weight, which is arranged in time sequence as a gated attention weight sequence; The weight value of each time tag in the gated attention weight sequence is multiplied with the bidirectional hidden state vector of the corresponding time tag to obtain an attention head weighted feature vector, and the weighted feature vectors of all attention heads are spliced in sequence as a multi-head attention fusion vector; The multi-head attention fusion vector is input into a full connection structure, and the weighted sum and nonlinear activation are sequentially performed in the full connection structure to obtain a driving action change score, a reaction time delay score and an operation stability score, and the driving behavior feature sequence is output in time tag sequence. 8.The digital-twin-based intelligent networked vehicle active safety teaching control method of claim 1, wherein The S6 specifically comprises: S61, read the driving behavior feature sequence in turn according to the sampling time label, and compose the driving action change score, the reaction time delay score and the operation stability score corresponding to each time label into a driving behavior feature vector; read the active safety control instruction sequence in turn according to the same time label, and compose the brake instruction amount, the steering instruction angle and the throttle instruction opening into an instruction feature vector; S62, at each time label, subtract the driving action change score in the driving behavior feature vector from the brake instruction amount, the steering instruction angle change and the throttle instruction change in the instruction feature vector item by item, calculate the driving action deviation value, and sort the absolute values from small to large; S63, subtract the reaction time delay score in the driving behavior feature vector from the instruction change time at the corresponding time label of the instruction feature vector item by item to obtain the reaction time delay deviation value, and subtract the operation stability score in the driving behavior feature vector from the stable execution time interval corresponding to the instruction feature vector item by item to obtain the operation stability deviation value; S64, weight and sum the action deviation value, the reaction time delay deviation value and the operation stability deviation value according to the preset weight as the comprehensive deviation score of the current time label, divide the teaching task level according to the comprehensive deviation score from low to high, and set the current time label with the maximum comprehensive deviation as the key practice content.

Images