A vehicle control method and system considering the effectiveness of the pre-warning of the operating driver

Abstract

This invention belongs to the field of road vehicle control and relates to a vehicle control method and system that considers the effectiveness of warnings for commercial drivers. The method acquires vehicle trajectory data, warning event data, and static feature data; constructs an adaptive reinforcement learning trajectory reconstruction model based on physical feasible domain constraints; outputs the reconstructed velocity using the trajectory reconstruction model; then backtracks and updates the position and heading angle to obtain the reconstructed trajectory; and constructs a prediction model for the effectiveness of warnings for commercial drivers, using the reconstructed kinematic data within a fixed short window before the warning is triggered as the temporal backbone input, introducing warning types, incorporating static feature data into a unified model, and outputting a predicted value and level of warning effectiveness based on the prediction model. Different vehicle control strategies are adopted according to the warning effectiveness level, realizing a shift from a control paradigm of "passive triggering based on operating condition thresholds" to "active collaborative intervention based on driver reaction prediction," ensuring safe driving.

Description

Technical Field

[0001] This invention belongs to the field of road vehicle control, specifically, it relates to a vehicle control method and system that takes into account the effectiveness of warnings for commercial drivers. Background Technology

[0002] Commercial vehicles play a crucial role in my country's road passenger and freight transport network, providing essential transport capacity and supporting the transportation industry. However, commercial vehicles operate under high intensity, cover long distances, and carry a high concentration of passengers and goods. Consequently, traffic accidents involving these vehicles often result in concentrated casualties, significant social impact, and complex rescue and response procedures. To improve the safety of commercial vehicle operations, it is necessary to equip key commercial passenger vehicles with features such as lane departure warning and forward collision warning.

[0003] Currently, proactive safety warning functions are widely used in the commercial vehicle sector. However, the deployment of warning functions does not necessarily mean that risks can be mitigated in a timely and effective manner. Due to significant individual differences among commercial drivers and the complex and varied nature of vehicle types and operating scenarios, the same warning may exhibit different effectiveness under different "person, vehicle, road, and environment" conditions: it may prompt drivers to take timely evasive action after the warning is triggered, or it may fail due to ignoring the warning, delayed reaction, or insufficient reaction. The essence of these differences lies in whether the driver can generate a risk-avoidance behavior within a limited time after the warning is triggered, and the intensity level of this reaction; both of these factors jointly characterize the effectiveness of the warning. Judging and predicting this effectiveness can provide a basis for adjusting the intensity and intervention of the warning strategy as needed. When the predicted driver warning effectiveness is low, the warning can be strengthened in a timely manner and necessary auxiliary control measures can be taken to form a safety net; when the effectiveness is high, the conventional strategy should be maintained to avoid unnecessary interference. Without the ability to judge and predict the effectiveness of early warnings, the system struggles to determine whether established strategies can mitigate risks. Furthermore, early warning strategies cannot be dynamically adjusted according to changes in driver status and scenarios, easily leading to an imbalance between "over-intervention" and "lack of backup in case of failure." On the one hand, strong prompts and interventions continue even when the driver is capable of making a sufficient response, increasing driving interference and reducing system acceptability. On the other hand, strategies cannot be strengthened in a timely manner when the driver is unable to make an effective response, resulting in the continuous accumulation of risks and weakening the actual prevention and control effect of the active safety system.

[0004] Meanwhile, the National Road Transport Dynamic Monitoring and Supervision Platform continuously accumulates high-frequency trajectory data and early warning event data of operating vehicles. It can obtain static information related to "people, vehicles, roads, and environments" and vehicle time-series trajectories, providing a data foundation for early warning effectiveness judgment and prediction. It is important to note that the training of early warning effectiveness judgment and prediction models highly depends on the quantitative representation of changes in vehicle motion states before and after an early warning is triggered. This requires extracting key response signals such as velocity, acceleration, and angular velocity from trajectory time-series data and constructing judgment criteria and training samples accordingly. However, due to factors such as sensor accuracy, environmental obstruction, and signal transmission delays, trajectory data generally suffers from high-frequency noise, random frame drops, and abnormal mutations. These defects are easily superimposed in complex scenarios, easily causing distortion or loss of response signals, leading to biases in early warning effectiveness judgment, further affecting the training quality and output accuracy of prediction models, and reducing the accuracy and reliability of early warning effectiveness prediction.

[0005] Therefore, it is necessary to establish a quantitative representation and prediction method for the effectiveness of early warnings for commercial drivers based on high-quality trajectory time-series data and by integrating the static characteristics of people, vehicles, roads, and the environment at the time of warning triggering. Based on the prediction results, the intensity of the prompts and the strength of the intervention under different warning types can be dynamically adjusted to achieve personalized early warning intervention for commercial drivers, thereby improving the effectiveness of active safety early warnings for commercial vehicles, reducing the risk of accidents caused by warning failures, and ensuring road traffic safety.

[0006] Research on predicting the effectiveness of warnings for commercial drivers is relatively limited. Existing similar studies typically use "whether the driver reacts" and "the intensity of the reaction" after a warning is triggered as target variables, and modeling is mainly based on static information at the moment the warning is triggered. For example, in 2019, Xu Yueru of Southeast University published a doctoral dissertation entitled "Analysis and Prediction of the Influence of Active Safety Systems on the Behavior of Commercial Vehicle Drivers." This dissertation selected factors such as vehicle type, driver age, weather conditions, road type, warning duration, warning time, and warning speed. First, it used statistical tests to identify variables that were significantly related to driver reactions. Then, it used random forest to screen key influencing factors and constructed a BP neural network model to predict the reactions and reaction intensity of different drivers to different types of warnings under different road and environmental conditions. While the aforementioned studies have achieved reaction discrimination and reaction intensity prediction to some extent, providing a foundation for evaluating the effectiveness of early warnings, the following shortcomings remain: First, the characterization of reaction intensity does not fully consider the differences in natural driving fluctuation levels among different drivers and under different road conditions. It directly uses the absolute value of acceleration or angular velocity as the magnitude of the reaction, which can easily misinterpret natural fluctuations caused by differences in scenario backgrounds such as slope and curves or driving styles as early warning reactions, resulting in limited transferability and generalization ability of the indicators. Second, the discrimination of "whether to react" relies on fixed threshold rules. The threshold setting lacks data support and the constraints on the time window for determining the occurrence of the reaction are insufficient, which can easily lead to missed or incorrect judgments. Third, modeling methods such as BP neural networks are essentially static mappings of static features and instantaneous states at the moment of early warning triggering. They do not fully utilize the continuous evolution information of trajectory time series data before the early warning, making it difficult to depict the dynamic process of the driver from perception to operation, and thus difficult to support stable and reliable early warning effectiveness discrimination and prediction.

[0007] Existing research on microscopic vehicle trajectory reconstruction methods can be broadly categorized into three types: statistical inference-based model methods, geometric and numerical computation-based interpolation methods, and data-driven deep learning methods. Regarding statistical models and interpolation, Chinese patent CN120256897A discloses a "sub-second vehicle trajectory reconstruction method based on high-frequency satellite positioning data," which utilizes the self-similarity of trajectory signals through fractal interpolation for encryption, combined with Butterworth filtering for noise reduction, smoothing, and equidistant alignment. This method has low computational complexity and strong physical interpretability. However, under the complex operating conditions of commercial vehicles, defects such as noise, frame drops, and anomalous mutations often overlap, limiting the accuracy and detail fidelity of the fixed interpolation and filtering framework. This can easily lead to over-smoothing, loss of detail, or anomalous amplification, thus affecting the reliability of downstream early warning effectiveness judgment and prediction. In the area of ​​deep learning, Chinese patent CN116578661B proposes "a vehicle trajectory spatiotemporal reconstruction method and system based on an attention mechanism," which embeds the current trajectory spatial sequence and historical trajectories, and infers the time series through an attention-based temporal model to fuse and obtain the reconstructed trajectory. These methods have advantages in fitting complex spatiotemporal relationships and reconstruction accuracy, but they mainly rely on black-box models for mapping, lack explicit physical constraints, and have relatively limited physical interpretability. Under conditions of superimposed defects, they may produce deviations that violate the laws of vehicle motion, thus affecting the determination of the effectiveness of warnings and the accuracy of predictions. Therefore, existing trajectory reconstruction methods generally face a trade-off between "reconstruction accuracy" and "physical interpretability": interpolation or filtering methods have strong physical interpretability, but in complex scenarios with superimposed defects such as noise, frame drops, and anomalous changes, reconstruction accuracy and detail fidelity are limited, and they often rely on fixed repair frameworks and parameters, making it difficult to adapt to the diversity and time-varying nature of trajectory defects. While data-driven deep learning methods can improve fitting ability and reconstruction accuracy, they are usually black-box models, lacking explicit kinematic boundary constraints. Their reconstruction results have relatively insufficient physical consistency, interpretability, and controllability, and the model training and deployment costs are also high.

[0008] In summary, the main problems faced by existing methods for judging and predicting the effectiveness of warnings for commercial drivers are: (1) Existing methods for judging the effectiveness of warnings mostly rely on fixed rules, making it difficult to simultaneously characterize the differences in the reaction mechanisms of different warning types and adapt to individual and scenario heterogeneity; (2) There is a lack of a high-quality trajectory reconstruction method for complex trajectory defects, making it difficult to provide a continuous and reliable time-series data foundation for judging and predicting the effectiveness of warnings; (3) It is difficult to achieve dynamic adjustment and personalized intervention of warning strategies. Therefore, there is an urgent need for a vehicle control method and system that considers the effectiveness of warnings for commercial drivers, which is of great practical significance for improving the effectiveness of proactive safety warnings and reducing the occurrence of traffic accidents. Summary of the Invention

[0009] In view of the shortcomings and deficiencies of the existing technology, the present invention provides a vehicle control method and system that considers the effectiveness of warnings to operating drivers. The method takes the prediction result of the effectiveness of driver warnings as the core of decision-making, and dynamically adjusts the intensity of the warning and the intensity of intervention according to the sufficiency of the predicted response. This realizes the transformation of the control paradigm from "passive triggering based on operating condition thresholds" to "active collaborative intervention based on driver response prediction", improves the pertinence and stability of the warning strategy, and ensures safe driving.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A vehicle control method and system considering the effectiveness of warnings for commercial drivers, the method comprising the following steps:

[0012] Step S1. Acquire vehicle trajectory data, early warning event data, and static feature data;

[0013] Step S2. Construct a trajectory reconstruction and repair model, train it based on the speed data in the vehicle trajectory data, and use the trained trajectory reconstruction and repair model to output the reconstructed speed. Then, backtrack and update the position and heading angle to obtain the reconstructed trajectory. The trajectory reconstruction and repair model is an adaptive reinforcement learning trajectory reconstruction model based on physical feasible domain constraints.

[0014] Step S3. Establish a system for judging the effectiveness of early warnings, and obtain a set of early warning effectiveness characterization results for each early warning event;

[0015] Step S4. Construct a prediction model for the effectiveness of warnings for operating drivers. Train the model based on the reconstructed kinematic data, static feature data, warning type, and the warning effectiveness representation results determined in step S3 within a fixed short window before the warning is triggered. Then, use the trained prediction model to output the predicted value of the warning effectiveness for each candidate warning type.

[0016] Step S5. Determine the warning effectiveness level based on the predicted value of the warning effectiveness level. For different warning types, adopt different vehicle control strategies according to the warning effectiveness level.

[0017] As a preferred embodiment of the present invention, the vehicle trajectory data includes positioning time, speed, heading angle, and latitude and longitude; the warning event data includes warning trigger time and warning type; wherein, the warning type includes forward collision warning, speeding warning, and lane departure warning; the static feature data includes driver information, vehicle type, road type, and meteorological environment data.

[0018] As a preferred embodiment of the present invention, the method for constructing the adaptive reinforcement learning trajectory reconstruction model based on physical feasible region constraints is as follows: first construct the state space. and action space ;in, Represents the velocity sequence within the window; This represents the sequence of defect markers within the window; Represents the acceleration feature sequence within the window; action Lagrange interpolation, kinematic low-pass filtering, and robust linear bridging operators are used to output a candidate repair velocity sequence within the output window after the action is executed. Then, based on the acquired candidate repair velocities and physical feasible region constraints, the original observation velocity of the defect points in the center area of ​​each window is updated while keeping the original observation velocity of normal and reliable points unchanged, thus obtaining a global reconstruction velocity sequence. A reward function is designed to calculate the immediate reward after obtaining the reconstruction velocity. Based on the samples in the empirical replay buffer, a dual-depth Q-network is trained to obtain a reinforcement learning policy model for trajectory window defect repair, namely the trajectory reconstruction and repair model.

[0019] As a preferred embodiment of the present invention, the method for constructing the driver warning effectiveness prediction model is as follows: First, the reconstructed kinematic data is processed based on Transformer to extract temporal motion features; static feature data is divided according to semantic source and embedded and mapped to obtain semantic tokens of the same dimension; then, semantic tokens from different sources are combined into a static token sequence; warning types are embedded and mapped to obtain warning type tokens, and gating vectors are generated from the warning type tokens; the static token sequence is fused with type conditionalization to obtain a type-conditionalized static token sequence; cross-modal attention is used to adaptively retrieve the cross-modal static context vector most relevant to the current motion state from the type-conditionalized static tokens, and a channel-gating vector is introduced to control the injection intensity; finally, the static context is injected into the temporal hidden state in the form of gating residuals to obtain the fused time-by-time representation; then, the temporal-level representation is aggregated into a fixed-length event-level representation, and the event-level representation vector is used to generate the final time-by-time representation. A joint output head is constructed, which simultaneously outputs the probability of reaction occurrence, the intensity of conditional reaction, and its fluctuation scale, and the predicted value of the effectiveness of the early warning is calculated based on this.

[0020] As a preferred embodiment of the present invention, in step S5, for forward collision warning and overspeed warning, when the warning effectiveness level is 2, a voice reminder is given to the driver; when the warning effectiveness level is 1, a voice reminder is given to the driver, and a visual reminder is superimposed on the central control display screen, while the throttle opening is reduced to its original value. When the warning effectiveness level is 0, a voice reminder is given to the driver, combined with tactile feedback from seat and steering wheel vibrations, while the accelerator pedal opening is reduced to its original value. And increase the electronic braking force to the original level. times;

[0021] For lane departure warning, when the warning effectiveness level is 2, a voice reminder is given to the driver; when the warning effectiveness level is 1, a voice reminder is given to the driver, and a visual reminder is superimposed on the central control display screen, while tactile reminders are provided in conjunction with steering wheel vibration; when the warning effectiveness level is 0, a voice reminder is given to the driver, and tactile reminders are provided in conjunction with seat and steering wheel vibration, while also providing return-to-center damping and return-to-center torque to guide the driver to complete the correction.

[0022] As a further preferred embodiment of the present invention, the candidate repair velocity generated by the action is mapped to a reconstruction velocity that satisfies the velocity boundary and the acceleration boundary through a recursive dynamic feasible region projection, and the expression is:

[0023] ;

[0024] ;

[0025] ;

[0026] in, Indicates the first The corresponding trajectory for each warning event is in the index. The reconstruction speed value at that location, Indicates the first The corresponding trajectory for each warning event is in the index. Candidate repair speed at the location These are the lower and upper bounds of velocity, respectively. For the physical limit of acceleration, The time interval between adjacent positioning, Indicates will Cut off to interval Inside, These are the upper and lower boundaries.

[0027] As a further preferred embodiment of the present invention, instant feedback The expression is:

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] in, These are the weighting coefficients. This indicates that the reference velocity value was obtained after rigorous physical verification and smoothing. Indicates the number of points in the central area. Index after trajectory reconstruction longitudinal acceleration at that point Reconstructing error terms for the central region; The speed of motion after trajectory reconstruction. For the smoothness term of the central region, The central area of ​​the window The number of times the internal candidate repair speed triggers physical constraint projection truncation.

[0034] As a further preferred embodiment of the present invention, the expression for the cross-modal static context vector is:

[0035] ;

[0036] ;

[0037] ;

[0038] in, Indicates time Attention weights for the four sets of static tokens This is the cross-modal static context vector obtained through aggregation. This represents the dimension of the hidden state vector; It is a learnable mapping matrix; For a moment The query vector, and These are the key matrix and value matrix obtained by mapping four sets of static tokens, respectively. Represents the transpose symbol. It is a static token sequence. Indicates the first The trajectory short-window sequence corresponding to each early warning event is in the... The hidden state vector obtained by extracting temporal features from each sampling point.

[0039] As a further preferred embodiment of the present invention, the fused time-by-time representation for:

[0040] ；

[0041] ；

[0042] in, For channel gating vectors, This represents vector concatenation. For the mapping matrix, These are bias terms, all of which are learnable parameters.

[0043] As a further preferred embodiment of the present invention, the event-level representation vector... The expression is:

[0044] ;

[0045] ;

[0046] ;

[0047] in, For a moment Event-level attention weights Indicates time Score the importance of event-level prediction; Mapping matrix For bias vector This is the projection vector.

[0048] As a further preferred embodiment of the present invention, the predicted value of the effectiveness of the early warning is... ;in, The probability of the reaction occurring. The strength of the conditional reaction.

[0049] This invention also provides a vehicle control system for implementing the above-described control method, considering the effectiveness of driver warnings. The system includes a data acquisition module, a trajectory reconstruction module, a warning effectiveness prediction module, a central control module, an in-vehicle communication module, a throttle control module, a braking control module, a voice control module, a central control screen control module, a vibration control module, and a steering control module. The data acquisition module acquires driver, vehicle, road, and environmental data from a road transport dynamic monitoring and supervision platform. The trajectory reconstruction module processes vehicle trajectory data and provides the reconstructed trajectory data to the warning effectiveness prediction module. The warning effectiveness prediction module integrates vehicle trajectory time-series data with static information on "people, vehicles, roads, and the environment," respectively... The system predicts the effectiveness of driver warnings under different warning types and outputs the corresponding warning effectiveness level. The central control module receives data uploaded by the data acquisition module and issues control strategies to the vehicle based on the driver's warning effectiveness level. The vehicle communication module receives the control strategies issued by the central control module. The throttle control module controls the throttle opening. The brake control module controls the electronic power assist provided by the electro-hydraulic braking system. The voice control module controls the vehicle's voice system. The central control screen control module controls the display content of the vehicle's central control screen. The vibration control module controls the vibration of the steering wheel and seat. The steering control module controls the return damping and torque of the steering wheel.

[0050] Advantages and beneficial effects of the present invention:

[0051] (1) To address the issues of noise interference, random frame loss, and abrupt changes in the trajectory data of operating vehicles, which often overlap under complex operating conditions, this invention proposes an adaptive reinforcement learning trajectory reconstruction method based on physical feasible domain constraints. This method models the trajectory repair process as a Markov decision process of "defect morphology recognition - repair strategy selection - reward feedback," constructing an action set composed of Lagrange interpolation, kinematic low-pass filtering, and robust linear bridging, enabling the agent to adaptively select the optimal repair strategy for different defect morphologies. Furthermore, this invention introduces a central region write-back mechanism and physical constraint projection: on the one hand, write-back updates are performed only in the central region of the window, taking advantage of the more complete context information in the central region to avoid repair deviations caused by insufficient context at the window boundaries, and reducing repeated modifications and coverage conflicts at boundary points, thereby improving global stitching stability and effectively suppressing oscillations and discontinuities caused by stitching at both ends; on the other hand, the repaired velocity sequence is projected into the physical feasible domain jointly defined by the velocity boundary and the acceleration boundary, ensuring that the reconstruction result satisfies the kinematic boundary constraints and maintains physical rationality. Simultaneously, a physical violation penalty term is set in the reward function to guide the model to actively move towards the physically feasible region during the generation phase, forming a two-layer mechanism of "hard constraints guaranteeing feasibility and soft penalties promoting endogenous feasibility": the projection stage ensures that the final output always meets the velocity and acceleration boundary constraints, while the penalty term prompts the agent to reduce its dependence on projection truncation and improve the endogenous feasibility of candidate results, thereby reducing the risk of step jumps and splicing discontinuities caused by frequent truncation. The above mechanism, together with the reconstruction error term and the smoothness term, achieves a joint trade-off between reconstruction accuracy, smoothness, and physical consistency, thus improving reconstruction accuracy while taking into account local motion trends and microscopic details, making the reconstructed trajectory have good physical interpretability and controllability.

[0052] (2) To address the lack of a comprehensive method for determining the effectiveness of early warnings in existing technologies, this invention proposes a method for determining the effectiveness of early warnings based on a threshold and time window synergy using natural fluctuation baseline calibration. First, at the reaction judgment logic level, using the early warning trigger time as a benchmark, a short-term trajectory before the trigger is extracted as a benchmark operating condition segment. The average longitudinal motion level and average lateral motion level of this segment are statistically obtained as the background driving fluctuation benchmark for the driver in the current road segment and traffic environment. Based on this, the longitudinal deceleration deviation and lateral steering deviation of the trajectory after the alarm relative to this benchmark are calculated, thereby reducing the background influence of driving style differences and road alignment conditions (such as slope, curves, etc.) on the reaction judgment, making the judgment result more focused on the actual operational changes after the early warning is triggered. Second, at the time domain search level, addressing the risk of missed detection and false detection caused by the reliance on experience-based setting of the judgment time window length in existing technologies, this invention, based on the statistical law of reaction delay of historical samples, aims to maximize the effective reaction coverage and minimize misjudgment interference, and data-drivenly determines the optimal judgment time window. This mechanism can reasonably limit the reaction search range in the time domain, reducing judgment bias caused by differences in reaction delay or natural operational fluctuations, thereby improving the consistency and robustness of judgment results under different scenarios and warning types. Finally, at the quantitative grading level, to avoid limited applicability due to the use of fixed absolute thresholds, this invention uses the statistical quantile of natural driving fluctuations before the alarm as the threshold benchmark to construct a relative threshold criterion, dividing the reaction intensity into three levels: "no reaction—weak reaction—strong reaction," and outputting the warning effectiveness level accordingly. This grading standard can adaptively calibrate according to the driver's background fluctuation level, thereby improving the stability of evaluation results under cross-driver and cross-scenario conditions.

[0053] (3) To address the problem that existing early warning effectiveness prediction models mainly rely on static features and do not make sufficient use of trajectory time-series information before the early warning is triggered, resulting in insufficient prediction accuracy and reliability, this invention proposes an early warning effectiveness prediction method that integrates trajectory time-series data with static features such as "people, vehicles, roads, and rings". This method uses the short-window kinematic sequence before the early warning as the time-series backbone input, explicitly introduces the early warning type, and semantically represents the static information of "people, vehicles, roads, and rings" and incorporates it into the unified modeling. Through type-guided static semantic selection and stable fusion mechanism, it can effectively adapt to the differences in response mechanisms of different early warning types, suppress the interference of redundant or noisy static information on model training, and thus improve the fusion and utilization effect of static information.

[0054] (4) To address the problem that existing technologies simplify early warning effectiveness prediction to a single regression or single classification, making it difficult to adapt to complex reaction data distributions, this invention innovatively constructs a joint output modeling architecture for zero-inflation data, based on the true statistical characteristics of driver reaction data. This prediction task exhibits typical "two-stage mixed distribution" characteristics: on the one hand, there are a large number of zero values ​​with no significant reaction, i.e., the reaction intensity is 0; on the other hand, when a reaction occurs, the reaction intensity usually exhibits a significantly right-skewed and long-tailed distribution, with a small number of strong reactions corresponding to a relatively large range of values, resulting in overall data simultaneously possessing sparsity, class imbalance, and heteroscedasticity. Based on this, this invention decouples the prediction objective into two sub-tasks: one is reaction occurrence probability estimation, used to characterize the mechanism of whether an effective reaction occurs; the other is conditional intensity regression, used to predict the magnitude of the reaction intensity under the condition that a reaction occurs. Furthermore, a fluctuation scale parameter is introduced to characterize the log-normal distribution of the conditional intensity, achieving simultaneous modeling of the "mean intensity" and "fluctuation uncertainty." This zero-inflation joint modeling approach can effectively alleviate the mean smoothing effect caused by the high proportion of zero values ​​in traditional single-task regression, while avoiding the problem of losing intensity details when only performing classification prediction. Thus, it can jointly characterize the two types of information, "whether to respond" and "response intensity", and significantly improve the accuracy, stability and interpretability of prediction results in sparse response and long-tail intensity scenarios.

[0055] (5) In order to solve the problem that existing active safety systems mainly rely on fixed physical risk index thresholds for mechanical triggering, which makes it difficult to balance "over-intervention" and "lack of backup after warning failure", this invention proposes a human-machine collaborative hierarchical intervention method based on warning effectiveness prediction: the system takes the driver's warning effectiveness prediction result as the decision core, and dynamically adjusts the prompt intensity and intervention intensity according to the sufficiency of the predicted response. Specifically, when the predicted driver reaction is sufficient, conventional voice prompts are used as the primary method while maintaining driver control to reduce unnecessary system interference. When the predicted reaction is insufficient, visual prompts and necessary tactile reminders are added on top of voice prompts. In longitudinal risk scenarios such as forward collision warning and overspeed warning, auxiliary intervention measures such as mild throttle limitation are introduced to guide the driver to make timely corrections. When the prediction is difficult to produce an effective reaction, a strong prompting and strong intervention linkage strategy is adopted. By enhancing voice and tactile warnings and combining them with measures such as enhanced braking, throttle limitation, return damping, and return torque guidance, a safety safety net is formed. This achieves a shift in the control paradigm from "passive triggering based on operating condition thresholds" to "active collaborative intervention based on driver reaction prediction," improving the pertinence, stability, and overall safety of the warning strategy. Attached Figure Description

[0056] Figure 1 A flowchart of a vehicle control method that considers the effectiveness of early warnings from commercial drivers, provided by the present invention;

[0057] Figure 2 This is a schematic diagram of the trajectory reconstruction model of the present invention;

[0058] Figure 3 A schematic diagram of the structure of the early warning effectiveness prediction model provided by the present invention;

[0059] Figure 4 This invention provides a structural block diagram of a vehicle control system that considers the effectiveness of early warning for commercial drivers. Detailed Implementation

[0060] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0061] Example 1:

[0062] like Figure 1 As shown, this embodiment provides a vehicle control method that considers the effectiveness of warnings from commercial drivers. The method includes the following steps:

[0063] Step S1. Data Acquisition and Processing:

[0064] Step S1.1. Obtain data;

[0065] Specifically, vehicle trajectory data, early warning event data, and static feature data are obtained from the road transport dynamic monitoring and supervision platform. The vehicle trajectory data includes positioning time, speed, heading angle, and latitude and longitude, with a sampling frequency of 10Hz. The early warning event data includes early warning trigger time and early warning type; the early warning types include forward collision warning (FCW), overspeed warning (OSW), and lane departure warning (LDW). The static feature data includes driver information, vehicle type, road type, and meteorological environment data; the driver information includes age and driving experience; the vehicle type includes scheduled passenger transport, tourist charter buses, and hazardous materials transport vehicles; the road type includes elevated and expressways, urban arterial roads, urban secondary arterial roads, and urban branch roads; the meteorological environment data includes weather conditions, temperature, wind speed, precipitation, and snow accumulation, with weather conditions including sunny, cloudy, and rain / snow.

[0066] Step S1.2. Process the data:

[0067] With the first One warning event Triggering time Based on this, the early warning event data and vehicle trajectory data are time-aligned and correlated, and the interval is extracted. The trajectory points within the area constitute the trajectory sequence before and after the warning is triggered. , Trajectory points The expression is:

[0068]

[0069] in, Indicates the location time; Indicates location information; Indicates speed (m / s); Indicates the heading angle (°). These are latitude and longitude coordinates, where Longitude Latitude.

[0070] At the same time, static features such as drivers, vehicles, roads, and weather conditions are encoded into static feature vectors. This forms the data sample corresponding to a single early warning event. ,remember For the first The warning type for each warning event is identified. This provides a data foundation for subsequent trajectory reconstruction, assessment of warning effectiveness, and prediction.

[0071] Step S2. Establish a trajectory reconstruction and repair model based on reinforcement learning, and obtain the reconstructed trajectory based on the constructed trajectory reconstruction and repair model:

[0072] This step uses trajectory sequences The velocity components in the input are denoted as the velocity subsequence. The trajectory reconstruction and repair model outputs the repair results, and the reconstructed velocity sequence is obtained. Then, the position sequence and heading angle sequence are further backtracked and updated to obtain the reconstructed trajectory. .

[0073] Furthermore, such as Figure 2 As shown, in this embodiment, the trajectory reconstruction and repair model is an adaptive reinforcement learning trajectory reconstruction model based on physical feasible region constraints. The specific construction method is as follows:

[0074] Step S2.1. Markov Decision Process Modeling:

[0075] The trajectory reconstruction and repair process is represented as a transfer sample. ;in, This represents the state characterization of the velocity defect morphology within the current sliding window. This indicates the repair action to be taken for the current window. This indicates the immediate reward obtained by performing the action. This represents the next state after an action is performed. The action value function is learned. To obtain the optimal repair action under different defect morphologies:

[0076]

[0077] Step S2.2. Establish the trajectory reconstruction model:

[0078] Step S2.2.1. Construct the state space:

[0079] Use a length of sliding window right Perform segment-by-segment repair, take This indicates the number of trajectory points contained within the window. To characterize the defect morphology within the window, each sampling point is first... Construction defect marking ,in Indicates the first Each speed point is a defect point. Indicates the first Each velocity point is considered a normal, reliable point. Specifically, a point is judged to be defective if it meets any of the following conditions: (1) (2) The value is null or invalid. ;in, and These are the lower and upper bounds of the velocity, respectively; (3) ;in, The physical limit of acceleration is acceleration. Calculated using the velocity difference between adjacent time points:

[0080]

[0081] in, This represents the time interval between adjacent positioning.

[0082] Based on this, the first A sliding window Constructing the state space:

[0083]

[0084] in, Represents the velocity sequence within the window; This represents the sequence of defect markers within the window; This represents the acceleration feature sequence within the window.

[0085] Step S2.2.2. Construct the action space:

[0086] Constructing a collection of actions:

[0087]

[0088] Output window after action is executed Candidate repair velocity sequence within ,in , Indicates the first Each warning event is indexed in the window. Candidate repair speed value at the location, Display window The set of indices.

[0089] (1) Action —Lagrange interpolation operator:

[0090] Define the set of trusted anchor indexes within the window. And sort the anchor points by time as follows ,in The number of anchor points. For any index and The defects, in order to locate the time As the independent variable, Lagrange interpolation provides the candidate repair speed:

[0091]

[0092] in, anchor point The velocity observation value at the location. This action is suitable for repairing trajectory defects with sufficient anchor points and short gaps, used to restore local motion trends and preserve micro-motion details.

[0093] (2) Actions —Kinematic low-pass filter operator:

[0094] To ensure the continuity of the filtered input, a temporary continuous sequence is constructed. , among which when season ;when Linear prefilling is then applied to this location to obtain... Based on this, a low-pass filter operator is applied. Candidate outputs are obtained:

[0095]

[0096]

[0097] in, This represents a convolutional smoothing low-pass filter operation; express Middle and Index The corresponding element. This action is suitable for repairing trajectory defects where velocity observations are basically continuous but contain high-frequency noise or anomalous changes, used to suppress high-frequency noise and local oscillations and improve the smoothness and stability of the trajectory.

[0098] (3) Actions — Robust linear bridging operator:

[0099] Define the left and right endpoint indices of the window as and ,in And the endpoints are trustworthy points. For any index and Defect point calculation uses linear interpolation at both endpoints to obtain candidate repair speeds:

[0100]

[0101] in, The endpoint velocity value. For positioning time. This action is suitable for repairing trajectory defects with long gaps or insufficient reliable anchors, providing stable, interpretable, and controllable candidate connections through linear bridging of the trends of the two endpoints.

[0102] Step S2.2.3. Central area write-back and physical constraints:

[0103] (1) Central area write-back mechanism:

[0104] To avoid sliding windows The candidate repair speed at the boundary is affected by incomplete context, resulting in splicing oscillations. A center region write-back mechanism with overlapping sliding windows is adopted. A center region index set is defined. ;in, This represents the middle subset of the indexes used for write-back within the window, with a length of... .

[0105] Constructing a global reconstruction speed sequence ;in, Indicates the first The corresponding trajectory for each warning event is in the index. The reconstruction speed value at each window. Only for the central area The point within the write-back update is executed when At that time, based on the candidate repair speed And physical feasible region constraints to update the original observation rate , and record as ;when Maintain the original observation speed This avoids unnecessary modifications to reliable observations and reduces the risk of error propagation. After the write-back is complete, the defect markers of repaired points within the central region are updated to... This is to avoid subsequent windows repeatedly repairing the same point.

[0106] (2) Physical constraint projection:

[0107] Perform physical constraint projection on candidate repair velocities in the central region. For any The candidate repair speed generated by the action The recursive dynamic feasible region projection mapping is used to reconstruct the velocity that satisfies the velocity and acceleration boundaries. The key to this projection lies in the upper and lower bounds of the constraints at each moment. and It is not a fixed threshold, but rather the result projected from the previous time step. Combined with vehicle maximum acceleration constraints and time interval The system recursively updates point by point, thus constructing a dynamic physically feasible region that updates progressively with the repair results of the previous time step. A time-by-time consistency constraint is imposed on the candidate repair velocity sequence to ensure that the entire sequence satisfies the kinematic boundaries. The projection adopts a recursive truncation form.

[0108]

[0109] in, Indicates will Cut off to interval Inside, upper and lower boundaries Defined as:

[0110]

[0111]

[0112] Through the above projection, the candidate repair results generated by any action are always constrained within an explicitly controllable dynamic physical feasible domain, thereby improving the reconstruction accuracy while ensuring that the velocity sequence satisfies the kinematic boundary constraints and has physical rationality and controllability.

[0113] Step S2.2.4. Reward Function Design:

[0114] In this embodiment, the sliding window After performing the repair action and physical constraint projection, the reconstruction speed is obtained. Calculate the instant return accordingly. Define the central region reconstruction error term:

[0115]

[0116] in, Indicates the number of points in the central area. This indicates that the reference speed value was obtained after rigorous physical verification and smoothing.

[0117] Furthermore, to suppress unnecessary drastic fluctuations, a smoothness term for the central region is defined. :

[0118]

[0119]

[0120]

[0121] in, The longitudinal acceleration after trajectory reconstruction; The jerk (Jerk) is the jerk after trajectory reconstruction.

[0122] Finally, to further suppress non-physical trajectories caused by candidate repair results deviating from the vehicle's feasible motion boundary, this invention introduces a physical violation penalty term into the reward function and defines a sliding window. central area The number of times the internal candidate repair rate triggers physical constraint projection truncation is: The design concept is that physical constraint projection is used to force the candidate repair velocities to be mapped into a dynamic physically feasible domain that satisfies both velocity and acceleration boundaries. This is used to quantify the degree to which "candidate results need to be forced to correct". Specifically, when the candidate repair speed frequently touches or exceeds the dynamic feasible region, the projection will intervene multiple times and produce significant results. Conversely, if the candidate repair speed is naturally located within the feasible region, then projection intervention is less necessary. The penalty is relatively small. Unlike static penalties that only apply to single-point exceedances, this penalty term transforms the model's dependence on projection truncation into a countable and cumulative sequence-level index, enabling the reward signal to directly characterize the inconsistency intensity between the candidate repair result and the vehicle kinematic boundary.

[0123] Furthermore, since this invention employs a recursive dynamic feasible region projection, the upper and lower bounds of the feasible region at each time step within the central region are recursively updated based on the repair state of the previous time step. It measures not an occasional out-of-bounds event at a particular moment, but rather the overall degree of conformity and physical consistency of the candidate repair velocity sequence with dynamic constraints throughout the entire central region.

[0124] This leads to the overall feedback:

[0125]

[0126] in, For the weighting coefficients, take respectively , and This is used to balance the constraints of reconstruction accuracy, smoothness, and physical consistency, and to scale the differences in dimensions among different terms. Through the above reward design, the agent, while improving the reconstruction accuracy of the central region, will tend to generate smoother candidate repair results that satisfy the kinematic boundaries, thereby enhancing the interpretability and controllability of the reconstruction process.

[0127] Step S2.3. Model Training:

[0128] Transfer sample Store in the experience replay buffer ,in For the first The state representation of a sliding window; The repair actions taken for this window; For immediate feedback after the action is performed, This is the next state after performing an action. During training, it is retrieved from the experience replay buffer. Random sampling (e.g., 64) mini-batch samples are used for training to construct an online action value network. With the target network ,in For online action value network parameters; For target network parameters; This indicates the current window's state; 'a' indicates the action to be taken on the current window.

[0129] For the first batch of samples Sample If the next state If the state is terminated, the objective value is the immediate reward; otherwise, the temporal difference objective value is constructed using the DDQN (Double Deep Q-Network) method.

[0130]

[0131] in, As a discount factor, For a set of actions, These are candidate action variables used for action enumeration.

[0132] The loss function is defined using mean squared error:

[0133]

[0134] And update the online network parameters using gradient descent. To maintain the stability of the target valuation, every [period]... The step is to synchronize the online network parameters to the target network:

[0135]

[0136] Repeat the above iterative process until the model converges, thereby obtaining the reinforcement learning policy model for trajectory window defect repair, i.e., the trajectory reconstruction and repair model.

[0137] Step S2.4. Reconstruct the trajectory sequence:

[0138] Regarding the first Trajectory sequence corresponding to each early warning event , with its velocity subsequence The trajectory reconstruction and repair model takes the input as input and outputs the repair result. Then, by further backtracking, the reconstructed position sequence and heading angle sequence are obtained, thereby reconstructing the entire trajectory. Let the reconstructed trajectory points be:

[0139]

[0140] in, To reconstruct latitude and longitude coordinates, This indicates the reconstructed heading angle.

[0141] To facilitate discrete integral updates, latitude and longitude are mapped within a local neighborhood using coordinate mapping. Convert to local plane coordinates:

[0142]

[0143] in, and These are the horizontal and vertical plane coordinates, respectively. With 0° due north and 90° due east, the reconstructed position is calculated using the following formula:

[0144]

[0145] Then through inverse mapping Will Reconstruct latitude and longitude:

[0146]

[0147] Simultaneously, the heading angle sequence is extrapolated back based on adjacent positions:

[0148]

[0149] in, For the four quadrant arctangent function; defined Map the angle to Used to standardize the range of heading angle values ​​and avoid crossing over [a certain range]. The expression "time" is ambiguous.

[0150] Get the starting index of the window Using the original observations as initial values, the reconstructed trajectory sequence is obtained by back-deriving from the above formula. This is used for subsequent determination of the effectiveness of early warnings and predictive modeling.

[0151] Step S3. Establish a system for determining the effectiveness of early warnings:

[0152] Step S3.1. Establish reaction characterization indicators:

[0153] Different warning signals correspond to different risk scenarios, and drivers adopt different risk avoidance strategies. Forward Collision Warning (FCW) and Overspeed Warning (OSW) often require drivers to control the vehicle longitudinally by releasing the accelerator and braking to adjust the vehicle's longitudinal running state. Lane Departure Warning (LDW), on the other hand, requires drivers to turn the steering wheel to correct the steering, thereby achieving lateral control. Based on these differences, the reconstructed trajectory sequence obtained in step S2... We extract the speed and heading angle changes between adjacent moments before and after the warning is triggered, and construct longitudinal and lateral indicators to characterize the driver's reaction.

[0154] (1) Longitudinal response characterization indicators:

[0155] Pre-baseline window triggered by early warning The average acceleration is taken as the baseline, denoted as . And define baseline correction acceleration :

[0156]

[0157] The baseline correction process described above is equivalent to applying a "high-pass" effect to the longitudinal acceleration sequence: by subtracting the short-window mean, low-frequency background fluctuations caused by long-term differences in driver driving style and road longitudinal slope are filtered out, retaining only the acceleration change signal with abrupt changes after the warning is triggered. Since only changes in the direction of deceleration represent that the driver has adopted a risk avoidance strategy, the longitudinal operation intensity is further defined as follows:

[0158]

[0159] in, Indicates index The longitudinal operating intensity at a given point indicates that a larger value indicates a more significant deceleration operation relative to the baseline.

[0160] (2) Lateral response characterization indicators:

[0161] Define the reconstructed angular velocity as:

[0162]

[0163]

[0164] in, This indicates mapping the angle difference to... Used to eliminate heading angle at Differential jumps caused by nearby crossings; Indicates the difference in angle between adjacent headings; For index The angular velocity after trajectory reconstruction, in deg / s. The baseline window before warning triggering. The average angular velocity is taken as the baseline, denoted as . And define the lateral operation strength as:

[0165]

[0166] in, Indicates index The lateral movement intensity at a given point indicates a more significant lateral correction relative to the baseline. Since this indicator measures the angular velocity deviation from the baseline, it can uniformly quantify the driver's lateral correction actions regardless of whether the vehicle is on a straight or curved road, effectively solving the problem of determining the effectiveness of warnings in curved scenarios.

[0167] Step S3.2. Determine the reaction threshold:

[0168] This step calculates the response threshold based on trajectory data from historical early warning samples. To avoid interference from early warning signals, a baseline time window prior to the early warning trigger is used. The internal reaction characterization index samples were used as a natural driving fluctuation sample set, and then summarized according to the warning type:

[0169]

[0170]

[0171] in, and These represent the longitudinal and lateral natural driving fluctuation sample sets, respectively. Based on this, the empirical distribution quantiles of historical samples are used as thresholds: let the quantile level be... =0.85, then the longitudinal response threshold With lateral reaction threshold They are defined as follows:

[0172]

[0173] in, Indicates the sample set at the quantile level The quantile value below, , The above thresholds can be continuously adjusted as historical warning samples are updated.

[0174] Step S3.3. Determine the reaction determination window length:

[0175] To avoid missed detections due to an excessively short reaction decision window, or misclassification of unrelated actions as reactions due to an excessively long window, this step statistically analyzes the reaction delay distribution after an alert is triggered based on historical data. While ensuring the highest possible effective reaction coverage, the shortest possible decision window is selected to reduce misjudgment interference, thus determining the reaction decision window length. .

[0176] Let the upper limit of the search response time after the warning is triggered be . ,Pick In the interval Internally, the moment when the reaction occurs is defined as the moment of the first threshold crossing. :

[0177]

[0178] If the corresponding set is empty, then the warning event is considered to be in No significant reaction was observed internally. If it exists, the reaction time delay is further defined as follows: For all historical early warning samples Statistical analysis was performed to obtain the frequency distribution and cumulative distribution of reaction time delay. Since a longer decision window increases the probability of incorporating natural driving fluctuations and unrelated warning-related maneuvers, and thus leads to greater misjudgment interference, this step selects the shortest decision window as the decision window, while ensuring a high level of coverage. Based on the statistical results, the reaction determination window length is determined. At that time, FCW could cover 86.5% of significant responses, OSW could cover 87.4%, and LDW could cover 83.5%, thus determining the setting. .

[0179] Step S3.4. Define whether a reaction occurs and the intensity of the reaction:

[0180] (1) Definition of whether it reacts:

[0181] Define whether to react as ,in This indicates that the driver reacted significantly to the warning within the judgment window. This indicates that no significant response was produced.

[0182]

[0183] (2) Definition of reaction intensity index:

[0184] To characterize the magnitude of the effective response, the reaction intensity is defined as... .when When the threshold is first crossed, the response characterization index value is taken as the intensity; when At that time, let the intensity be 0. Specifically, let... This is the index of the first threshold crossing time, which is... The reaction strength is:

[0185]

[0186] Based on the above definition, for each early warning event A set of early warning effectiveness characterization results were obtained for each. ,in Used to indicate whether an alert triggers an effective response. Used to indicate the intensity of conditional response for different warning types.

[0187] Step S3.5. Define the effectiveness level and grade of the early warning:

[0188] To comprehensively characterize whether and how strong a response is triggered after an early warning is given on the same scale, let the first... One warning event The effectiveness of the early warning is Defined as:

[0189]

[0190] Among them, when hour, This indicates that the warning did not trigger a significant response within the judgment window; when hour, This indicates that the warning triggers a significant response, and its effectiveness is determined by the intensity of the response at the moment the threshold is first crossed.

[0191] To facilitate subsequent adjustments to the early warning strategy based on the effectiveness of the early warning, the effectiveness of the early warning will be assessed. The system is divided into three levels. ,set up This represents the quantile of response intensity for a specific warning type, obtained based on historical natural fluctuation samples. and At that time, respectively obtained and .

[0192] For early warning events Based on its warning type The corresponding thresholds are used to classify the warning effectiveness levels. :

[0193]

[0194] in, These represent the levels of no response, weak response, and strong response to the early warning, respectively, and can serve as the basis for adjusting subsequent early warning strategies. , , The above thresholds can be continuously adjusted as historical warning samples are updated.

[0195] Step S4. Establish a prediction model for the effectiveness of warnings for commercial drivers, and predict the degree of effectiveness of driver warnings based on the constructed prediction model:

[0196] This step is for single warning events. Construct a predictive model for the effectiveness of early warning systems. The model structure is as follows: Figure 3 As shown, the model input consists of three parts:

[0197] (1) Short window timing input: take the fixed short window before the warning is triggered. The kinematic sequence within, for the first Construct a kinematic input vector from each sampling point. :

[0198]

[0199] in, For speed, For heading angle, For acceleration, Let be the angular velocity. Given the sequence length, the short window will contain... The points are stacked to form a dynamic input matrix. ;

[0200] (2) Static input: static feature vector ;

[0201] (3) Warning type input: Warning type .

[0202] Model outputs the probability of reaction occurrence. Prediction of reaction intensity under reaction conditions Based on this, the effectiveness of the early warning can be calculated.

[0203] Step S4.1. Construction of the early warning effectiveness prediction model:

[0204] Step S4.1.1. Temporal motion feature extraction based on Transformer:

[0205] right Sequence modeling is performed to obtain the time-by-time hidden state representation:

[0206]

[0207] in, This represents an encoder consisting of multiple layers of self-attention and feedforward networks stacked together. This represents the dimension of the hidden state vector; Indicates the first The trajectory short-window sequence corresponding to each early warning event is in the... The hidden state vector obtained by extracting temporal features from each sampling point; It is used to characterize the continuous motion trend and key fluctuation features within a short window before the warning is triggered, and serves as the input for subsequent static semantic fusion and joint output head. The self-attention mechanism can adaptively focus on the kinematic changes at key moments within the same window, taking into account both long-term and short-term dependencies and abrupt change pattern extraction, thus providing a robust temporal representation foundation for subsequent prediction.

[0208] Step S4.1.2. Semantic grouping and vectorization mapping of static heterogeneous information:

[0209] To reduce feature confounding caused by differences in the dimensions, semantics, and intensity of static information from different sources, this invention semantically groups heterogeneous static features such as people, vehicles, roads, and rings, and maps them to corresponding group semantic tokens to form structured input representations, providing a foundation for subsequent refined feature interaction and fusion modeling. Divided into four groups according to semantic origin:

[0210]

[0211] in, 、 、 、 These represent the static feature sub-vectors of the driver, vehicle, road, and environment corresponding to the warning event, respectively.

[0212] Embedding mapping is performed on the four sets of static sub-vectors respectively to obtain semantic tokens of the same dimension:

[0213]

[0214] in, Indicates the first Embedding mapping function of group static features, For the first Group static semantic token, =Driver, =Vehicles, =road, =Environment, which combines four static semantic tokens into a static token sequence:

[0215]

[0216] In order to enable cross-modal interactive fusion with time-series representations in subsequent steps, and in the early warning type To achieve differentiated utilization under certain conditions.

[0217] Step S4.1.3. Construct the fusion of early warning type token and type conditional gating:

[0218] The risk formation mechanisms and driver reaction mechanisms differ across different warning types, resulting in varying degrees of contribution from static information. To avoid cross-type averaging and information dilution caused by uniformly fusing multiple types of samples, this invention introduces a warning type token and proposes a warning type-driven adaptive gating mechanism. This mechanism dynamically adjusts the activation intensity of semantic tokens in each group, adaptively weights static semantic information, and injects it into the temporal backbone as residuals. While sharing transferable information, it preserves type-specific expressions, thereby achieving mechanistic modeling that focuses on different static elements in different warning scenarios. This mechanism enables the model to proactively enhance or suppress the contribution of static features of different categories based on the warning type, thereby improving the targeting, stability, and interpretability of the fusion process.

[0219] Warning type The alert type token is obtained through the embedded mapping:

[0220]

[0221] in, For embedded functions, For the first A type representation vector for each early warning event.

[0222] Based on this, a gating vector is generated from the warning type token:

[0223]

[0224] in, For the Sigmoid function; For the mapping matrix, These are bias terms, and all can be learned parameters; The value represents the injection strength coefficient of static semantic differences in fusion. The closer the value is to 0, the weaker the injection of difference information. The closer the value is to 1, the stronger the injection.

[0225] right Perform type-conditional fusion to obtain a type-conditional static token:

[0226]

[0227] in, A shared mapping matrix is ​​used to extract common static semantics that are transferable across types; This is a difference mapping matrix used to extract type-related static semantic differences. This represents element-wise multiplication, used to adaptively adjust the injection strength of differential semantics based on the gating vector. This yields the type-conditionalized static token sequence. :

[0228]

[0229] Used for subsequent and time-series characterization Cross-modal interaction fusion.

[0230] Step S4.1.4. Cross-modal attention interaction and stable gating injection of dynamic and static features:

[0231] This module enables temporal motion features to be queried at each time step using the current hidden state as the query. It adaptively retrieves the most relevant static context information to the current motion state from the type-conditionalized static tokens, thereby improving the effective utilization of static information. Considering that directly injecting static information into the temporal backbone in a strongly coupled manner can easily introduce noise interference and cause unstable fusion in the early stages of training, while weak fusion may lead to insufficient contribution or even neglect of static branches, this step adopts a combination of cross-modal attention and channel-gated residual injection: while selecting key static semantics as needed, the gating adaptively adjusts the injection intensity to form a stable and beneficial fusion process, thereby avoiding static information from interfering with temporal feature extraction and alleviating the problem of insufficient utilization of static information in traditional fusion methods.

[0232] (1) Cross-modal attention:

[0233] For any given moment ,by For querying, use a static token as the key and value:

[0234]

[0235] in, It is a learnable mapping matrix; For a moment The query vector, and These are the key matrix and value matrix obtained by mapping four sets of static tokens, respectively.

[0236] Calculate the cross-modal attention weights and aggregate the static semantics to obtain the cross-modal static context vector:

[0237]

[0238] in, Indicates time Attention weights for the four sets of static tokens This is the cross-modal static context vector obtained through aggregation. This represents the dimension of the hidden state vector. Represents the transpose symbol.

[0239] (2) Stable gated injection:

[0240] To avoid excessive perturbation of the temporal backbone by the static context in the early stages of training, a channel-gated vector is introduced. Controlling injection intensity:

[0241]

[0242] in, This represents vector concatenation. For the mapping matrix, These are bias terms, all of which are learnable parameters. Finally, the static context is injected into the temporal hidden state in the form of gated residuals to obtain the fused time-series representation:

[0243]

[0244] This yields the fused temporal representation sequence:

[0245]

[0246] Step S4.1.5. Event-level attention aggregation of temporal features:

[0247] The model predicts single early warning events. The probability of occurrence and the intensity of conditional responses are event-level outputs. Therefore, time-series representations need to be aggregated into fixed-length event-level representations to highlight the contributions of key moments and avoid information dilution caused by simple averaging. Based on fused sequences Construct event-level attention weights and complete weighted convergence:

[0248]

[0249]

[0250]

[0251] in, Indicates time Score the importance of event-level prediction; Mapping matrix For bias vector These are projection vectors, all of which are learnable parameters; For a moment Event-level attention weights are used to measure the contribution of different time points to the prediction; For a single warning event The event-level representation vector is used for subsequent joint output modules.

[0252] Step S4.1.6. Joint Prediction and Validity Quantification for Zero-Inflation Data:

[0253] After a warning is triggered, the driver may not react significantly, or they may react but with varying intensity; the reaction intensity exhibits zero inflation and a long tail. To simultaneously characterize whether a reaction occurs and the intensity of the reaction, this step is based on event-level representation vectors. A joint output head is constructed to simultaneously output the probability of reaction occurrence, the intensity of conditional reaction, and its fluctuation scale. Based on this, a predicted value of the effectiveness of the early warning is calculated, providing a continuous quantitative basis for subsequent classification and strategy adjustment.

[0254] (1) Output of reaction occurrence probability:

[0255]

[0256] in, Output weight vector for probability. These are bias terms, all of which are learnable parameters; Indicates the first The predicted probability of a significant response to a warning event.

[0257] (2) Conditional response intensity output:

[0258]

[0259] in, ; Output a weight vector for the intensity. These are bias terms, all of which are learnable parameters; Indicates the warning type is And the predicted intensity under reaction conditions.

[0260] (3) Conditional intensity fluctuation scale output:

[0261]

[0262] in, For weight vector, These are bias terms, all of which are learnable parameters; Used to characterize the degree of dispersion of intensity distribution under reaction conditions.

[0263] (4) Calculation of the predicted value of the effectiveness of the early warning:

[0264]

[0265] in, For the first The warning event type is: The predicted value of the effectiveness of the early warning at that time is used for subsequent early warning effectiveness classification and early warning strategy adjustment.

[0266] Step S4.2. Model Training and Parameter Update:

[0267] sample set Divided into training set and validation set, where This represents the total number of samples in the training set. As input to the model, the model output is , , The monitoring signal is whether a reaction label has occurred. With reaction strength Based on the above data, mini-batch stochastic gradient descent was used to train the model end-to-end.

[0268] To simultaneously characterize the structural zero-inflation features of unreacted samples and the continuous long-tail features of reacted samples, Modeled as a zero-inflated log-normal random variable, its probability mass and probability density are written as:

[0269]

[0270] in, Indicates the warning type is Temporal reaction strength The likelihood values; Let be the probability density function of the log-normal distribution. To align with the conditional strength output, let . .

[0271] During the training phase, the negative log-likelihood is used as the loss function. Random samples are taken from the training set each time. The samples constitute a small batch, the first... Loss in batches for:

[0272]

[0273] in, Indicates the warning type is Temporal reaction strength The likelihood value, This indicates a small batch number. This indicates the sample number within that small batch.

[0274] And the set of learnable parameters of the model renew:

[0275]

[0276] in, Let be the learning rate, and take... Repeat the above sampling and update process until the loss converges to obtain the predictive model for the effectiveness of the warning for operating drivers.

[0277] Step S5. Control the vehicle according to the effectiveness level of the warning:

[0278] In practical operation, the warning type is treated as a candidate conditional variable. For each candidate warning type, its predicted warning effectiveness and warning effectiveness level are calculated, thereby enabling dynamic adjustment of the multi-type warning strategy. Let the set of candidate warning types be... For any Output from the prediction model And according to the grading threshold given in step S3.5 , Calculate the effectiveness level of the early warning:

[0279]

[0280] in, This indicates that under the current driving conditions, this type of warning signal is unlikely to elicit an effective response, and there is a significant risk of ignoring it. This indicates that a response is expected but its intensity is insufficient, limiting the effectiveness of early warning and intervention, and risks may still exist. This indicates that the expected response is adequate, the early warning is highly effective, and the risk can be mitigated in a timely manner. The system adjusts its approach based on different early warning types. and corresponding levels The early warning strategy for this type of event will be dynamically adjusted.

[0281] Specifically, in this embodiment, the dynamic adjustment strategy for different warning types is as follows:

[0282] 1. Forward Collision Warning (FCW) tiered strategy:

[0283] (1) If In this case, a standard warning intensity will be used to alert the driver, with a voice prompt saying, "There is a collision risk ahead, please slow down immediately."

[0284] (2) If The warning intensity is increased, with a voice reminder to the driver stating "There is a collision risk ahead, please slow down immediately," and a visual reminder is overlaid on the central control display screen. Simultaneously, the accelerator pedal opening is reduced to its original position. .

[0285] (3) If The system employs a combination of strong prompts and strong interventions. In addition to increasing the volume of the voice prompts, it verbally reminds the driver, "The risk of a collision ahead is high; please slow down immediately," and combines this with tactile feedback through seat and steering wheel vibrations. Simultaneously, it reduces the accelerator pedal opening to its original position. And increase the electronic braking force to the original level. Double the distance to reduce the risk of rear-end collisions.

[0286] 2. Overspeed warning OSW (Overspeed Warning System) tiered strategy:

[0287] (1) If If the speed limit is exceeded, a standard warning will be issued, with a voice reminder to the driver: "The current speed has exceeded the limit, please slow down immediately."

[0288] (2) If The alert intensity is increased, with a voice reminder to the driver stating, "Current speed has exceeded the limit, please slow down immediately," and a visual reminder is overlaid on the central control display screen. Simultaneously, the accelerator pedal opening is reduced to its original position. This is to guide drivers to slow down voluntarily.

[0289] (3) If The system employs a combination of strong prompts and strong interventions, increasing the volume of the voice prompt and reminding the driver, "The vehicle speed is severely exceeded; please slow down immediately." This is combined with tactile feedback through seat and steering wheel vibrations, while simultaneously reducing the accelerator pedal opening to its original position. And increase the electronic braking force to the original level. This is to reduce the risk of continuous speeding and decrease the risk of accidents.

[0290] 3. Lane Departure Warning (LDW) tiered strategy:

[0291] (1) If In this case, a standard warning intensity will be used to remind the driver, with a voice prompt saying, "The vehicle is deviating from its lane; please correct your course."

[0292] (2) If The system will enhance the alert intensity by providing a voice reminder to the driver that "the vehicle is deviating from the lane, please pay attention and correct it," along with a visual reminder on the central control display screen and a tactile reminder via steering wheel vibration to prompt the driver to straighten the steering wheel.

[0293] (3) If The system employs a combination of strong prompts and strong interventions, increasing the volume of voice prompts and reminding the driver that "the vehicle has seriously deviated from the lane, please correct it immediately." It also combines tactile reminders with seat and steering wheel vibrations, while providing return-to-center damping and torque to guide the driver to complete the correction, thereby reducing the lateral risks caused by deviation.

[0294] Example 2:

[0295] like Figure 4 As shown, this embodiment proposes a vehicle control system that considers the effectiveness of warnings for operating drivers in order to implement the control method described above. The system includes a data acquisition module, a trajectory reconstruction module, a warning effectiveness prediction module, a central control module, an in-vehicle communication module, a throttle control module, a braking control module, a voice control module, a central control screen control module, a vibration control module, and a steering control module.

[0296] The data acquisition module is used to acquire driver, vehicle, road and environmental data from the road transport dynamic monitoring and supervision platform.

[0297] The trajectory reconstruction module is used to process vehicle trajectory data and provide high-quality trajectory data (reconstructed trajectory data) to the early warning effectiveness prediction module.

[0298] The warning effectiveness prediction module is used to integrate vehicle trajectory time series data with static information such as "people, vehicles, roads, and rings" to predict the degree of warning effectiveness for drivers under different warning types and output the corresponding warning effectiveness level.

[0299] The central control module is used to receive data uploaded by the data acquisition module and issue control strategies to the vehicle based on the driver's warning effectiveness level.

[0300] The vehicle communication module is used to receive control strategies issued by the central control module;

[0301] The throttle control module is used to control the throttle opening.

[0302] The braking control module is used to control the electronic power assist provided by the electro-hydraulic braking system;

[0303] The voice control module is used to control the vehicle's voice system;

[0304] The central control screen control module is used to control the display content of the vehicle's central control screen;

[0305] The vibration control module is used to control the vibration of the steering wheel and seat;

[0306] The steering control module is used to control the return damping and torque of the steering wheel.

[0307] In this embodiment, the data acquisition module is connected to the road transport dynamic monitoring and supervision platform to acquire vehicle trajectory time-series data and static information such as "people, vehicles, roads, and rings," and uploads it to the central control module. The trajectory data is stored in the first memory C1, and the static information is stored in the second memory C2. The trajectory reconstruction module retrieves vehicle trajectory data from the first memory C1, processes the data using a reinforcement learning-based trajectory reconstruction model, and obtains high-quality reconstructed trajectory data, which is stored in the third memory C3. The early warning effectiveness prediction module retrieves static information from the second memory C2, calls the reconstructed trajectory data from the third memory C3, uses a prediction model to predict the driver's early warning effectiveness under different early warning types, outputs the corresponding early warning effectiveness level, and stores the prediction results in the fourth memory C4. The central control module retrieves the driver's early warning effectiveness level from the fourth memory C4 and issues early warning strategies based on the different early warning types.

[0308] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the vehicle control method considering the effectiveness of warnings for operating drivers as described above.

[0309] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the vehicle control method described above that takes into account the effectiveness of warnings for operating drivers.

[0310] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0311] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0312] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A vehicle control method that takes into account the alert effectiveness of an operating driver, characterized by, The method includes the following steps: Step S1. Acquire vehicle trajectory data, early warning event data, and static feature data; Step S2. Construct a trajectory reconstruction and repair model, train it based on the speed data in the vehicle trajectory data, and use the trained trajectory reconstruction and repair model to output the reconstructed speed. Then, backtrack and update the position and heading angle to obtain the reconstructed trajectory. The trajectory reconstruction and repair model is an adaptive reinforcement learning trajectory reconstruction model based on physical feasible domain constraints. Step S3. Establish a system for judging the effectiveness of early warnings, and obtain a set of early warning effectiveness characterization results for each early warning event; Step S4. Construct a prediction model for the effectiveness of warnings for operating drivers. Train the model based on the reconstructed kinematic data, static feature data, warning type, and the warning effectiveness representation results determined in step S3 within a fixed short window before the warning is triggered. Then, use the trained prediction model to output the predicted value of the warning effectiveness for each candidate warning type. Step S5. Determine the warning effectiveness level based on the predicted value of the warning effectiveness level. For different warning types, adopt different vehicle control strategies according to the warning effectiveness level. The method for constructing the driver warning effectiveness prediction model is as follows: First, the reconstructed kinematic data is processed based on Transformer to extract temporal motion features; static feature data is divided according to semantic source and embedded and mapped to obtain semantic tokens of the same dimension; then, semantic tokens from different sources are combined into a static token sequence; warning types are embedded and mapped to obtain warning type tokens, and gating vectors are generated from the warning type tokens; the static token sequence is fused with type conditionalization to obtain a type-conditionalized static token sequence; cross-modal attention is used to adaptively retrieve the cross-modal static context vector most relevant to the current motion state from the type-conditionalized static tokens, and a channel-gating vector is introduced to control the injection intensity; finally, the static context is injected into the temporal hidden state in the form of gating residuals to obtain the fused time-by-time representation; then, the temporal-level representation is aggregated into a fixed-length event-level representation, and the event-level representation vector is used to generate the final time-by-time representation. A joint output head is constructed, which simultaneously outputs the probability of reaction occurrence, the intensity of conditional reaction, and its fluctuation scale, and the predicted value of the effectiveness of the early warning is calculated based on this.

2. The vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 1, characterized in that, The vehicle trajectory data includes positioning time, speed, heading angle, and latitude and longitude; the warning event data includes warning trigger time and warning type; the warning types include forward collision warning, speeding warning, and lane departure warning; the static feature data includes driver information, vehicle type, road type, and meteorological environment data.

3. The vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 1, characterized in that, The construction method of the adaptive reinforcement learning trajectory reconstruction model based on physical feasible region constraints is as follows: first construct the state space. and action space ;in, Represents the velocity sequence within the window; This represents the sequence of defect markers within the window; Represents the acceleration feature sequence within the window; action Lagrange interpolation, kinematic low-pass filtering, and robust linear bridging operators are used to output a candidate repair velocity sequence within the output window after the action is executed. Then, based on the acquired candidate repair velocities and physical feasible region constraints, the original observation velocity of the defect points in the center area of ​​each window is updated while keeping the original observation velocity of normal and reliable points unchanged, thus obtaining a global reconstruction velocity sequence. A reward function is designed to calculate the immediate reward after obtaining the reconstruction velocity. Based on the samples in the empirical replay buffer, a dual-depth Q-network is trained to obtain a reinforcement learning policy model for trajectory window defect repair, namely the trajectory reconstruction and repair model.

4. A vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 1, characterized in that, In step S5, for forward collision warning and overspeed warning, when the warning effectiveness level is 2, a voice reminder is given to the driver; When the warning effectiveness level is 1, a voice prompt will remind the driver, along with a visual reminder displayed on the central control screen, and the accelerator pedal opening will be reduced to its original position. ; When the warning effectiveness level is 0, a voice reminder is given to the driver, combined with tactile feedback from seat and steering wheel vibrations, while the accelerator pedal opening is reduced to its original position. And the electronic braking force will be increased to the original level. times; For lane departure warnings, when the warning effectiveness level is 2, a voice reminder will be given to the driver. When the warning effectiveness level is 1, a voice reminder is given to the driver, and a visual reminder is superimposed on the central control display screen, while a tactile reminder is given in combination with steering wheel vibration; When the warning effectiveness level is 0, the system provides a voice reminder to the driver and a tactile reminder in conjunction with the vibration of the seat and steering wheel. It also provides return damping and return torque to guide the driver to make corrections.

5. A vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 3, characterized in that, The candidate repair velocities generated by the action are mapped to reconstructed velocities that satisfy both velocity and acceleration boundaries via a recursive dynamic feasible region projection, expressed as: ; ； ； in, Indicates the first The corresponding trajectory for each warning event is in the index. The reconstruction speed value at that location, Indicates the first The corresponding trajectory for each warning event is in the index. Candidate repair speed at the location These are the lower and upper bounds of velocity, respectively. For the physical limit of acceleration, The time interval between adjacent positioning, Indicates will Cut off to interval Inside, For upper and lower boundaries; Instant returns The expression is: ； ； ； ； ； in, These are the weighting coefficients. This indicates that the reference velocity value was obtained after rigorous physical verification and smoothing. Indicates the number of points in the central area. Index after trajectory reconstruction longitudinal acceleration at that point Reconstructing error terms for the central region; The speed of motion after trajectory reconstruction. For the smoothness term of the central region, The central area of ​​the window The number of times the internal candidate repair speed triggers physical constraint projection truncation.

6. A vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 1, characterized in that, The expression for the cross-modal static context vector is: ； ； ； in, Indicates time Attention weights for the four sets of static tokens This is the cross-modal static context vector obtained through aggregation. This represents the dimension of the hidden state vector; It is a learnable mapping matrix; For a moment The query vector, and These are the key matrix and value matrix obtained by mapping four sets of static tokens, respectively. Represents the transpose symbol. It is a static token sequence. Indicates the first The trajectory short-window sequence corresponding to each early warning event is in the... The hidden state vector obtained by extracting temporal features from each sampling point.

7. A vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 6, characterized in that, The merged time-by-time representation for: ； ； in, For channel gating vectors, This represents vector concatenation. For the mapping matrix, These are bias terms, all of which are learnable parameters.

8. A vehicle control method considering the effectiveness of early warning for commercial drivers according to claim 7, characterized in that, Event-level representation vector The expression is: ； ； ； in, For a moment Event-level attention weights Indicates time Score the importance of event-level prediction; Mapping matrix For bias vector The projection vector; Predicted value of early warning effectiveness ;in, The probability of the reaction occurring. The strength of the conditional reaction.

9. A vehicle control system for implementing the vehicle control method considering the effectiveness of warnings for commercial drivers as described in any one of claims 1 to 8, characterized in that, The system includes a data acquisition module, a trajectory reconstruction module, a warning effectiveness prediction module, a central control module, a vehicle communication module, a throttle control module, a braking control module, a voice control module, a central control screen control module, a vibration control module, and a steering control module. The data acquisition module acquires driver, vehicle, road, and environmental data from a road transport dynamic monitoring and supervision platform. The trajectory reconstruction module processes vehicle trajectory data and provides the reconstructed trajectory data to the warning effectiveness prediction module. The warning effectiveness prediction module integrates vehicle trajectory time-series data with static information on "people, vehicles, roads, and the environment" to assess the warning effectiveness of drivers under different warning types. The system performs predictions and outputs corresponding warning effectiveness levels. The central control module receives data uploaded by the data acquisition module and issues control strategies to the vehicle based on the driver's warning effectiveness level. The vehicle communication module receives the control strategies issued by the central control module. The throttle control module controls the throttle opening. The brake control module controls the electronic power assist provided by the electro-hydraulic braking system. The voice control module controls the vehicle's voice system. The central control screen control module controls the display content of the vehicle's central control screen. The vibration control module controls the vibration of the steering wheel and seat. The steering control module controls the return damping and torque of the steering wheel.

Images