A data-model backend fusion method for automatic driving decision control function

Abstract

A kind of data-model backend fusion method for automatic driving decision control function, comprising the following steps: step one, driving data acquisition and data set construction;Step two, vehicle dynamics theory model establishment and verification;Step three, the automatic driving decision control problem is constructed as integrated dynamic optimization problem, and the decision control problem is solved at each time;Step four, the dynamic optimization problem is solved by model predictive rolling horizon online optimization;Step five, the dynamic optimization problem is solved by data-driven reinforcement learning offline optimization;Step six, the confidence of pure model-driven solution is calculated;Step seven, the confidence of pure data-driven solution is calculated, step eight, the backend fusion factor of model data information and fusion decision control quantity are calculated.The application improves the safety and decision control optimality in unusual dangerous conditions while ensuring the decision control performance in common conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicles and relates to an autonomous driving decision control method. In order to improve the safety and optimality of decision control of autonomous vehicles under uncommon dangerous conditions, a new decision control method integrating vehicle dynamics model and driving data backend is proposed. Background Technology

[0002] Most existing autonomous driving decision-making and control technologies are based on human driving experience rules or vehicle dynamics theoretical models for algorithm design. They utilize heuristic rules or vehicle motion characteristic information to solve for decision and control outputs, and the developed decision-making and control algorithms have a certain applicability to common autonomous driving conditions. Currently commonly used decision-making and control methods include expert decision systems, PID control, fuzzy rule control, and model predictive control. However, in actual autonomous driving processes, existing decision-making and control algorithms often cannot adapt to the uncertainties of road driving conditions and traffic driving environments, leading to weakened decision-making and control performance of autonomous vehicles and even posing a serious threat to road traffic safety.

[0003] Existing autonomous driving decision-making and control technologies have the following main drawbacks: First, they are designed based on human driving experience rules, resulting in complex decision-making and control layers and high costs for algorithm maintenance and updates; Second, they are designed based on vehicle dynamics theoretical models, making the performance of decision-making and control algorithms highly dependent on the accuracy of the models, and they are usually only applicable to linear state regions; Third, decision-making and control algorithms that simply use driving experience or theoretical model information do not make full use of information and lack design for uncertainties in road driving conditions and traffic driving environment, resulting in poor autonomous driving performance in less frequent scenarios. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a data-model backend fusion method for autonomous driving decision-making and control functions. This method combines the advantages of existing technologies and is convenient for practical engineering applications. It aims to ensure decision-making and control performance under common operating conditions while improving safety and optimal decision-making and control under uncommon and hazardous conditions.

[0005] The technical solution adopted by this invention to solve its technical problem is: A data-model backend fusion method for autonomous driving decision-making and control functions includes the following steps: Step 1: Driving Data Collection and Dataset Construction Using the onboard sensor system of autonomous vehicles, vehicle dynamics data and driving scenario data are collected under typical driving conditions, including data at various times ( t , t+1 , t+2 Autonomous driving scenario states (…) sInput and autonomous vehicle decision control quantity a Output; Autonomous driving scenario status s Including vehicle dynamics x The surrounding traffic environment and road information influence the vehicle's decision-making and control decisions. a Including decision-making behavior and chassis motion control. u Using the Markov decision process definition method, data from adjacent time points are combined to construct a driving database. s t , a t , s t+1 () s t+1 , a t+1 , s t+2 ...; Step 2: Establishment and Verification of Vehicle Dynamics Theoretical Model Since the decision-making and control of autonomous vehicles mainly focus on motion within the road plane, based on the knowledge of vehicle dynamics theory, a discrete analytical model of vehicle lateral and longitudinal dynamics is established, including differential dynamic models in the longitudinal, lateral, and yaw directions. x t+1 =Ax t +Bu t +Gξ t ,in x This is the vehicle dynamics state vector. u This is the motion control vector for the vehicle chassis. ξ To compensate for the linearization error of the dynamic model, A , B , G The vehicle parameter matrix is ​​then used; subsequently, the vehicle dynamics data stored in the driving database described in step one and the least squares algorithm are used to identify the parameters and verify the model accuracy of the vehicle dynamics model. Step 3: Construct the autonomous driving decision-making and control problem as an integrated dynamic optimization problem, and solve the decision-making and control problem at each time step. The input to the decision control problem is the autonomous driving scenario state described in step one. s The output is the decision control quantity for autonomous vehicles. a The optimization objective is a multi-performance weighted index of autonomous vehicles in the current driving scenario, including safety, comfort, smoothness, and energy efficiency. For this dynamic optimization problem, different solution methods will correspond to different problem construction forms. Step 4: Solve the above dynamic optimization problem by performing model prediction and rolling time-domain online optimization. At the current moment, the decision control optimization objective is the sum of performance-weighted indices over the finite future prediction time domain. J Using the aforementioned discrete vehicle dynamics model and existing optimization problem solvers, the minimum... J To optimize the objective, solve for... N The optimal control sequence in the finite prediction time domain ( u t , u t+1 , u t+2 , …, u t+N The vehicle behavior decision corresponding to this control sequence is related to the first control variable. u t The combination of these is the optimal decision control quantity for pure model-driven solution. u MB ; Step 5: Solve the above dynamic optimization problem offline using data-driven reinforcement learning. By employing an offline-optimized, online application approach, the Actor-Critic reinforcement learning method is first used to sample and iteratively solve Markov decision data from the aforementioned driving database. This yields the optimal performance index evaluation function within the infinite future prediction time domain. Q The optimal action strategy under the current driving scenario is the output of the optimal action strategy in the online application process, which is the corresponding optimal decision control quantity obtained through pure data-driven solution. a DB ; Step 6: Calculate the confidence score of the pure model-driven solution. Because the decision control quantity based on the model solution in step four u MB The performance generally depends on the objective function after online optimization. J Size, J A smaller value indicates better optimality of the solution; therefore, the value calculated online at the current time can be utilized. J t Value and nominal value` J The optimality and confidence of the solution are determined by the gap judgment, and the confidence of the solution is calculated by the calculation model. c x , c x The value range is [0, 1], and it is represented by a quadratic function. c x With |` J - J t | Relationship; Step 7: Calculate the confidence score for pure data-driven solutions. Because the decision control quantity is obtained based on data in step five. a MB The performance generally depends on the evaluation function after offline optimization. Q Size, Q A larger value indicates better optimality of the solution, therefore, the online application computation at the current time can be utilized. Q t Value and nominal value` Q The optimality and confidence of the solution are determined by the difference in the data, and the confidence level is calculated. c s , c s The range of values ​​is [0,1], and it is represented by a quadratic function. c s With |` Q - Q t | Relationship; Step 8: Calculate the back-end fusion factor and fusion decision control quantity of the model data information. To fully leverage the advantages of both model and data information, the optimal decision control quantities of the two are complemented; the model-data fusion factor is calculated. K F = c s / (c s + c x ) Its value range is also [0, 1]. K F The backend integrates data and model information to calculate the autonomous driving decision-making and control quantities based on the fused model and data. u F = u MB + K F * (a DB - u MB ) .

[0006] Furthermore, in step one, the vehicle's dynamic state x The data includes position, velocity, and acceleration in each direction of motion; the surrounding traffic environment includes the position and velocity of surrounding traffic participants within the road plane; the road information includes road structure, road gradient, and traffic signals; the decision-making behavior includes lane selection, acceleration / deceleration, and steering; and the chassis motion control parameters. u This includes throttle opening, braking intensity, and steering wheel angle.

[0007] The beneficial effects of this invention are mainly reflected in: 1. In typical autonomous driving scenarios, by fusing driving data information and dynamic model information, the model information is used to ensure the basic performance and robustness of the decision control function, and the data information is combined to improve the optimality of the decision control function. 2. By leveraging the diversity of collected driving data, the scenario coverage of autonomous driving decision-making and control functions can be increased, and by utilizing the authenticity of the data, safety in actual application can be improved. 3. By leveraging the continuous updating capability of driving data and the iterative learning capability of the reinforcement learning method, the learning and evolution of autonomous driving decision-making and control capabilities can be achieved by combining vehicle-cloud collaborative technology. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the main calculation process of the method of the present invention; Figure 2 This is a schematic diagram of the back-end fusion factor calculation in the method of this invention; Figure 3 This is a comparison chart of the implementation effects of the method of this invention and the classical method. Detailed Implementation

[0009] The present invention will now be further described with reference to the accompanying drawings.

[0010] Reference Figures 1-3 A data-model backend fusion method for autonomous driving decision-making and control functions includes the following steps: Step 1: Driving Data Collection and Dataset Construction Using the onboard sensor system of autonomous vehicles, vehicle dynamics data and driving scenario data are collected under typical driving conditions, including data at various times ( t , t+1 , t+2 Autonomous driving scenario states (…) s Input and autonomous vehicle decision control quantity a Output; Autonomous driving scenario status s Including vehicle dynamics x (e.g., position, speed, acceleration, etc. in each direction of motion), surrounding traffic environment conditions (e.g., position and speed of surrounding vehicles, pedestrians, and other traffic participants on the road plane), and road information (e.g., road structure, road gradient, traffic signals, etc.), the vehicle's decision-making and control quantities. a This includes decision-making behaviors (such as lane selection, acceleration / deceleration, and steering) and chassis motion control parameters. u (e.g., throttle opening, braking intensity, steering wheel angle, etc.); using the Markov decision process definition method, data from adjacent time points are combined to construct a driving database. st , a t , s t+1 () s t+1 , a t+1 , s t+2 ...; Step 2: Establishment and Verification of Vehicle Dynamics Theoretical Model Since the decision-making and control of autonomous vehicles primarily focus on motion within the road plane, discrete analytical models of vehicle dynamics in both the longitudinal and lateral directions are established based on vehicle dynamics theory. These models include differential dynamic models in the longitudinal, lateral, and yaw directions. x t+1 =Ax t +Bu t +Gξ t ,in x This is the vehicle dynamics state vector. u This is the motion control vector for the vehicle chassis. ξ To compensate for the linearization error of the dynamic model, A , B , G The vehicle parameter matrix is ​​then used; subsequently, the vehicle dynamics data stored in the driving database described in step one and the least squares algorithm are used to identify the parameters and verify the model accuracy of the vehicle dynamics model. Step 3: Construct the autonomous driving decision-making and control problem as an integrated dynamic optimization problem, and solve the decision-making and control problem at each time step. The input to the decision control problem is the autonomous driving scenario state described in step one. s The output is the decision control quantity for autonomous vehicles. a The optimization objective is a multi-performance weighted index of autonomous vehicles in the current driving scenario, including safety, comfort, smoothness, and energy efficiency. For this dynamic optimization problem, different solution methods will correspond to different problem construction forms. Step 4: Solve the above dynamic optimization problem by performing model prediction and rolling time-domain online optimization. At the current moment, the decision control optimization objective is the sum of performance-weighted indices over the finite future prediction time domain. J Using the aforementioned discrete vehicle dynamics model and existing optimization problem solvers (such as Adam, OSQP, and IPOPT), the minimum... J To optimize the objective, solve for... N The optimal control sequence in the finite prediction time domain ( u t ,u t+1 , u t+2 , …, u t+N The vehicle behavior decision corresponding to this control sequence is related to the first control variable. u t The combination of these is the optimal decision control quantity for pure model-driven solution. u MB ; Step 5: Solve the above dynamic optimization problem offline using data-driven reinforcement learning. By employing an offline-optimized, online application approach, we first utilize Actor-Critic reinforcement learning methods (such as DDPG, TD3, SAC, etc., which involve iteratively solving the evaluation function Critic and the action function Actor) to sample and iteratively solve Markov decision data from the aforementioned driving database. This allows us to obtain the optimal performance index evaluation function in the infinite future prediction time domain. Q The optimal action strategy under the current driving scenario is the output of the optimal action strategy in the online application process, which is the corresponding optimal decision control quantity obtained through pure data-driven solution. a DB ; Step 6: Calculate the confidence score of the pure model-driven solution. Because the decision control quantity based on the model solution in step four u MB The performance generally depends on the objective function after online optimization. J Size, J A smaller value indicates better optimality of the solution; therefore, the value calculated online at the current time can be utilized. J t Value and nominal value` J The optimality and confidence of the solution are determined by the gap judgment, and the confidence of the solution is calculated by the calculation model. c x like Figure 2 As shown, c x The range of values ​​is [0, 1], and it can be represented by a quadratic function. c x With |` J - J t | Relationship; Step 7: Calculate the confidence score for pure data-driven solutions. Because the decision control quantity is obtained based on data in step five. a MB The performance generally depends on the evaluation function after offline optimization. Q Size, QA larger value indicates better optimality of the solution, therefore, the online application computation at the current time can be utilized. Q t Value and nominal value` Q The optimality and confidence of the solution are determined by the difference in the data, and the confidence level is calculated. c s like Figure 2 As shown, c s The range of values ​​is [0, 1], and it can be represented by a quadratic function. c s With |` Q - Q t | Relationship; Step 8: Calculate the back-end fusion factor and fusion decision control quantity of model data information. To fully leverage the advantages of model information and data information, the optimal decision control quantities of the two are complemented, and the model data fusion factor is calculated. K F = c s / (c s + c x ) ,like Figure 1 and 2 As shown, its value range is also [0, 1]. K F The backend integrates data and model information to calculate the autonomous driving decision-making and control quantities based on the fused model and data. u F = u MB + K F * (a DB - u MB ) .

[0011] Example: Control scheme for autonomous vehicles driving on partially slippery roads.

[0012] After a period of use, asphalt or cement pavements on urban roads often suffer from partial damage, potholes, and debris accumulation, making certain areas slippery during rainy or snowy weather. The tire adhesion coefficient and grip are significantly reduced in these slippery areas, and the location of these slippery areas is often random. Therefore, controlling autonomous vehicles on slippery surfaces is more challenging, increasing the risk of skidding, rollovers, and lateral rotation—unstable and unsafe conditions. Current methods have not yet adequately addressed safe driving control in such situations.

[0013] The data-model backend fusion method proposed in this invention for autonomous driving decision-making and control functions can effectively utilize vehicle dynamics information and driving data information simultaneously. It is an optional solution method for this type of challenging autonomous driving control task and will have better decision-making and control performance.

[0014] The autonomous driving scenario considered in this embodiment for curved driving on partially slippery road surfaces is as follows: A B-class sedan with front-wheel steering, rear-wheel drive, and independent suspension needs to have an autonomous driving algorithm designed to control its steering system and drive safely on a road with undulations, curves, and slippery areas.

[0015] According to the implementation steps of the present invention, the implementation effects are as follows: Figure 3 As shown, the effects of pure model-driven decision control, pure data-driven decision control, and model-data fusion decision control in this invention are compared. It can be seen that in the curved driving conditions on partially slippery road surfaces in this embodiment, the vehicle path tracking effect of the method of this invention is better than the other two methods, and it avoids yaw rate errors caused by inaccuracies in the vehicle dynamics model under slippery road conditions.

[0016] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A data-model backend fusion method for autonomous driving decision-making and control functions, characterized in that, The method includes the following steps: Step 1: Driving Data Collection and Dataset Construction Using onboard sensor systems in autonomous vehicles, vehicle dynamics data and driving scenario data are collected under typical driving conditions, including data at various times ( t , t+1 , t+2 Autonomous driving scenario states (…) s Input and autonomous vehicle decision control quantity a Output; Autonomous driving scenario status s Including vehicle dynamics x The surrounding traffic environment and road information influence the vehicle's decision-making and control decisions. a Including decision-making behavior and chassis motion control. u Using the Markov decision process definition method, data from adjacent time points are combined to construct a driving database. s t , a t , s t+1 () s t+1 , a t+1 , s t+2 ...; Step 2: Establishment and Verification of Vehicle Dynamics Theoretical Model Since the decision-making and control of autonomous vehicles mainly focus on motion within the road plane, based on the knowledge of vehicle dynamics theory, a discrete analytical model of vehicle lateral and longitudinal dynamics is established, including differential dynamic models in the longitudinal, lateral, and yaw directions. x t+1 = Ax t +Bu t +Gξ t ,in x This is the vehicle dynamics state vector. u This is the vehicle chassis motion control vector. ξ To compensate for the linearization error of the dynamic model, A , B , G The vehicle parameter matrix is ​​then used; subsequently, the vehicle dynamics data stored in the driving database described in step one and the least squares algorithm are used to identify the parameters and verify the model accuracy of the vehicle dynamics model. Step 3: Construct the autonomous driving decision-making and control problem as an integrated dynamic optimization problem, and solve the decision-making and control problem at each time step. The input to the decision control problem is the autonomous driving scenario state described in step one. s The output is the decision control quantity for autonomous vehicles. a The optimization objective is a multi-performance weighted index of safety, comfort, smoothness and energy efficiency of autonomous vehicles in the current driving scenario. For this dynamic optimization problem, different solution methods will correspond to different problem construction forms. Step 4: Solve the above dynamic optimization problem by performing model prediction and rolling time-domain online optimization. At the current moment, the decision control optimization objective is the sum of performance-weighted indices over the finite future prediction time domain. J Using the aforementioned discrete analytical model of vehicle lateral and longitudinal dynamics, and employing an optimization problem solver, to minimize J To optimize the objective, solve for... N The optimal control sequence in the finite prediction time domain ( u t , u t+1 , u t+2 , …, u t+N The vehicle behavior decision corresponding to this control sequence is related to the first control variable. u t The combination of these is the optimal decision control quantity for pure model-driven solution. u MB ; Step 5: Solve the above dynamic optimization problem offline using data-driven reinforcement learning. By employing an offline-optimized, online application approach, the Actor-Critic reinforcement learning method is first used to sample and iteratively solve Markov decision data from the aforementioned driving database to obtain the optimal performance index evaluation function in the future infinite prediction time domain. Q The optimal action strategy under the current driving scenario is the output of the optimal action strategy in the online application process, which is the corresponding optimal decision control quantity obtained through pure data-driven solution. a DB ; Step 6: Calculate the confidence score of the pure model-driven solution. Because the decision control quantity based on the model solution in step four u MB The performance depends on the objective function after online optimization. J Size, J A smaller value indicates better optimality of the solution; therefore, the value calculated online at the current time should be used. J t Value and nominal value The optimality and confidence of the gap judgment are determined, and the confidence of the calculation model is calculated. c x , c x The value range is [0, 1], and it is represented by a quadratic function. c x and Relationship; Step 7: Calculate the confidence score for pure data-driven solutions. Because the decision control quantity is obtained based on data in step five. a MB The performance depends on the evaluation function after offline optimization. Q Size, Q A larger value indicates better optimality of the solution; therefore, the value calculated using the online application at the current time is optimal. Q t Value and nominal value The optimality and confidence of the solution are determined by the difference in the data, and the confidence level is calculated. c s , c s The value range is [0, 1], and it is represented by a quadratic function. c s and Relationship; Step 8: Calculate the back-end fusion factor and fusion decision control quantity of the model data information. To fully leverage the advantages of both model and data information, the optimal decision control quantities of the two are complemented; the model-data fusion factor is calculated. K F = c s / (c s + c x ) Its value range is also [0, 1]. K F The backend integrates data and model information to calculate the autonomous driving decision-making and control quantities based on the fused model and data. .

2. The data-model backend fusion method for autonomous driving decision-making and control functions as described in claim 1, characterized in that, In step one, the vehicle dynamics state x The data includes position, velocity, and acceleration in each direction of motion; the surrounding traffic environment includes the position and velocity of surrounding traffic participants within the road plane; the road information includes road structure, road gradient, and traffic signals; the decision-making behavior includes lane selection, acceleration / deceleration, and steering; and the chassis motion control parameters. u This includes throttle opening, braking intensity, and steering wheel angle.

Images