Deep reinforcement learning based active safety control for distributed drive electric vehicles

Abstract

Distributed drive electric vehicles are regarded as the promising transportation due to the advanced power flow architecture. Optimizing the yaw motion to enhance vehicle safety is a challenging job. Besides, the nonlinear features in vehicles affect the control accuracy of the yaw motion controllers. To this end, a deep reinforcement learning (DRL) based direct yaw moment control (DYC) strategy is put forward here. Vehicle dynamics can be approximated with the DRL algorithm, which reduces the complex nonlinear solving process. Concretely, the DYC problem is formulated as Markov Decision Process in which the observed signals and external yaw moment are incorporated as the state and action sets. Thereupon, actor-critic network is exhibited to approximate action-value function and policy function for better control performance. Furthermore, to guarantee the continuous solution of external yaw moment, the deep deterministic policy gradient algorithm is employed, in which target and online network parameters are simultaneously trained to maintain learning process stability. The proposed DRL based DYC strategy is verified using the Carsim/Simulink platform under the typical lane change manoeuvres. Numerical test results demonstrate that the proposed DYC strategy outperforms the linear approaches on taking full advantage of understeer features and enhancing the lateral stability, especially under the critical steering manoeuvres.

Optimal energy management strategy for dual-power coupling tractor based on the adaptive control technology

“…The proposed energy management approach can adaptively allocate output power between the engine and motor using the integrated strategy by reading up table written to VCU. Furthermore, the computer can store the state variables and parameters [33]. To demonstrate the effectiveness of the proposed technique, the typical rule-based optimization algorithm is employed for comparison.…”

Torque coordinated control of the through-the-road (TTR) 4-wheel-drive (4WD) hybrid vehicle under extreme road conditions

“…Similarly, in 26 , the RL scheme was used to tune PID controller parameters. Wei et al 27 , 28 combined with vehicle safety and energy utilization efficiency, a torque coordination control strategy based deep RL was proposed, subsequently, the effectiveness of this strategy was proved by the simulation. However, the RL approach is not suitable for the online operation due to its increased computational overhead.…”

Active Steering Controller for Driven Independently Rotating Wheelset Vehicles Based on Deep Reinforcement Learning

“…In the current research, DRL has been successfully applied to the control of different nonlinear systems, such as the intelligent driving of automobiles [21][22][23][24], active control of railway vehicles [25][26][27][28], and robotic control [29,30]. In automotive fault-tolerant control, a double Q-Learning algorithm has been proposed for the online determination of optimization weight factors by integrating DRL into a fault-tolerant coordinated controller, which ensures that vehicles can achieve optimal control strategies across various operating conditions [21].…”

“…In automotive fault-tolerant control, a double Q-Learning algorithm has been proposed for the online determination of optimization weight factors by integrating DRL into a fault-tolerant coordinated controller, which ensures that vehicles can achieve optimal control strategies across various operating conditions [21]. In active safety control, DRL has been utilized to enhance the yaw motion stability of distributed drive electric vehicles using the Deep Deterministic Policy Gradients (DDPGs) to learn and control the vehicle's nonlinear dynamics [22]. A controller that combines DRL with Nonlinear Model Predictive Control (NMPC) has been proposed in [23] to achieve safe highway autonomous driving.…”

“…For the actor network, the IS weight of each sample is employed to weight the policy gradient, yielding a weighted policy gradient. This weighted policy gradient is then used to update the parameters of the actor network, as illustrated in Equation (22).…”