Clutch Engagement Control Of Starting Process For AT With Reinforcement Learning

Automatic transmission can guarantee the stability of the starting process relaying on the torque converter. But before that, it needs to complete the clutch engagement, which will likely cause impact or cost too much time if the clutch can’t be engaged rapidly and smoothly.To meet the control quality requirements, a common approach is to modify the control parameters with the bench calibration and the vehicle calibration. On one hand, it takes lots of human and material resources. On the other hand, even if the calibration work makes the clutch engagement control to reach a good quality, the quality will gradually decline along the use of the vehicle. So in order to improve the reproducibility and consistency of the control quality, the control method of clutch engagement process of AT with reinforcement learning was studied.The basic control strategy is that using an open-loop control in filling phase and using a closed-loop control in inertia phase with turbine rotation speed as the feedback information. Filling time and the initial duty value was picked as the learning parameters, and then a turbine speed curve-tracked agent was set up. The whole learning system insists of two parts: the simulation training part and the parameter modifying part. First learning a policy on the simulation model and then implementing the obtained policy to modify the control parameter of clutch control. The learning system was tested on the HIL(hardware in the loop) bench.The experimental results showed that within a few clutch engagement processes the parameter modifying part can make it reach an ideal turbine rotation speed curve, which means the impact was efficiently reduced and the adaptive ability of the clutch engagement control was improved.