Learning Methods for the Feedforward Control of a Hydraulic Clutch Actuation Path

Hydraulic clutch actuation paths in heavy duty transmissions show a lot of variability due to series production variance, e.g. different friction. Additionally, there are external factors, such as the operators choice of hydraulic fluid or the compliance with specified service intervals. A further factor is the amount of entrained air that can vary from clutch to clutch in a single transmission. To address these effects, the authors recently proposed feedforward control concepts for the hydraulic clutch actuation path. While the feedforward controllers were shown to significantly improve the control quality, a manual parameter adaptation is necessary for optimal performance. Therefore, this paper presents and evaluates different learning algorithms. One algorithm based on rules and one based on multi-armed bandits that are suitable for a model-based control scheme. The third, a reinforcement learning algorithm based on policy gradients is suitable for a data-based feedforward control concept. In simulations with a high fidelity model, the improvement achievable in comparison with the state-of-the-art approach with all three algorithms is shown. In addition, the strengths and weaknesses of each algorithm are discussed.