Control of digital electrohydraulic systems driven by poppet-type hydraulic valves using reinforcement learning

In this thesis, a control method is proposed for the position control of digital electrohydraulic systems (D-EHS) that could replace conventional electrohydraulic systems (C-EHS) in terms of energy efficiency and cost. This offers an alternative approach to the complex controller design process frequently encountered in the control of the D-EHS. The double deep Q-network (DDQN) reinforcement learning approach, known for its effectiveness in controlling large and discrete action spaces, is employed for the position control of the D-EHS. The mathematical model of the D-EHS is obtained and employed for training the DDQN controller. Various training approaches were explored, and the best result was achieved by retraining the multiple DDQN controllers, which had been pretrained in a simulation environment, using an experimental setup. Additionally, the D-EHS's performance is evaluated against the C-EHS, showing comparable results in both simulation and experiments. A deep deterministic policy gradient based controller is employed for the position control of the C-EHS. The behavior of the proposed DDQN controller and the D-EHS under non-training conditions is also examined through experimental studies. The results show that, the D-EHS offers comparable performance to the C-EHS when used with the proposed controller, and the position control performance is maintained even under non-training conditions.