Deep reinforcement learning–based digital twin system of a multi-cylinder hydraulic press with adaptive synchronous control

Abstract

It is well known that due to the complex modeling process, including the over-actuated and eccentric load characteristics of multi-cylinder hydraulic press systems, designing their control systems is very challenging. In reality, engineers must undertake considerable design and debugging tasks. As a result, we recommend using deep reinforcement learning to address these difficult modeling problems. The high failure rate and risk, on the contrary, have hindered the industrial application process. We present a digital twin solution for multi-cylinder hydraulic presses in this paper. The deep reinforcement learning agent is trained in the digital twin environment to dynamically adjust the synchronous controller parameters to fulfill various forging tasks. For multi-cylinder hydraulic presses, the digital twin system performs production operations such as virtual testing, fault analysis, optimization design, intelligent decision-making, and online operation and maintenance. To evaluate the digital twin system, we performed quick deployment in a real-world setting and created a partial load forging experiment for the multi-cylinder hydraulic press. Throughout the experiment, there was no loss of control or collision, and the forging process obtained good results.