Decreasing ramp-up durations of ultraprecision machine tools using reinforcement learning

Abstract

Ramp-up phases in machining processes are often required to enable the desired workpiece quality. Within conventional CNC processes these phases mainly depend on the corresponding CNC program. But, especially in the field of ultra-precision manufacturing these ramp-up phases require adaptions deep within the control system of the machine tools. The control parameters of the axis servo-drives need to be optimized regarding the corresponding manufacturing process. This process inhibits an iterative nature, which causes multiple rejects because of not sufficient workpiece quality. Naturally this optimization process adds up to the overall manufacturing time. One of the main factors to reduce this time span is the machine operator's experience. But within the currently expanding industry of high-tech products, which require functional workpiece surfaces inhibiting complex geometries, requirements in workpiece tolerances are constantly increasing, resulting in individual and mostly unknown axis control settings for most newly developed workpieces. In a first approach of automating and accelerating the ramp-up phase, an artificial intelligence solution, based on reinforcement learning techniques is introduced. One of the main advantages of using reinforcement learning (RL) based models for this problem is their capability to adapt to feedback from their environment. The installed machine drives can serve as learning environments, but this approach results in extraordinary high training durations. The behavior of the machine axis can be efficiently simulated by applicating techniques from the field of control theory, which results in a drastic reduction of training times while the behavior of the real axis can still be emulated.