Efficient Deep Q-Learning for Industrial Equipment Calibration in Elevator Manufacturing

Nikolaos Dimitriou, Paschalis Charalampous, Theofilos D. Mastos, Dimitris Tzovaras

Keywords: Automatic calibration, deep Q-learning, parameter estimation, reinforcement learning

Abstract: Industrial equipment calibration is an essential element for the proper functioning of any production plant. Without frequent and proper calibration, the quality and efficiency of the overall production process are threatened. Despite its significance, it is often performed manually based on qualitative measures that may lead to suboptimal behavior. In this article, we propose a deep reinforcement learning (RL)-based methodology to automate the calibration process and evaluate it in an elevator control use case. Moreover, to overcome data scarcity we develop a simulation environment based on generative modeling that creates synthetic RL episodes. The proposed methodology relies on a minimal set of sensors and actuators, i.e., a webcam, an Arduino board, three stepper motors, and an edge computational unit. Nevertheless, experimental evaluations indicate that it can perform in real-time applications achieving accurate calibration results.

Submission Number: 70