Robot Fault Detection Using Digital Twin and Deep Learning

Reliability and lifetime of robots is critical for modern manufacturing systems, as robots have been widely used in manufacturing, and their failure could lead to substantial financial losses. Predictive Maintenance (PdM) has emerged as an effective strategy, utilizing historical data and prognostic models to anticipate maintenance needs. Digital twins simulate the behavior of real systems and connect to the real system using sensors in real-time, and have shown potentials to booster the performance of predictive maintenance algorithms. In this paper, we present an open-source demonstration platform of using digital twin for robot PdM. A data-driven digital twin is developed first for a robot to predict its temperature during operation. By leveraging time-series data collected in real-time, including motor temperature, voltage, and position, a data-driven algorithm is developed to detect abnormality in motor temperature response. An innovative Recursive Prediction Update (RPU) technique is proposed, which replaces fault-contaminated data with predicted values in real-time, significantly enhancing accuracy of abnormility detection. Results show that the integration of Digital Twins and RPU improves fault detection performance, offering valuable insights for predictive maintenance under limited failure data conditions.

Keywords: Digital twin, Predictive maintenance, Robotic arm, Deep learning, Fault diagnosis.