On the relevance of using temperature for mechanical fault detection in industrial robots based on dynamic modeling

This paper addresses a key blind spot in industrial robot fault detection and dynamic modeling: the modeling of temperature effects. Although often neglected, ambient temperature and joint temperature significantly influence robot dynamics and vary widely in practice. To address this challenge, a one-class anomaly detection approach is proposed based on an adaptation of a previously introduced hybrid inverse dynamic model combining physics-based and data-driven components, which constructs tight envelopes of normality . The resulting anomaly detector is capable of detecting very slight torque deviations in a real TX2-90 Stäubli 6-axis robot, while remaining robust across a wide trajectory space and continuous joint temperature ranges. The behavior of the envelopes is further analyzed, their effectiveness for anomaly detection on the considered robot is illustrated through preliminary experiments, and the importance of accounting for temperature for mechanical fault detection and localization is highlighted.