Fault Detection in Gauge-Sensorized Strain Wave Gears

Strain wave gears (SWG) are employed in most robot joints, and hence monitoring their condition gets more important in robotic applications. The condition of SWGs can, e.g., be observed by sensor signals of strain gauges that are mounted on the flex spline, the deformable part of the gear, for torque prediction purposes. In this paper, the feasibility of utilizing these sensor signals for fault detection in SWGs is shown and meaningful features tailored to a specific sensor setup are proposed. As a first important step towards fault detection in a real world application, synthetically generated sensor signals are considered that are derived from a simulation chain allowing the injection of different faults. In total, five distinct and practically relevant faults are considered and different algorithms are applied to classify them. In addition, robustness regarding disturbed synthetic data is investigated and the classifiers' potential for out-of-distribution prediction is evaluated.