Active learning for gear defect detection in gearboxes

Abstract

Condition monitoring of gears in gearboxes is crucial to ensure performance and minimizing downtime in many industrial applications including wind turbines and automotive. Monitoring techniques using indirect measurements (i.e. accelerometers, microphones, acoustic emission sensors and encoders, etc.) face challenges, including the defect interpretation and characterization. Vision-based gear condition monitoring, as a direct method to observe gear defects, has the capability to give a precise indication of the starting point of a potential surface failure, but suffers from the image annotations (to train a reliable vision model for automatic defect detection of gears). In this paper, we propose an active learning framework for vision-based condition monitoring, to reduce the human annotation effort by only labelling the most informative examples. In particular, we first train a deep learning model on limited training dataset (annotated randomly) to detect pitting defects. To select which samples have the highest priority to be annotated, we compute the model's uncertainty on all remaining unlabeled examples. Bayesian active learning by disagreement is exploited to estimate the uncertainty of the unlabeled samples. We select the samples with the highest values of uncertainty to be annotated first. Experimental results from defect detection of gears in gearboxes show that with less than 6 times image annotations, we can achieve similar performances.