Machine Learning for Detecting and Locating Damage in a Rotating Gear

This paper describes a multi-disciplinary damage detection methodology that can aid in detecting and diagnosing a damage in a given structural system, not limited to the example of a rotating gear presented here. Damage detection is performed on the gear stress data corresponding to the steady state conditions. The normal and damage data are generated by a finitedifference solution of elastodynamic equations of velocity and stress in generalized coordinates1. The elastodynamic solution provides a knowledge of the stress distribution over the gear such as locations of stress extrema, which in turn can lead to an optimal placement of appropriate sensors over the gear to detect a potential damage. The damage detection is performed by a multi-function optimization that incorporates Tikhonov kernel regularization reinforced by an added Laplacian regularization term as used in semi-supervised machine learning. Damage is mimicked by reducing the rigidity of one of the gear teeth. Damage detection models are trained on a subset of the normal data and are then tested on the damage solution. The precision with which the damaged tooth and the extent of the damage are identified is very encouraging. The present methodology promises to lead to a significant damage detection, diagnosis and prognosis technology for structural health monitoring.