Image-Based Damage Quantification and Vibration-Based Estimation of Gear Tooth Degradation

Abstract

Estimating gear damage remains a complex task due to significant variability in gearbox lifespans. The ability to accurately predict gear damage is crucial in various industries such as aerospace and automotive, where gearbox failure can have catastrophic consequences. While conventional data acquisition methods rely on accelerometer data to monitor gearbox health, this data is indirect and can be affected by noise and lead to inaccurate gear damage predictions. Optical images of the tooth surface provide a reference for the gear spall damage but image acquisition is expensive as it disrupts the operation.
The 2026 PHM Data Challenge aims to develop models that can predict gear damage using accelerometer data and only use imaging data during training.
A high-level detail of the methodology is depicted in Figure 1. In this work, an image-based damage indicator (DI) is constructed to quantify gear degradation where candidate image features are evaluated against the Coble metrics [1]. Subsequently, the DIs are used as a reference for developing partial least squares (PLS) regression model using vibration-based condition indicators for gear degradation estimation [2, 3]. The method is assessed using leave-one-out cross validation (LOOCV) and applied to unseen experiments to assess the potential.