Online damage severity level classification in gears under natural damage progression

Effective diagnosis of the gear damage stages is critical for the industries to reduce unexpected failures and maximise life utilisation. In geared systems, pitting is one of the most common failure modes observed, which originates from the surface/subsurface cracks. The gear damage levels were classified using seeded defect data instead of naturally progressed in the reported works. It is difficult to simulate a natural pitting failure on the gear tooth using an artificial process. As implemented in prior experimental studies involving seeded defects, a sudden change in the gear pitting area may not occur in practice. This study presents an ensemble decision tree-based random forest (RF) classifier methodology for the online classification of gear damage stages under natural pitting progression. A health indicator (HI) termed CCR (i.e. correlation coefficient of residual vibration signal) is extracted using a raw vibration signal to represent the pitting progression in spur gears. The exact relationship between the HI and gear tooth degradation stages is crucial during the implementation of the classifier model. Hence, a binary segmentation (BS) methodology identifies the relationship between HI and gear health stages (i.e. healthy, initial pitting, medium pitting and severe pitting). The output of BS methodology is used for classifier model training, and later based on the trained model, gear pitting severity levels were estimated for a newly installed gear. The performance of the proposed framework (i.e. combining BS and RF methodology) is validated through six accelerated runs to failure gear pitting experiments.