Generalized ML Approach for Fault Diagnosis of Rotating Machinery using Domain-based Features

The recent advancement of Industry 4.0 has increased the attention to predictive maintenance (PdM) and fault diagnosis. In majority of the existing industrial setups, the choice of condition-based approaches is very popular, but it is limited because of a fixed threshold used to raise failure alerts. For reliable development and practical deployment of such models, it is essential to have a generalization capacity. However, it is challenging to build such models in production setups due to the non-stationary behavior of signals, noisy environment of the facilities, and scarcity of labeled datasets. In this work, the early failure detection of rotating component such as bearing is focused and generic domain inspired machine learning (ML) model is introduced. The proposed approach includes significant dedicated efforts to manually label publicly available vibration datasets which increases the volume and reliability of the data. The domain-based features such as impact root mean square (RMS), fatigue RMS and band features are applied. The modeling strategies are improved by considering exclusive bearing-wise data splits and training settings like individual and cross-dataset settings. This approach is tested on open-source datasets such as Franche-Comté Électronique Mécanique Thermique et Optique (FEMTO) and Xi’an Jiaotong University (XJTU) bearing dataset. The experimental findings demonstrate an improved generalization compared to other baselines.