Degradation Assessment of Bearings Using Deep Convolutional Inner-Ensemble Learning with Outlier Removal

Degradation assessment plays a significant role in increasing and managing the reliability and safety of mechanical systems, especially for key components. Hence, Health Index (HI) construction is a research field of interest. In conventional HI construction methods, performance features have to be selected manually based on experience and expert knowledge. This is a time-consuming procedure requiring a detailed understanding of the component. Furthermore, outliers usually exist in any HI constructed, and these can result in false alarms when HIs are subsequently used. To solve those problems, a new 'end-to-end' methodology, which combines Deep Convolutional Inner-Ensemble Learning (DCIEL) with an Outlier Removal method, is proposed for the degradation assessment of roller bearings. In the DCIEL algorithm, an inner-ensemble structure including multiple Deep Learning (DL) blocks is used to extract features directly from raw vibrational signals obtained from bearings. A novel DL block including two convolution layers and one nonlinear pooling layer is proposed. The obtained features are mapped into the HI using a Global Average layer, an Averaging layer, and a Logistic Regression layer. Finally, a novel outlier removal method, based on sliding thresholds, is proposed in order to detect and remove outliers in the constructed HIs. The effectiveness of the proposed method is verified using an open-source bearing health dataset. The proposed method demonstrates significant advantages across multiple HI assessment metrics in comparison with different DL blocks and commonly used HI construction methods.