Development of a hybrid model for estimating the non-linear rolling bearing remaining lifetime

In recent years, the data-driven approach has emerged as the predominant method for estimating the Remaining Useful Life (RUL) of bearings. This approach offers notable advantages, including reduced reliance on expert knowledge in bearing health analysis. However, there are three significant drawbacks associated with data-driven models that have impeded the advancement of bearing RUL estimation. These drawbacks encompass limited data availability, constrained model generalization ability, and inadequate adaptability to the nonlinear degradation process of bearings. To address these limitations, this thesis presents an improved nonlinear method for predicting bearing RUL. The proposed model leverages multiple critical time-domain characteristics of bearings, which are identified using the Random Forest (RF) algorithm. These identified characteristics are then employed in constructing a Multi-Feature Fusion Neural Network (MFFNN), trained through the implementation of a Transfer Learning (TL) strategy. The prediction results obtained from the proposed model exhibit superior accuracy when compared to other data-driven methods, accompanied by enhanced generalization capabilities. Furthermore, this study delves into the exploration of hybrid models based on feature fusion networks. The fusion of the MFFNN with the Long-Short Term Memory (LSTM) model is particularly emphasized, as it outperforms other hybridizations in terms of prediction accuracy. By combining the strengths of these models, a more robust and accurate approach for estimating bearing RUL can be achieved. Fundamental to the above research, one scheme to utilise the ensemble strategy to broaden the perspective of developing the nonlinear bearing RUL prediction model.