Deep Spatiotemporal Convolutional-Neural-Network-Based Remaining Useful Life Estimation of Bearings

Abstract The remaining useful life (RUL) estimation of bearings is critical for ensuring the reliability mechanical systems. Owing to rapid development deep learning methods, a multitude data-driven RUL approaches have been proposed recently. However, following problems remain in existing methods: 1) Most network models use raw data or statistical features as input, which renders it difficult extract complex fault-related information hidden signals; 2) current observations, dependence between states emphasized, but their on previous often disregarded; 3) output neural networks directly used estimated most studies, resulting extremely volatile prediction results that lack robustness. Hence, novel prognostics approach based time–frequency representation (TFR) subsequence, three-dimensional convolutional (3DCNN), and Gaussian process regression (GPR). primarily comprises two aspects: construction health indicator (HI) using TFR-subsequence–3DCNN model, GPR model. signals are converted into TFR-subsequences by continuous wavelet transform dislocated overlapping strategy. Subsequently, 3DCNN applied spatiotemporal from construct HIs. Finally, can also define probability distribution potential function confidence. Experiments PRONOSTIA platform demonstrate superiority TFR-subsequence–3DCNN–GPR approach. degradation-related herein achieve highly accurate bearing with uncertainty quantification.