Spatiotemporal non-negative projected convolutional network with bidirectional NMF and 3DCNN for remaining useful life estimation of bearings

Abstract

Remaining useful life (RUL) estimation for bearings is crucial in guaranteeing the reliability of rotating machinery. With the rapid development of information science, deep-learning-based RUL estimation has become more appealing as it can automatically establish the mapping relationship between the monitored data and the degradation states through feature learning. Vibration analysis via time–frequency representation (TFR) has shown great advantages for the detection of bearing damage in deep-learning-based prognostics. However, the following two problems remain: 1) insufficient or ineffective utilization of the data feature information, and 2) the requirement for huge computational resources, which still present challenges for the accuracy and efficiency of TFR-based prognostics. A novel RUL estimation approach called spatiotemporal non-negative projected convolutional network (SNPCN) is hence proposed. The approach can fully learn the spatiotemporal degradation features of bearing TFRs with high computational efficiency. In detail, the continuous wavelet transform (CWT) was applied as a TFR analysis method to reveal the nonstationary properties of the bearing degradation signals. Then, a newly proposed bidirectional non-negative matrix factorization (BiNMF) method was used to obtain the low-rank eigenmatrices of the TFRs and greatly compress the calculations in TFR-based prognostics. A three-dimensional convolutional neural network (3DCNN) was next constructed to learn the spatiotemporal degradation features in adjacent BiNMF eigenmatrices and construct the mapping relationship between the bearing RUL and current monitored data. Experiments on the PRONOSTIA platform demonstrate the feasibility and superiority of the proposed SNPCN-based bearing RUL estimation approach.