A New Convolutional Prognostics Network with Adaptive Kernels for Estimating Remaining Useful Life of Bearings Considering Variable Speed

Remaining useful life (RUL) prediction of rolling element bearings has become an indispensable technology in modern industries for safety guarantee and maintenance optimization of machines, and deep learning-based prognostics methods are the research hotspot in recent years. Existing deep prognostics networks, however, usually focus on RUL prediction of bearings under constant operating conditions, and the variation of rotational speed of bearings is not explicitly considered during network construction and representation learning. To address this limitation, a new convolutional prognostics network with adaptive kernels, named adaptive multi-scale convolutional network (AMSCN), is proposed in this paper. The proposed AMSCN uses vibration and speed signals as the network inputs, and builds two novel layers, i.e., an adaptive convolutional layer and an information merging layer, to enhance the representation learning ability, including automatic scale-varying degradation feature extraction and representation recalibrations based on time-series information. The proposed AMSCN is evaluated using life-cycle monitoring data from accelerated degradation tests of bearings under variable speed conditions, and is also compared with some existing prognostics methods. Experimental results validate the advantages of the proposed AMSCN in terms of accuracy and convergence for RUL prediction of bearings considering variable speed.