A Data-Driven Method for Remaining Useful Life Prediction of Rolling Bearings Under Different Working Conditions

Efficiently extracting long-term sequence features is the key for deep learning models to accurately predict the remaining useful life (RUL) of mechanical bearings. However, completely ignoring local features may result in losing a large amount of prediction information. In addition, the existing soft thresholding function loses a large number of feature values when denoising, which affects the prediction performance of the network. A new temporal convolutional neural network with an adaptive activation function and attention mechanism (TCNAAM) for RUL prediction is proposed in this article. The network consists of two branches; one branch is to be utilized to adaptively select key local information in the time series using an adaptive convolutional scale attention layer. The other branch is feature denoising by adding an improved activation function after the temporal convolutional neural network during training, to preserve most of the original features. Therefore, TCNAAM can simultaneously extract long-term and short-term information and constrain all eigenvalues within a certain interval. The bearing dataset on Xi'an Jiaotong University and IEEE PHM Challenge 2012 is used to validate the effectiveness of the proposed TCNAAM. Experimental results show that the proposed TCNAAM performs better than some existing prediction methods for bearings RUL prediction applications under different operating conditions.