Multiscale attentional residual neural network framework for remaining useful life prediction of bearings

Traditional deep learning methods do not effectively extract degradation features from vibration signals widely used for remaining useful life (RUL) prediction while avoiding the gradient problem. This paper proposes a residual neural network framework with multiscale attention mapping (ResNet-MA) based on the vibration signal's characteristics to solve the above problems. In the framework, we first use the decomposed vibration signal processed by the ensemble empirical mode decomposition method (EEMD) as the model's input. To improve the neural network's ability to extract degradation signals, channel attention mapping, time attention mapping, and multiscale pooling methods are used in ResNet-MA. Experimental verification and analysis are carried out with the available data sets, which show that the proposed method's prediction accuracy is 14% higher than the residual neural network of deep learning before the improvement, and 3-6% higher than the state-of-the-art related algorithms.

Automated summary

The paper introduces a residual neural network framework ResNet-MA based on vibration signal characteristics for effective extraction of degradation features in deep learning. Experimental results demonstrate a significant increase in prediction accuracy compared to previous methods.