Frequency Hoyer attention based convolutional neural network for remaining useful life prediction of machinery

Remaining useful life (RUL) prediction based on vibration data is a vital part of bearing prognostic and health management, which can be applied to formulate a suitable maintenance strategy. Recently, attention mechanism has been widely studied and applied in the field of prognostics, which can adaptively enhance the features for RUL prediction and weaken the features interfering with the accurate estimation of the condition. However, insufficient priori information has been provided to the deep learning model via the attention mechanism. The domain knowledge of the special structure and the characteristics of the bearing vibration signal is underutilized in the conventional attention mechanism. An innovative attention based RUL prediction model, called frequency Hoyer attention based convolutional neural network (FHA-CNN), is proposed in this study, which combines a deep learning model and signal processing method organically. The 1D convolutional layer and isometric empirical wavelet transform are developed to extract the latent representation of vibration signals from different scales. The proposed FHA is applied to calculate the weight of the feature map adaptively, in which three types of Hoyer index are adopted to comprehensively evaluate the contribution of each frequency part to the degradation of rolling bearings from the frequency domain perspective. To verify the superiority of the proposed method, two run-to-failure experimental dataset case studies are analysed. The obtained results indicate that the proposed FHA-CNN model exhibits a better performance than conventional deep learning-based RUL prediction methods. In addition, the proposed method concentrates on the special structure of the bearing vibration signal and provides a novel insight into the decision-making processes of deep neural networks.