Remaining useful life prediction of rotating machinery based on a multi-scale separable convolutional network

The remaining useful life (RUL) prediction models of rotating machinery based on deep learning usually take the data from different sensors as their input. However, these models cannot effectively consider the multi-scale natures of degraded features and lack explicit learning mechanisms, thus compromising the accuracy and generalization of prediction models. This paper proposes a multi-scale separable convolutional network (MSSCN) model to track the degradation state of rotating machinery, which can consider the cross-channel correlations at multiple scales and process feature recalibrations. Firstly, multiple parallel separable convolutional channels are employed to extract the degradation features of different scales from the original signals. Secondly, a multi-scale convolution mechanism and the squeeze and excitation (SE) units are used to effectively integrate these degraded features and enhance informative feature maps but suppress useless ones. Finally, the automatically learned high-level representations are input to the regression layer to predict the RUL of a rotating machinery. The proposed MSSCN model was validated with the 2010 PHM tool wear dataset. The experimental results demonstrate that the MSSCN model has better prediction accuracy and application potential in RUL prediction of rotating machinery.