Remaining Useful Life Prediction of Bearings Based on Multi-head Self-attention Mechanism, Multi-scale Temporal Convolutional Network and Convolutional Neural Network

Abstract

Accurate remaining useful life (RUL) prediction of bearings is critical to the safety of the modern manufacturing industry. In this study, an effective end-to-end deep learning framework consisting of a feature extraction module (FEM) and a feature fusion module (FFM) is proposed to predict bearings RUL. In the FEM, an intelligent network architecture fusing a temporal convolutional network and a multi-head self-attention mechanism (MSM) is designed to extract multi-scale temporal features from raw data. Subsequently, the MSM is introduced to capture attention representations of features at different scales from multiple perspectives. All the attention representations are concentrated into an advanced feature vector as the input of the FFM. In the FFM, one-dimensional convolutional neural network layers with the ability of continuous down-sampling are designed to deeply fuse the internal information contained in the feature vector and output the predicted bearings RUL. Extensive experiments on PRONOSTIA dataset analyze the contribution of the framework's components and indicate that the proposed framework greatly outperforms the existing methods.