A Novel Convolution Network Based on Temporal Attention Fusion Mechanism for Remaining Useful Life Prediction of Rolling Bearings

Abstract

Rolling bearing is one of the core components of modern machinery and is widely used in rotating machinery. It is of great significance to judge the running state and predict the remaining useful life (RUL) of bearings for preventive maintenance of rotating machinery. Due to the complexity of the fault mechanism, traditional prediction methods cannot clearly describe the relationship between local and global temporal features in bearing vibration signals. To overcome this shortcoming, a novel convolution network based on temporal attention fusion (TAF) mechanism, i.e., TAF convolutional network (TAFCN), is proposed in this article. Its core part is a TAF module, which consists of a separable temporal self-attention (STSA) submodule and a competitive TAF (CTAF) submodule. In particular, the STSA submodule focuses on the internal correlation of local temporal features, and the CTAF submodule aims to enhance the extraction and fusion of global temporal features across different levels. The comparison results based on XJTU-SY datasets show that the TAF module is robust to signal disturbance and noise, and the prediction accuracy of TAFCN for the RUL of rolling bearings is also better than some existing models.