LBTCN with Self-Attention for Bearing's Remaining Useful Life Prediction

Bearings are the core components of mechanical equipment, making it crucial to monitor their performance during operation. Currently, deep learning-based methods for predicting remaining useful life (RUL) provide a data-driven approach to assess bearing performance. Temporal Convolutional Networks (TCNs) are one such architecture capable of parallel computation of sequence data, strictly adhering to causality. While TCNs prevent future data leakage, they overlook the value of nearby and past data in offline modeling. This study proposes an LBTCN architecture that utilizes bidirectional information within local windows while maintaining global causality, enhanced by a self-attention mechanism to highlight valuable information. Experimental results demonstrate that the proposed method achieves higher prediction accuracy compared to typical prediction networks.