Res-TCEANet: An expansive attention mechanism with positional correspondence based on semi-supervised temporal convolutional network for RUL estimation

Abstract

The accurate prediction of Remaining Useful Life (RUL) is crucial for various applications, and the relative position of time steps between features plays a significant role in this process. However, traditional deep learning models often struggle with extracting and positional corresponding information in temporal features, especially in the presence of noise and limited labeled data. To overcome these challenges, we propose a novel semi-supervised Residual-denoising Temporal Convolutional Expansive Attention Network (Res-TCEANet). This approach introduces a unique expansive attention mechanism (EAM) that enhances the modeling of long-term dependencies by addressing the positional correspondence of features across layers. The proposed EAM distinguishes itself from existing attention mechanisms by enabling TCEANet to model long sequences with a focus on positional coherence, resulting in more robust feature extraction. The Root Mean Square Error and Score of the proposed method on C-MAPSS dataset are 10.75, 12.27, 10.82, 12.33 and 114.82, 427.05, 161.96 746.93, respectively, which have demonstrated that our method achieves the start-of-the-art performance and outperforms other models.