Dual-branch attention convolutional network with adaptive loss for remaining useful life estimation

Abstract The prediction of remaining useful life (RUL) of aero-engines is a central task in prognostics and health management, which is essential for ensuring safety, reliability, and cost-efficiency. This study proposes a convolutional network with dual-branch attention (DBAN) and adaptive loss to effectively capture both local and global features from multi-sensor degradation data for accurate RUL estimation. A local channel dynamic adjustment strategy is introduced to sequentially integrate channel and spatial attention within the convolutional backbone, thereby enhancing feature representation. Furthermore, a parallel dual-branch attention module is developed, in which the perturbation-enhanced branch employs dynamic masking to emphasize local features and suppress noise or outliers, while the trend-modeling branch utilizes dynamic channel weighting to focus on global features, enabling adaptive modeling of sensor contributions and long-term degradation patterns. In addition, an adaptive latency-aware weighted loss is defined to penalize delayed predictions and to dynamically adjust training weights. Experimental evaluations on the commercial modular aero-propulsion system simulation dataset demonstrate the effectiveness and superiority of the proposed DBAN compared to existing state-of-the-art methods.