Aero-engine prognosis strategy based on multi-scale feature fusion and multi-task parallel learning

Aero-engine prognosis is helpful to ensure its safety and reliability, and effectively reduce the maintenance cost. However, the existing works only perform RUL prediction, ignoring the fault factors that lead to engine degradation. In addition, most prognosis methods can only extract single-scale features, ignoring the potential degradation features at other scales and layers. Therefore, this work proposes an aero-engine prognosis framework based on multi-scale feature fusion and multi-task parallel learning. In the proposed framework, multi-scale feature fusion blocks are designed to explore and fuse the potential degradation features of samples under different scales. And a layers concatenation block is constructed to integrate feature details from different layers and avoid losing useful information. Then a multi-task parallel learning block is constructed, and a joint loss function is developed for parallel learning of RUL prediction and fault diagnosis tasks. Meanwhile, a stacked image conversion method is proposed to integrate multi-sensor data with multiple cycles into image sample and make it contains more information beneficial to engine degradation. Finally, experimental results on CMAPSS and NCMAPSS datasets show that the proposed framework exhibits superiority over other state-of-the-art methods and demonstrates good generalization and robustness.