Deep multi-scale multi-head attention network for aero-engine remaining useful life prediction

An aero-engine serves as the “heart” of an aircraft, playing a pivotal role in its operation and performance. The primary research area in aero-engine prognostics and health management is the prediction of remaining useful life (RUL), a crucial aspect of aircraft navigation safety. This research proposes a deep multi-scale multi-head attention network (DMMAN) to predict aero-engine RUL. Firstly, a multiscale feature extraction module is designed by adopting convolution kernels of different sizes, which allows the network to focus on both global and local degradation information at the same time. Then, a feature enhancement module is proposed using a multi-head attention strategy, which amplifies the significance of pivotal features while mitigating the influence of superfluous ones. Finally, a bidirectional gate recurrent unit-based feature fusion module is constructed for feature fusion, which can consider both historical and future information. To verify the performance of the network, an extensively used dataset is employed for experimentation. The experimental results conclusively demonstrate the superior performance of DMMAN compared to contemporary state-of-the-art (SOTA) methods for RUL prediction.