A generative dual-input model based on architectural computational optimization and multi-attention mechanism for remaining useful life prediction

Abstract

Remaining Useful Life (RUL) prediction is one of the key technologies to ensure the smooth operation of aircraft engines. Previous studies have focused on improving prediction accuracy but overlooked issues related to model efficiency with long sequences and the bottleneck in information utilization. This paper proposes a generative dual-input model based on architectural computation optimization and a multi-attention mechanism. The model aims to break the inherent architectural limitations that reduce the model’s computational load and provides an update to the traditional Transformer at the architectural level. Specifically, a probabilistic sparse self-attention mechanism and an additive attention mechanism are updated as the basic internal components of the model. This design allows the model to achieve outstanding long-sequence processing efficiency and strong heterogeneous information fusion capabilities. Additionally, the concept of hierarchical decomposition is implemented in the data embedding process. This pattern cleverly connects the additive attention block while providing it with heterogeneous attention inputs to discover the intrinsic characteristics within the sequence. Finally, the complete degradation subsequences and short-term degradation slices are synchronously input into the model. This mechanism enables dependency discovery at the subsequence level. Satisfactory results were achieved on the C-MAPSS dataset, demonstrating the superiority of the model. Moreover, it outperforms many existing models in terms of balancing prediction accuracy and model scale.