An efficient turbofan engine RUL prediction framework based on AFE-LSTM architecture

Effective turbofan engine maintenance requires precise remaining useful life (RUL) forecasting, yet many existing deep learning approaches encounter a fundamental trade-off between predictive accuracy and computational efficiency, which restricts their adoption in industrial environments. Although complex Transformer architectures and multi-scale Convolutional Neural Network (CNN) models achieve strong benchmark performance, their computational demands and redundancy hinder practical application. To address this challenge, a computationally efficient RUL prediction framework is proposed, integrating an Adaptive Feature Enhancement (AFE) module with a streamlined attention-augmented Long Short-Term Memory (LSTM) model. The AFE module employs residual denoising connections to adaptively identify and suppress noise features across sensor channels while preserving and amplifying critical signals indicative of degradation trends. Dilated convolutions are utilized to capture multi-scale temporal dependencies. Unlike conventional static denoising techniques, the AFE module dynamically adjusts feature enhancement strategies in response to input data characteristics. The lightweight LSTM attention-based design efficiently captures temporal dependencies, sustaining high prediction accuracy while markedly reducing computational complexity. The framework demonstrates competitive accuracy on C-MAPSS dataset evaluation, achieving RMSE values of 11.50, 12.68, 10.29, and 13.39 for FD001–FD004, respectively, while substantially lowering computational requirements. The proposed framework offers an effective balance between computational efficiency and predictive performance, supporting real-time processing in resource-constrained environments and providing a viable solution for industrial deployment.