An adaptive dual distillation framework for efficient remaining useful life prediction

作者：Xiang Cheng, Jun Kit Chaw, Shafrida Sahrani, Mei Choo Ang, Saraswathy Shamini Gunasekaran, Moamin A. Mahmoud, Halimah Badioze Zaman, Yanfeng Zhao, Fuchen Ren · 发表于：Complex & Intelligent Systems · 年份：2025 · DOI：10.1007/s40747-025-01886-w · 被引用次数：7 · 研究领域：Fault Detection and Control Systems、Process Optimization and Integration、Advanced Control Systems Optimization

Predicting the Remaining Useful Life (RUL) of industrial equipment is essential for proactive maintenance and health assessment, particularly under the computational constraints of edge devices. While deep learning methods, such as Long Short-Term Memory (LSTM) networks, excel at modeling complex time series, their high computational cost often restricts real-time deployment. To address this challenge, we present an Adaptive Dual Distillation Framework (A-DDF) that transfers knowledge from a large LSTM teacher model to a lightweight bidirectional Gated Recurrent Unit (GRU) student model. Soft-target distillation refines predictive distributions to provide robust supervision and our correlation-based feature alignment preserves inter-feature relationships and prevents information loss. An adaptive weighting mechanism balances these two distillation strategies, enabling the student model to maintain high predictive accuracy while reducing model complexity. We validate our approach on NASA’s C-MAPSS dataset, which includes diverse operating conditions. A-DDF outperforms previous methods, achieving a 12% decrease in relative error (MAPE), improving prediction accuracy and stability. Ablation experiments show the dual distillation strategy improves predictive accuracy, surpassing single distillation approaches. Notably, the student model achieves a 5.34-fold compression rate, reducing parameters by 83%, while maintaining or exceeding the performance of the LSTM teacher model. Thes...