Traditional hyperparameter optimization (HPO) for aero engine Prognostics and Health Management (PHM) suffers from excessive computational overhead and poor knowledge reuse, hindering real-time deployment in complex aviation scenarios. To develop an efficient HPO algorithm that balances search speed and accuracy for medium-dimensional spaces in mission-critical PHM tasks. A Score-Guided Heuristic Algorithm (SGHA) is proposed, featuring a fine-grained knowledge-reusable framework. Core mechanisms include: (1) independent performance profiling to quantify individual hyperparameter contributions; (2) a targeted rotation optimization strategy to eliminate inter-parameter coupling; and (3) a score-weighted heuristic sampling mechanism with UCB-style exploration to dynamically adjust search trajectories. SGHA incorporates local fine-tuning and task-adaptive quantization (bin size 0.1) to optimize resolution and efficiency. Validations on the C-MAPSS FD004 dataset show that SGHA reduces RMSE by 18.06% and 23.12% compared to GA-Transformer and HyperOpt-LSTM, respectively. For thrust prediction, it achieves up to 90.2% error reduction over baseline LSTM. Furthermore, SGHA shortens optimization time by 30.87% compared to mainstream HPO methods (GA, QGA, RF) and exhibits strong cross-architecture generalization across LSTM, GRU, TCN, and lightweight Transformers, maintaining stable RMSE levels (32.5–34.2). SGHA provides a practical, robust, and high-efficiency solution for aero engine RUL and thrust prediction. By overcoming the limitations of traditional HPO in knowledge reuse, it establishes a solid foundation for the industrial deployment of real-time PHM systems.
Zhou et al. (Sun,) studied this question.