Answers are grounded in this paper and cite their sources.
Predictive maintenance (PdM) is essential for reliable and cost-efficient turbine-based power generation, where thermo-mechanical degradation, nonlinear dynamics, and sensor noise challenge conventional maintenance strategies. This paper proposes a unified PdM framework that combines physics-motivated stochastic degradation modeling, deep and ensemble learning, probabilistic filtering, and reinforcement learning. Degradation is modeled as a partially observable stochastic process; a hybrid of a bidirectional LSTM with multi-head attention and a gradient-boosted ensemble estimates the remaining useful life (RUL); Kalman and particle filters with a monotonic degradation constraint quantify predictive uncertainty; a penalized Cox model provides interpretable hazard estimates; and maintenance scheduling is solved as a Markov decision process with a deep Q-network. On NASA C-MAPSS FD001, the hybrid achieves a test RMSE of 14.37±0.30 cycles over five seeds, competitive with attention-based state-of-the-art models, and the 90% prediction intervals of the particle filter reach an empirical coverage of 94%. On the multi-condition and multi-fault subsets FD002, FD003, and FD004, the RMSEs are 19, 12, and 20 cycles. Hazard ratios and SHAP attributions both identify Ps30 (s11) as the dominant degradation indicator, and the learned maintenance policy lowers the cost rate by 45% relative to corrective maintenance, but it does not surpass a tuned threshold rule and is sensitive to the discount factor. Transfer to field turbine data remains to be demonstrated.
No takes yet. Share an insight, caveat, or question.
Aqeel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: