The aviation industry faces mounting pressure to improve safety standards while reducing maintenance costs and extending aircraft service life. Conventional scheduled maintenance practices, though regulatory-compliant, routinely result in unnecessary component replacements and costly operational disruptions. More critically, unexpected in-service failures — those occurring without prior warning — impose severe schedule penalties and safety risks that purely schedule-driven approaches cannot address. Existing data-driven solutions to this problem have concentrated on improving the predictive accuracy of Remaining Useful Life (RUL) estimation for aircraft engines, leaving the broader safety and operational interpretability requirements of maintenance engineers largely unaddressed. This work proposes a physics-informed CNN-LSTM architecture for RUL prediction on the NASA CMAPSS benchmark that prioritises safety and operational interpretability over raw accuracy. The model integrates four deliberate design decisions: (1) physics-derived features encoding thermodynamic degradation mechanisms, providing domain-grounded inputs rather than purely statistical ones; (2) a temporal attention mechanism, enabling the model to weight the most degradationrelevant time steps; (3) an asymmetric aviation penalty loss function, which penalises overestimation of RUL more heavily in the critical failure zone, directly encoding the asymmetric cost of unsafe predictions; and (4) degradation velocity tracking and SHAP-based sensor attribution, giving maintenance engineers an operationally actionable view of which sensors are driving each engine’s health trajectory. An ablation study quantifies the individual contribution of each component across all four CMAPSS sub-datasets. This work proposes not a more accurate RUL predictor, but a safer and more operationally interpretable one — where every design decision, from loss function to explainability structure, is motivated by what a maintenance engineer actually needs.