Causal Meta Multiple-Instance Learning for Interpretable Stall Precursor Detection

Abstract Aerodynamic stall-induced loss-of-control in-flight (LOC-I) remains a leading cause of fatal aviation accidents, where the timely identification of stall precursors is critical for hazard mitigation. However, existing identification methodologies struggle with data scarcity and the lack of robust validation frameworks. To address these challenges, we propose causal meta multiple-instance learning (CaM-MIL), a novel framework designed to identify stall precursors in data-scarce scenarios while ensuring both interpretability and cross-aircraft generalization. The methodology leverages causal-aware multiple-instance learning (MIL) to derive precursor probabilities from bag-level labels, thereby preserving essential temporal causal relationships. Furthermore, meta learning is integrated to facilitate efficient model adaptation across different aircraft types. To address the gap in precursor validation, we refine the area over perturbation curve (AOPC) metric with mean imputation to evaluate precursor utility in the absence of ground-truth labels. Additionally, a reinforcement learning (RL)-based intelligent agent is developed to verify whether the identified precursors provide actionable early warnings through closed-loop simulations. Experimental results demonstrated the superiority of CaM-MIL, which achieved an AOPC score of 106.88, outperforming state-of-the-art baselines by 14.69%. The RL-based validation further confirmed that the system provides effective early warnings, bridging the gap between offline precursor discovery and real-time stall risk monitoring.