Quantifying Pilot Operational Deviations Based on Flight Data via Inverse Reinforcement Learning and Conformal Prediction

Abstract

Threshold-based flight data monitoring (FDM) identifies parameter exceedances but provides limited visibility into gradual operational deviation within nominal operating bounds. This study presents a three-layer framework for continuous operational deviation quantification from quick access recorder (QAR) data. A standard operating procedure (SOP)-constrained Gaussian mixture model–hidden Markov model (GMM-HMM) infers 27 procedural states from flight parameters, aligning latent state representation with operational procedure. From these state sequences, maximum entropy inverse reinforcement learning (IRL) recovers an expert value function that estimates execution quality at each time step. Split conformal prediction calibrates detection thresholds with distribution-free coverage guarantees, while counterfactual analysis traces identified deviations to their procedural origin. Learned from real flight data, the recovered reward function assigns 1.6 times greater weight to lateral and speed stability than to vertical stability. Split conformal calibration yields an empirical false alarm rate of 4.2% under a nominal 5% significance level. Counterfactual analysis identifies vertical speed management as the dominant source of procedural deviation. These results support the proposed framework as a candidate continuous safety performance indicator that surfaces gradual operational deviation that threshold-based exceedance monitoring would not flag.