Gated Branched Convergence Network for Upset Recovery in Full-scale Transport Aircraft Using Reinforcement Learning

Aircraft loss of control in-flight remains the primary threat to modern aviation safety. This study proposes a reinforcement learning-based autonomous recovery system to address this challenging problem. Our approach first establishes a simulation environment incorporating a full-scale NASA Generic Transport Model with turbulence, pilot behavior, and uncertainties. Then, the gated branched convergence network (GBCN) with gated recurrent units-based temporal feature extraction and gated convergence mechanism is designed to optimize the tradeoff between altitude loss and recovery time. Monte Carlo simulations confirm GBCN's superior recovery performance, demonstrating an 81.61% success rate with balanced recovery time and altitude loss. The system maintains robustness against ±10% aerodynamic uncertainties and turbulence. The decision-making logic is interpretable through gradient-weighted class activation mapping visualizations. This research provides both the theoretical foundations and inspiration for flight training programs for the next-generation intelligent aviation safety solution.