A dynamic function allocation method for single pilot operations in commercial aircraft based on Bayesian-enhanced Q-learning

Human-machine collaboration is a key feature of Single Pilot Operations (SPO). With only a single pilot in the cockpit, workload monitoring and adjustment become even more critical compared to Dual-Pilot Operations (DPO). Hence, a dynamic function allocation mechanism must be established--increasing the Level of Automation (LOA) under high workload conditions and reducing it under low workload conditions to maintain situational awareness. To address the challenges of excessive subjectivity and limited knowledge transfer in the existing dynamic function allocation methods, this paper proposes a dynamic function allocation method based on Bayesian-enhanced Q-Learning (BQL). First, a Bayesian Network (BN) is constructed to predict Human-Machine System (HMS) performance, determining when reallocation should be triggered. Compared to the existing trigger mechanisms, this approach enables earlier activation while maintaining non-intrusive. Then, the BN-predicted HMS performance is integrated into the reward feedback for the reinforcement learning algorithm, allowing the system to continuously refine its strategy through interaction with the environment. Finally, flight experiments conducted in a low-fidelity SPO simulator, incorporating both objective physiological monitoring and subjective assessments, validate the effectiveness of the proposed method.