Q-Learning-Based Optimal Control via Adaptive Critic Network for a Wankel Rotary Engine

We propose a new Q-learning-based air-fuel ratio (AFR) controller for a Wankel rotary engine. We first present a mean-value engine model (MVEM) that is modified based on the rotary engine dynamics. The AFR regulation problem is re-formulated as an optimal PI controller for fuel tracking over the augmented error dynamics. Leveraging the generalized-Hamilton-Jacobi-Bellman (GHJB) equation, we propose a new definition of the Q-function with its arguments being the augmented error and the injected fuel flow rate. We then derive its Q-learning Bellman (QLB) equation based on the optimality principle. This allows online learning of a controller via an adaptive critic network for solving the QLB equation, of which the solution satisfies the GHJB equation. The proposed modelfree Q-learning-based controller is implemented on an AIE 225CS Wankel engine, where the practical experiments validate the optimality and performance of the proposed controller.