Actuator fault recovery with deep reinforcement learning in a linear model-based control framework: Application to a physical AUV