Reinforcement Learning-Based Supervisor System Proposal for Fault-Tolerant Control of Direct Fired Heater

This work proposes a reinforcement learning-based supervisor system that incorporates automatic fault detection and fault-tolerant control in a fired heater plant, furnace, used to raise the temperature of crude oil for post-processing purposes. The faults considered are associated with the plant’s operating conditions, including the temperature sensor. The supervisor system contemplates supervised-trained neural networks to build a fault detector and an estimator of the controlled variable, a virtual sensor, and a reinforcement-trained neural network for the fault-tolerant controller; specifically, the Monte Carlo algorithm is implemented. Computational simulations illustrate the supervisor system’s functionalities, and a discussion of its physical implementation is presented.