Control industrial por aprendizaje reforzado

In this article, the implementation of the flow control technique with artificial intelligence (DDPG) is presented in a fully instrumented functional prototype with industrial sensors and actuators, simulating flow recirculation through three tanks. The methodology used for process identification (first-order plus dead time model (FOPDT)) through ClientServer OPC communication with Matlab® is presented. The design of the reinforcement learning algorithm and its adaptation in the learning environment with experimental data are also presented. The simulation results were satisfactory compared to traditional control techniques, demonstrating robustness against forced disturbances. Finally, the implementation of reinforced learning control integrating TIA Portal and Matlab (trough a PLC-S7-1500 controller) was evaluated with a reference of 600 l/h, achieving 0% overshoot with a settling time of 22s. Compared to other control systems, a better response in settling time and overshoot-free control was observed. Finally, perturbations were applied to the system, observing their effect in relation to the flow.