Model Predictive Control with Adaptive PLC-based Policy on Low Dimensional State Representation for Industrial Applications

Abstract

In the modern era of manufacturing automation, the integration of sensor technology into the system ensures that data acquisition and analysis from complex systems become more efficient than ever. With the support of such developments, artificial intelligence-powered control in industrial control domains gains popularity and enhances the traditional human-based PLC control, where the machines can monitor themselves, learn from the experience, and make their own decisions. However, despite advances in sensor technologies, there are some limitations of the current applications of sensors in industries, for instance, sensors for observing the current status of the system often provide Boolean output data instead of continuous output. Therefore, such limitation forms a low dimensional state representation of the system, which could be problematic to develop a self-control policy, e.g. using a model-free deep reinforcement learning. In this paper, we present an effective model predictive controller with adaptive PLC-based policy on low dimensional state representation specifically for industrial control domains. First, we learn the model of the production system using the deep learning method to get the representation of the system dynamics, in case its digital representation is not available. Second, we set up a native implementation of model predictive control. Third, we enhance the model predictive control with adaptive PLC-based policy. The proposed method is implemented into a bulk good system showing its potential to self-optimize the system by satisfying the production objective without overflow and low power consumption.