Utilizing Neural Networks for Image-based Model Predictive Controller of a batch Rotational Molding process

We present a data-driven modelling and control approach for batch processes utilizing information from thermal images for feedback control. This work is driven by the requirement of utilizing the thermal image data that is the sole output of the system for feedback control. The overall goal here, like in many batch processes, is to obtain products with quality variables which match the user’s specifications. The quality variables of the product cannot be measured online and is only measurable after the batch has terminated. The control problem is therefore not a setpoint tracking problem. We propose a multi-layered modelling approach. We first have a dimensionality reduction technique to reduce the high dimensional image to a set of few representative outputs. Then, we apply subspace Identification (SSID) to identify a Linear Time Invariant (LTI) State space (SS) model between the inputs and the reduced outputs, and finally we construct a Partial Least Squares (PLS) model between the terminal states of a batch (identified using SSID) and the product qualities obtained for that particular batch. This model is utilized in a Model Predictive Control (MPC) formulation. We demonstrate the working of the MPC by showing its ability to achieve products with good quality.