Beans to Bytes: Grey-Box Nonlinear System Identification Using Hybrid Physics-Neural Network Models

Abstract

The advancement of neural networks in the last several years has yielded some astonishing results. However, the applicability to system identification and modelling dynamical systems still has a great amount of room for exploration. This thesis reviews different neural network architectures and their application to complex non-linear dynamic system identification. In particular, it uses the intricate process of coffee roasting as a case study to explore and demonstrate these techniques. Coffee roasting is a complex process that requires precise control to achieve the desired coffee quality. The ability to develop models that represent a system, i.e. system identification, is of great value to industry. Coffee roasting poses several challenges for system identification from complex chemical reactions occurring inside the bean, to temperature trajectories being dependent on several states that cannot be explicitly measured, such as moisture content, or reaction rate, making it an ideal candidate for exploring the application and limitations of neural networks. The primary contributions of this study are a proposed "grey-box" model that augments previously established physics based models, as well as illustrating the limits of LSTM, Deep NARX models using "one-step" forward prediction techniques. Although the study focuses explicitly on coffee roasting, the conclusions drawn are applicable to other similarly complex industrial and manufacturing processes.