Executive Summary : | A model that maps the operating conditions with product properties is crucial in optimization and control algorithms for the energy and cost-efficient design of industrial processes. Often built using physics-based models, over time, these maps grew to be computationally intensive, thus becoming limited in scope and application. This limitation and the advent of Data Science gave way to Digital Twins, i.e., the data-driven surrogates that built mathematical models devoid of any physics of the system. While they were fast and had a significantly high scope, the lack of information about the first principles of the system led to several questions on the predictions and thereby the use of these models. This brought Physics-driven Machine Learning and Explainable Artificial Intelligence to the forefront of the Industrial Internet of Things. In this project, an idea to build an Explainable data-driven model based on Deep Learning using parallel programming, to map the operating conditions with product properties is proposed for application in the design of Coal Flotation Columns (CFCs). The Computer Vision-based model is then planned for real-time implementation in a Reinforcement Learning (RL) based nonlinear control algorithm to produce clean coal with the highest quality, lowest energy consumption, and least carbon footprint. Coal is an important energy source whose quality is often measured in terms of Ash content. With a mining rate of 716 million metric tons per year (measured in 2018), India is the second-largest producer and consumer of coal. Since the Indian coals are known to have high ash content that reciprocates inversely with the coal quality it is important to refine the coal characteristics such as ash content in a coal preparation plant using CFCs. However, the lack of a readily implementable end-to-end map of the process has led to the erroneous design and operation of CFCs using heuristics, creating several safety constraints in addition to energy and economic imperfections. The proposed idea aims to eliminate the heuristics by mapping the operating conditions of CFCs with coal characteristics through images of the froth in CFCs and human expertise incorporated using Computer Vision and Fuzzy Logic, respectively. The data of operating conditions, froth images, and coal characteristics will be generated in a lab-scaled CFC. Thus, this study will make a significant contribution to the development of a smart online system capable of generating froth images and predicting coal quality efficiently in real-time through Explainable Deep Learning. Such a trained model will be implemented to control the experimental setup of CFC using an online control system incorporating the logic of RL. This project has immense scope and applicability in an organization such as IIT-ISM Dhanbad, which is also in the vicinity of major mining industries that might be interested in future collaboration and large-scale implementation of the proposed idea. |