Industrial Data Science for Batch Manufacturing

• Published in: Computer Aided Chemical Engineering Vol. 53; pp. 2965 - 2970
• Main Authors: Vallerio, Mattia, Perez-Galvan, Carlos, Navarro-Brull, Francisco J.
• Format: Book Chapter
• Language: English
• Published: 2024
• Subjects: AutoML; Batch Manufacturing; Data Analytics; FPCA; Machine Learning; FPCA; Data Analytics; AutoML; Machine Learning; Batch Manufacturing
• DOI: 10.1016/B978-0-443-28824-1.50495-6
• ISBN: 9780443288241; 0443288240
• ISSN: 1570-7946

Batch processes are the main production methodology for a large number of manufacturing industries: e.g., chemicals, food and beverage, pharma, Levenspiel (1998). However, these processes are subject to high variability: raw material composition, initial condition, unit degradation, and their intrinsic nonlinear and dynamic nature. All these reasons make batch processes challenging to analyse, control and optimize, Kourti (2005), Rawlings (2017). Nowadays, data can be considered an additional asset of all manufacturing processes and with the introduction of IoT devices, the economic burden to capture even more data has dropped. This data overload does not help on its own. It generates an added value when it is turned into actionable information (e.g., reduce batch time, increased first-time-right production). This contribution aims at introducing the need to establish the field of industrial data science (Mowbray et al. (2022)), and by doing so it highlights the use of two well-established data analytics methods in the context of batch manufacturing analysis. A well-known industrial example will be used to demonstrate the newly defined workflows to apply these two machine learning methods to convert this high variability and apparent excess of data into valuable information. The first method involves AutoML analysis, Wilson (2017). We will present how to automatically summarize its properties into features and identify the most relevant ones via non-linear correlation analysis (e.g. random forest) for a batch process. Then, trajectory analysis and functional data exploration will be covered. However, before doing so, we will discuss the need to align the data time-wise to effectively use these techniques (Dynamic Time Warping). In this last step, we will use the Functional Data Explorer in JMP Pro to monitor and identify deviations, Silverman (2002). These two novel workflows show how to apply ML method to industrial data successfully.