Aplicação de metodologias estatísticas e machine learning para validação de analisadores contínuos para indústria de celulose

Abstract

Computational devices are more and more becoming the everyday tools that scientists and engineers turn to for efficiently and quickly handling large data sets. The quality of the data points allows for rapid decision making along the processes of industries. For example, in the paper milling and manufacturing process there is a high risk of variance and error in the quality assurance process, which typically occurs in laboratories and are performed by technicians. However, with the use of sensors and measuring devices these errors can be greatly reduced to produce greater returns on processes, cost reduction, resource allocation, and the statistical confidence and quality of data. The goal of this project is to validate the use of measuring devices on a network by using statistical modeling and machine learning, such as linear regression, decision trees, and random forest, to obtain the higher quality laboratory measurements and data points along the manufacturing process. The implementation of statistical methodologies was executed using the Python language for both statistical tools and construction of the virtual analyzer through Machine Learning. The statistical results were satisfactory, with some spurious errors. The soft sensor, on the other hand, presented poorly representative prediction results with an R² of 62,3% after optimization, requiring a greater amount of data for statistically satisfactory results.