Applied Predictive Machine Learning Algorithms in Bio-Composite Manufacturing and Development of Real-Time Digital Process Twin

Author

Chloe M. Ericksen, University of Tennessee, KnoxvilleFollow

Orcid ID

https://orcid.org/0000-0001-5030-2531

Date of Award

5-2022

Degree Type

Thesis

Degree Name

Master of Science

Major

Forestry

Major Professor

Timothy M. Young

Committee Members

Dr. Bogdan Bichescu, Dr. Sheng-I Yang, and Dr. Micholas Smith

Abstract

Much of the manufacturing industry is focused towards transitioning from Industry 3.0 towards Industry 4.0. Data science is the basis of Industry 4.0 for the bio-composites industry.

The bio-composites industry consists of enterprises which produce products from renewable raw materials. This research project was applied, and predictive models were developed of the strength properties of final products for a particleboard mill. Quality control measures in bio-composite manufacturing tend toward reactivity given the time from when samples are taken from the production line and destructive test data are presented to operations personal. The implementation of a feedback system for final board strength properties will allow for a more proactive approach to quality control and process improvement. A gap in the literature, identified by Liao et al (2017) found most Industry 4.0 research done in labs (95%), leaving little applied research (5%). This research aims for a better knowledge of how useful and feasible implementing machine learning algorithms in a bio-composite manufacturing setting are.

Quality control lab data was combined with process data to create fifteen predictive models. The merged data sets established a link between final board strength properties and process parameters from sensors along the process. The database starts March 3^rd, 2021 and ends January 4^th, 2022. The dataset was divided into seven product-grade categories by board strength property and grade. The algorithm ‘Multivariate Adaptive Regression Splines - four interaction terms’ (MARS4) performed best across all product categories in 10 validation studies. The R² of MARS4 in ten validation studies ranged from 0.588 to 0.845 (𝑋¯ = 0.731) for all internal bond product-grade categories. The R² of MARS4 in ten validation studies ranged from 0.745 to 0.923 (𝑋¯ = 0.826) of all modulus of rupture categories. Weighted predictions of the three top performing models (MARS4, MARS3, MARS2) in validation had promising results.

Automated decision making, which is powered by predictive analytics, is a foundation of Industry 4.0, which goes beyond Industry 3.0. The real-time feedback loop had to be simulated in this study due to an unforeseen occurrence and the subsequent disconnection of the modeling servers from the process servers.

Recommended Citation

Ericksen, Chloe M., "Applied Predictive Machine Learning Algorithms in Bio-Composite Manufacturing and Development of Real-Time Digital Process Twin. " Master's Thesis, University of Tennessee, 2022.
https://trace.tennessee.edu/utk_gradthes/6406