Combining Augmented Design and Statistical Learning Approaches to Address Multicollinearity in Small Data

• Activity Number: 341 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Tuesday, August 9, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Min Chen* and Christine A Zielinski and Charles L Baker
• Companies: ExxonMobil and ExxonMobil and ExxonMobil
• Keywords: Augmented Design; Statistical Learning; Multicollinearity; Small Data

Abstract

The GF-6 Sequence IIIH engine test evaluates automotive engine oil performance for high-temperature characteristics, including oil thickening (viscosity increase) and piston deposits. The existing small dataset (n=27) shows that engine oil performance varies with base oil formulations. However, among 30 predictors, many base oil properties are highly correlated, which causes issues in determining the effects of individual base oil properties on the engine oil performance. An augmented design was constructed for an additional 4 runs to reduce the multicollinearity among base oil properties. Various statistical/machine learning approaches such as random forest, LASSO, Elastic-Net, and stepwise selection were evaluated to identify the important base oil properties for engine oil performance. The parsimonious regression models based on stepwise selection show the best prediction accuracy and interpretability. NMR branching, 10% Boiling Point, Sulfur and Saturates are the best predictors for viscosity increase. Saturates, 90% Boiling Point and 254 nm UV are important predictors for weighted piston deposits.