%A M.A. Mohd fadzil
%A A.A. Razali
%A H. Zabiri
%A A.H. Che Hussin
%T Investigative Analysis of Automatic Mode Detection for a Lubricant Base Oil Production Plant Using PCA and Machine-Learning Models
%J ACS Omega
%L scholars18994
%O cited By 0
%R 10.1021/acsomega.3c07331
%D 2023
%X Lubricants are important fluids and are commonly used to suppress friction between two metallic surfaces and as a medium for heat transportation. In an industrial plant considered in this study, the base oil mode changes can only be detected based on the kinematic viscosity values obtained using lab analysis. Since the lab analysis data are only available every 8 h, detecting the change in the production modes for 4, 6, and 10 cSt and the transitions among them are significantly delayed, causing unnecessary off-spec products that have to be directed to the slopping tank. In this paper, the innovativeness of the work comes from the idea of trying to unravel the underlying pattern of the plant data that correlate to the changes in the base oil modes and using that to classify hourly the kinematic viscosity values. Hence, a novel industrial application is presented to predict the class of base oil mode change on an hourly basis that can significantly reduce the losses in terms of off spec products and sloping tank wastes. The modes are segregated into three classes based on the values of kinematic viscosity. The classes are C-1 (4 cSt), C-2 (6 cSt), and C-3 (10 cSt). Anything in between the stipulated thresholds is called transition T-12 (C-1 to C-2), T-21(C-2 to C-1), T-23 (C-2 to C-3), T-31 (C-3 to C-1), and T-32 (C-3 to C-2). To unravel the pattern, principal component analysis (PCA) is utilized on 42,000 operating plant data. After a thorough analysis, the third principal component provides the highest correlation to the eight classes of the base oil mode changes C-1 (4 cSt), C-2 (6 cSt), and C-3 (10 cSt) and the transitions T-12 (C-1 to C-2), T-21(C-2 to C-1), T-23 (C-2 to C-3), T-31 (C-3 to C-1), and T-32 (C-3 to C-2). This third principal component is then utilized together with plant process variable values as inputs to four machine learning models, namely, XGBOOST, Random Forest, and CatBoost algorithms to predict the mode of the base oil hourly. The overall comparison analysis shows that utilizing the XGBoost algorithm for the prediction of the eight classes of the base oil modes at a faster hourly rate results in the most consistent classification accuracy of 92.96% for the test set and 89.22% in the deployment set. This capability to predict the mode change in the hourly basis can significantly reduce the losses in terms of off spec products in the production line. © 2024 The Authors. Published by American Chemical Society.