Reshaping Industrial Maintenance with Machine Learning: Fouling Control Using Optimized Gaussian Process Regression

Abstract

Reliable maintenance scheduling is essential for complex industrial equipment, yet traditional condition-based strategies with static warning thresholds often fail in fouling-prone processes or when feedstock composition fluctuates. This paper presents a predictive maintenance strategy based on the automatic selection of optimal kernel combinations in Gaussian Process Regression (GPR) through a recursive algorithm. The approach is applied to a vacuum distillation column processing used oil, a fouling-prone waste stream with variable composition. The algorithm performs an automated search and optimization of models through recursive combination of kernels and operators, following a greedy search strategy. The algorithm’s predictive capabilities are validated on five distinct data sets representing the evolution of the column’s pressure differential, a key indicator of fouling. Results show significant improvements, with the strategy reducing suboptimal operating time by 30-40% and, in some cases, entirely avoiding such conditions. Automation of kernel search and optimization ensures general validity for the proposed method.