Application of a Predictive Maintenance Strategy Based on Machine Learning in a Used Oil Refinery

Aug 4, 2026

Upload by Maeve

The Itelyum Regeneration used oil re-refining plant in Pieve Fissiraga operates a condition-based maintenance policy for the thermodeasphalting (TDA) T-401 column, using the real-time pressure differential (ΔP) between column top and bottom as a fouling indicator. Maintenance triggers when ΔP exceeds an empirical threshold, keeping T-401 within normal operating limits but underperforming for non-conventional used oils. A Gaussian Process Regression machine-learning model predicts ΔP evolution, reducing suboptimal operation time and enabling a dynamic, data-driven maintenance strategy adaptable to variable feedstock composition.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

AccessFree to download

Introduction

Materials and Methods

Current Maintenance Strategy

Application of a Predictive Maintenance Strategy Based on Machine Learning in a Used Oil Refinery Francesco Negria,b, Andrea Galeazzib, Francesco Galloa, Flavio Manentib,* aItelyum Regeneration S.p.A., Via Tavernelle 19, Pieve Fissiraga 26854, Lodi, Italy bPolitecnico di Milano, CMIC Dept. “Giulio Natta”, Piazza Leonardo da Vinci 32, Milan 20133, Italy *[flavio.manenti@polimi.it](flavio.manenti@polimi.it) Abstract The Itelyum Regeneration used oil re-refining plant in Pieve Fissiraga currently employs a condition-based maintenance strategy for its thermodeasphalting (TDA) section, particularly focusing on the TDA T-401 column. This strategy involves monitoring the real-time pressure differential (ΔP) between the column's top and bottom, which increases in time due to fouling phenomena. Maintenance is scheduled when ΔP exceeds a predetermined empirical threshold, ensuring that the T-401 column operates within normal operations limits. However, this approach has limitations with non-conventional used oils. To address this, a data-driven machine learning algorithm, previously successful in predicting key performance indicators ofthe PH-401B furnace in the TDA section, was applied to the T-401 column datasets. This algorithm, based on Gaussian Process Regressions, effectively predicts the evolution of ΔP and reduces the time during which T-401 operates in suboptimal conditions. The implementation of this machine learning approach marks a significant improvement in the maintenance strategy, shifting from a static, condition-based approach to a dynamic, predictive one, thus ensuring more efficient and reliable operations, even with non-conventional used oil. Keywords: Data-driven, Machine learning, Predictive maintenance, Thermodeasphalting, Used oil. 1. Introduction Maintenance is a crucial aspect of every industrial plant to ensure continuity of operations and safety of the workers. Typical approaches in the European industrial context are corrective, preventive, opportunistic, condition-based, and predictive maintenance (Bevilacqua and Braglia, 2000) . The combination of a vacuum distillation column and its feedstock fired heater are critical pieces of equipment in crude oil refineries and used oil re-refineries that suffer from fouling due to the characteristics of the heavy hydrocarbon feed, thus requiring careful maintenance planning (Fuentes et al., 2007; Morales-Fuentes et al., 2014) . One of the most robust and efficient processes for the regeneration of used oil is based on the patented Revivoil® technology, and is currently operated in the Itelyum Regeneration re-refining facility in Pieve Fissiraga, Lodi, Italy (Gallo, 2016) . The maintenance approach in the case of the TDA section of the process is typically conditionbased, using the pressure differential across pieces of equipment as a sentinel key performance indicator to be monitored. Maintenance is planned once the parameter overcomes a warning threshold value. This static approach is typical for the refining industry, where fouling is an ever-present problem. Data-driven approaches have shown good results in modeling fouling in refinery equipment such as heat exchangers, with better fitting compared with equation-based, mechanistic modeling approaches, which instead often show poor results due to the extremely complex and partially random nature of fouling phenomena (Mei et al., 2023) . This work uses said data-driven algorithms to develop a dynamic predictive maintenance strategy that is more effective in avoiding runtime within suboptimal operating regions and it is more adaptable to changes in feedstock composition, a typical situation for used oil waste. 2. Materials and Methods 2.1. Current Maintenance Strategy The thermodeasphalting (TDA) section of the Revivoil® process works by fractionating dehydrated used oil in a vacuum column. The main products from the TDA T-401 column are three semi-finished base lube oil cuts. The TDA column sta

FAQs

What parameter does the current maintenance strategy monitor for the TDA T-401 column?

It monitors the real-time pressure differential (ΔP) between the column top and bottom, which increases over time due to fouling.

What is the limitation of the static condition-based maintenance approach?

It relies on a predetermined empirical ΔP threshold, which can be inadequate for non-conventional used oils.

How does the machine-learning approach improve maintenance decisions?

A Gaussian Process Regression model predicts the evolution of ΔP, shortening the time the column operates in suboptimal conditions and shifting toward a dynamic predictive policy.

Aug 4, 2026

Upload by Maeve