Fouling-characteristic transfer learning for improving remaining useful lifetime prediction in heat exchange unit

Accurately predicting the remaining useful lifetime (RUL) of heat transfer units is essential for optimizing maintenance schedules and ensuring efficient operation in industrial processes. Traditional models often struggle with varying components, operating conditions, and limited training datasets, while none have explored how fouling behavior can be shared across different fluid characteristics. The current study introduced a fouling factor transfer learning-based long short-term memory model, which utilized pre-trained fouling factor representation from crude oil to improve RUL predictions for its derivatives, such as asphaltene and olefin, and extended the approach to other fluids, such as glycerin, across different unit operations. The proposed model achieved notable improvements, with RUL prediction accuracy reaching up to 99.6% for asphaltene and 96.8% for olefin, while maintaining robust performance for glycerin (despite domain discrepancies), with an average prediction error of 7 days in glycerin case study. In addition, the model was computationally efficient, reducing training time by 50% for asphaltene and olefin and by 9% for crude oil, underscoring its adaptability. By applying shared fouling dynamics across different fluids, the proposed model effectively addresses challenges related to limited data availability, enhances generalization across chemical processes, and offers a more reliable and efficient tool for predictive maintenance strategies in petrochemical industries.