Data-Driven Predictive Maintenance for Heat Exchangers: Real-Time Monitoring and Long-Term Performance Prediction Using Integrated ML Models

Abstract

This study addresses the high maintenance costs of heat exchangers in petrochemical processes by developing a deep learning-based predictive maintenance (PdM) model for performance monitoring and scheduling. Using a mathematical model, the overall heat transfer coefficient (U) was derived to evaluate heat exchanger performance, resulting in a performance indicator (DI). An artificial neural network-genetic algorithm (ANN-GA) technique was employed to create a real-time DI prediction model based on industrial process data. A long short-term memory (LSTM) model was then used to predict heat exchanger performance over 3 days using short-term operating data (12 h). The model's hyperparameters were optimized, achieving a real-time monitoring model with a mean absolute percentage error (MAPE) of 0.59% and a maintenance-cycle prediction model with an MAPE of 2.41%. This integrated system, akin to soft sensors, accurately predicted a 72-h performance profile using 12-h history data owing to our implemented data augmentation strategies, demonstrating robustness and potential for improving uptime and maintenance scheduling.