Data-driven remaining useful life estimation of a fouled plate heat exchanger

Abstract

Fouling-induced heat exchanger performance degradation brings about the need for timely maintenance to minimize economic and energy losses as well as reduce chances of system failure during long-term operation. An accurate real-time estimation of remaining useful life can drastically shorten downtime and improve process control. In this paper, we propose a novel coupling of machine learning, Kalman filter and Monte Carlo simulations to predict the remaining useful life of a fouled brazed plate heat exchanger without the need for run-to-failure data, with emphasis on practical model deployment. Firstly, an offline-trained Bagged Tree ensemble is employed to predict the reference performance of the heat exchanger in real-time, which is then compared to current sensor-measured performance of the unit. The progressing mismatch between both metrics is exploited for underlying degradation drift identification with the adaptive linear Kalman filter. Finally, the identified degradation model is propagated to the terminal threshold within a Monte Carlo simulation, obtaining a population of predicted remaining useful life values. The framework’s prognostic capability is demonstrated on three distinct experimental case studies of brazed plate heat exchanger fouling. We illustrate an accurate adaptive prediction of remaining useful life, even after dynamic changes of degradation progression in presence of large flow-rate induced disturbances. In addition, we provide a comprehensive outlook for future work together with possible framework extensions

Topics & Concepts