Data Driven Leak Detection in a Real Heat Exchanger in an Oil Refinery

Abstract

This study focuses on implementation of a data-based leak detection method in a heat exchanger in a petroleum refinery. We have studied on the two real leakage cases in a heat exchanger in Izmit TUPRAS Refinery. Leaks are one of the major problems that occur in operations. The autoencoder (AE) method is implemented for leak detection. Reconstruction error is used as the leak indicator. In case of leakage, the reconstruction value is expected to increase. For both cases examined, the reconstruction error is found to be around 1-5 under normal operating conditions. On the other hand, reconstruction error is observed to change between 10 and 60 under the conditions with leakage. Besides, the AE is able to indicate the start of one leakage case before the process engineers noticed it. © 2023 Elsevier B.V.