Machine learning for early detection of distillation column flooding

Adebayo, Opeoluwa (2025) Machine learning for early detection of distillation column flooding. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (5MB)

Abstract

Distillation column significantly affects the overall energy efficiency of a process plant. Poor performance of a column can result from faults, such as reflux failure, change in tray efficiency, change in feed temperature, etc. Flooding is one of the severe consequences of the faults attributed to a distillation column. During flooding, products go off-specification, and there is a tendency for a complete shutdown of the production process. In recent years, machine learning (ML) methods have been widely employed in process engineering for their ability to discover important patterns in data. One of the applications of ML is in predicting distillation column flooding. The supervised ML methods can predict flooding by forecasting the pressure drop across the column. The challenges of applying supervised ML methods for predicting flooding in distillation columns include a lack of large volumes of flooding data, the potential for overfitting, and long training time in some cases. A large amount of flooding data combined with normal operation data is needed to train the supervised ML algorithms for flooding detection. Therefore, it is important to identify flooding data sets from the operational data. However, flooding data sets are rare compared to normal data sets, which leads to an imbalanced data set. In this research, we address the data scarcity issue surrounding the application of supervised ML for flooding prediction by utilizing time-series generative adversarial networks, a framework that uses deep learning algorithms to generate synthetic data by preserving the temporal order in the original data. Additional flooding data sets are generated using this framework. Supervised ML algorithms are trained and tested to forecast the pressure drop of the column. Classification of the column data (i.e., flooding or not flooding) is done using clustering. This method is compared with predicting flooding using popular unsupervised ML methods such as principal component analysis (PCA) and autoencoders; these are unaffected by the data imbalance. Results show that by applying supervised ML algorithms to the sensor data of the distillation column, flooding conditions can be detected 19 minutes in advance and up to 60 minutes before it fully develops. This outperforms the PCA and autoencoders, which are popular unsupervised ML methods.

Actions (login required)

View Item