Performance improvements in machine learning approaches for fault detection and soft sensing in the process industry
Loading...
Authors
Mazibuko, Tshidiso
Issue Date
2021-12
Type
Dissertation
Language
en
Keywords
Process industry , Fault detection and diagnosis , Soft sensing , Principal component analysis , Long short-term memory , Dynamic autoencoder , Tennessee Eastman chemical process , Steam boiler control , Feature selection
Alternative Title
Abstract
The main focus of this research is on the application of machine learning in solving problems that have not been solved by the advancement in process simulation and automation tools in the process industry. These problems are the fault detection and diagnosis, and soft sensing of variables that are difficult and/or expensive to measure.
A literature review was conducted in areas where the application of machine learning was used to solve the problems related to fault detection and diagnosis, and soft sensing of process variables. Two case studies from the literature review were further extended,
with the aim of improving the performance of the machine learning approaches to these problems.
The first case study is on the detection of process faults for the Tennessee Eastman chemical process. In this case study, unsupervised sequential data-driven models such as
the long short-term memory autoencoder (LSTM autoencoder), dynamic autoencoder and the dynamic principal component analysis (PCA) are explored. The results show that the LSTM and the dynamic autoencoder improved the detection of five faults that were poorly detected in the original case study by at least 60%.
The second case study is the optimisation of a steam boiler control system using machine learning. In this case study, the contribution made is the use of feature selection in improving the performance of the machine learning models used in predicting the temperature of six zones in the boiler (to minimise overheating of tubes) and the oxygen
content on both sides of the flue system (to maximise combustion efficiency).
The results show that feature selection decreased the mean squared error (MSE) and mean absolute percentage error (MAPE) by 60% and 50% respectively.