Creative Commons

This dissertation was written as a part of the MSc in Data Science at the International Hellenic University. Fault detection for industrial chemical processes has gained the interest of the scientific community for the last years. There have been carried out several attempts to early detect faults on Industrial chemical processes, as industries seek to optimize their production. Those conducted mainly in this century because of the exponential growth of computing power, Machine Learning, and Deep Learning concepts. Due to Machine learning and Deep Learning success over many different subjects related to discovering patterns in non-linear time related data, early fault detection could be possible. The present thesis implements Machine Learning algorithms and tries to early detect the occurrence of a fault, through experimentation using the real-life data that Tennessee Eastman Process provides. That is happening by using Recurrent Neural Network un extension of Deep Learning and especially a model called Long Short-Term Memory, which has good results on similar concepts related to time series data. To improve the efficiency of the model regarding pattern extraction on the data that could lead to detection, another algorithm that is derived from Statistics is implemented, the Principal Components Analysis. Combining those two models, interesting results have arisen. Finally, the thesis is enhanced by an extended research on literature to conclude in the models that have been used and compare and evaluate the results.