This dissertation was written as a part of the MSc in Data Science at the International
Hellenic University.
Machine learning is getting more and more developed in the industrial field to improve
its performance. During the last years, machine learning was used to address the problems that arose in a time series classification process. The main problem identified was
the management of large volumes of data at any given time. Deep learning is a part of
the machine learning methods that have been very successful in many fields of pattern
recognition. Recently, researchers have succeeded remarkable results in time series
classification using Deep Learning techniques.
The goal of the thesis is to present a fault diagnosis and anomaly detection in the
benchmark Tennessee Eastman process by using Convolutional Neural Network techniques. The Inception Module and Residual Network are developed methods of the
conventional Convolutional Neural Network that use fewer parameters but maintain the
accuracy of the model. Furthermore, the Statistical method of Principal Components
Analysis was used in the context of preprocessing to reduce the dimensionality of the
data.
These methods are going to be tested for their performance but the accuracy metric is
not sufficient to evaluate the models. For this reason, Fault Detection Rate and Fault
Positive Rate are the metrics that determine the performance of the models.
Collections
Show Collections