Fault Detection in Tennessee Eastman Process

Recently, the advent of a new simulated Tennessee Eastman Process (TEP) dataset made it possible to use deep learning techniques for fault detection in chemical processes, which had previously been unavailable. In this thesis we apply a set of models, which are typically associated with time-series data, such as recurrent neural networks, attention and transformer mechanisms, or computer vision, as convolutional neural networks. Another deep learning architecture that could be applied to this problem is a generative adversarial network, which can enrich training data for better performance and detect faults independently. Given sensory data is a long time series that could lead to typical problems of applying recurrent networks, therefore we propose two CNN architectures to solve them. We compare all models in terms of true positive rate (TPR), false-positive rate (FPR), and detection delay, the most crucial metrics in fault detection studies. And proposed models over-perform most previous studies. The dataset we used is publicly available. Key words: industrial machine learning, fault detection, tennessee eastman process, chemical processes, deep learning, generative adversarial networks