Classification of Profit-Based Operating Regions for the Tennessee Eastman Process Using Deep Learning Methods

Abstract

The focus of this work is on the classification of an input-design space into different regions of input conditions that result in different corresponding ranges of productivity costs in the Tennessee Eastman Process (TEP). Although similar classification tasks had been previously carried out using linear multivariate statistical data analysis methods, these are of limited efficacy when dealing with highly non-linear dynamics. In this work, we present two Deep Learning Tools for classification: using either supervised learning or unsupervised learning. For classification with supervised learning, a Recurrent Neural Network (RNN) known as Long Short-Term Memory (LSTM) is trained on normalized training data. Since deep learning networks generally involve a large number of nodes and parameters an algorithm named Sequential Layer-wise Relevance Propagation (SLRPFP) is proposed for selecting the relevant inputs and for pruning the LSTM network such that the test accuracy at each step is maintained or even improved. For classification with unsupervised learning, the main features from the input dataset are first extracted using an Autoencoder. Then a Multi-dimensional Support Vector Machines (MSVM) model is applied to the features identified by the autoencoder. The performance of the proposed supervised and unsupervised deep learning approaches are compared to an approach that combines linear Dynamic Principal Component Analysis (DPCA) and a MSVM based classification and conclusions are drawn on the relative advantages of the deep learning methods.