Deep Learning for Classification of Profit-Based Operating Regions in Industrial Processes

Abstract

A classification approach is proposed for finding ranges of process inputs that result in corresponding ranges of a process profit function using deep learning. Two deep learning tools are used to formulate models for use in classification, based on either supervised learning or unsupervised learning approaches. The supervised learning models are based on long short-term memory networks and multilayer perceptron networks while the unsupervised learning model consists of an autoencoder neural network connected to a support vector machine classifier. An algorithm referred to as sequential layer-wise relevance propagation for pruning (SLRPFP) is proposed and applied to the aforementioned models for selecting relevant inputs and for pruning the neural networks and its inputs such that the test accuracy at every step of the proposed sequential algorithm is maintained or even improved. It is also shown that the selected inputs from the proposed algorithm (SLRPFP) provide important process insights on the productivity, that is, the profit-based objective function. The approaches are illustrated for the Tennessee Eastman Process (TEP) and for an industrial vaccine manufacturing process (industrial process). The efficacy of the proposed supervised and unsupervised deep learning approaches over linear model-based classification methods that are based on linear dynamic principal component analysis combined with multiclass support vector machines classification is shown by comparing the performances of both a TEP and a vaccine manufacturing process.