Integrating dynamic slow feature analysis with neural networks for enhancing soft sensor performance

Abstract

This paper proposes integrating slow feature analysis (SFA) with neural networks (SFA-NN) for soft sensor development. Dynamic linear SFA is applied to the easy to measure process variable data. Then the dominant slow features are selected as the inputs of a neural network to predict the difficult to measure product quality variables. SFA can capture underlying dynamics of industrial processes through the extraction of slowly varying latent variables, known as slow features. Selection of dominant slow features using scree plot is proposed. Neural networks are utilised to cope with nonlinearities present in many real industrial processes. The effectiveness of the proposed method is evaluated on two real industrial processes and is compared with slow feature regression, partial least square regression, traditional feedforward neural networks, and using principal component analysis prior to a neural network. The proposed SFA-NN gives the best generalisation performance amongst these techniques in both case studies.