Deep Learning with Local Spatiotemporal Structure Preserving for Soft Sensor Development of Complex Industrial Processes

Abstract

Recently, deep learning has attracted increasing attention for soft sensor applications. However, traditional deep learning methods capture the hierarchical data features by minimizing the global fitting errors, neglecting the local structure characteristics implied in the original data. In this paper, we propose a deep learning approach for soft sensor development. Using the autoencoder as the basic structure of the network, the layer-wise pretraining optimization goal is integrated with data neighborhood feature learning and a supervised tuning strategy, and a local spatiotemporal structure-preserving stacked semi-supervised autoencoder (LSP-SuAE) is established. Specifically, representative data features are extracted by learning the global and local characteristics of data. A layer-wise semi-supervised learning mode is adopted to improve the correlation between data features and quality variables, thereby obtaining optimal network parameters and improving the accuracy of quality prediction. Experimental results showed that LSP-SuAE had higher accuracy and better stability than five data-driven methods.

Keywords: soft sensor, Data-driven modeling, Deep Learning