Enhanced Soft Sensor with Qualified Augmented Data Using Centroid Measurement Criterion

Abstract

Development of reliable soft sensors using limited labeled samples is not an easy task in industrial processes. A selective generative adversarial network (SGAN)-based support vector regression (SGAN-SVR) soft sensor is proposed for quality prediction using limited labeled training data. Specifically, SVR is considered as a base prediction model. The Wasserstein GAN (WGAN) is adopted to capture the distribution of available labeled data and generate virtual candidates. Subsequently, using a proposed similarity measurement strategy, those synthetic data with more information are selected and introduced into the training set. Using the designed data augmentation approach, the SGAN-SVR model can achieve better prediction performance compared with the SVR soft sensor. The quality prediction results on an industrial polyethylene process demonstrate the effectiveness and advantages of the proposed method.