Improving the Performance of Just-In-Time Learning-Based Soft Sensor Through Data Augmentation

Abstract

Just-in-time learning (JITL) is a widely used method for online soft sensing. The limitation of available data and the increase of sample dimensions will make the historical dataset sparse, seriously impair the reliability and usability of JITL applications in the industry. To this end, data augmentation technology (DA) is introduced into the JITL framework for the first time to improve the performance of JITL-based soft sensors. In this article, a novel causality-informed variational autoencoder (CIVAE) is developed to generate virtual samples to augment the historical dataset. Based on this, a complete framework of data augmentation just-in-time learning (DA-JITL) is formulated offline and online and implemented with two different strategies. Finally, a numeral example and an actual industrial example are applied to verify the effectiveness of the proposed method. Besides, the effect of virtual data on JITL and the influence of virtual data volume, virtual data ratio and other factors are carefully discussed on the proposed method. In the industrial case, two indicators (rmse and R2) of the proposed method have been improved by an average of 25% and 17%, respectively, compared to the traditional JITL approach.