Data Preprocessing for Soft Sensor Using Generative Adversarial Networks

Abstract

Soft sensors have been widely used in industrial processes in the past two decades, using easy-to-measure process variables to predict hard-to-measure ones. Sufficient training data can significantly improve the prediction performance as well as speed up the computation with low cost in the modeling process of the soft sensor. However, data collection is often difficult due to the harsh environment of the industrial process. Generative adversarial networks (GANs) is a prominent method for learning generative models in recent years. In this paper, a generative model named DWGAN based on improved Wasserstein generative adversarial networks (WGAN) is proposed to generate new samples for soft sensors. For evaluating the proposed method, a practical industrial experiment is studied in the case of paucity of data, and three other state-of-the-art generative models are adopted for comparison. The experimental results show that the samples yielded by the proposed method behave more similarly with the real samples compared to samples provided by other methods. Additionally, these synthetic samples can make the training data sufficient and significantly improve the prediction accuracy of the soft sensors.