Variational Adversarial Active Learning Assisted Process Soft Sensor Method

Abstract

Soft sensor methods have been widely applied in process industries to predict key quality variables that cannot be measured online. However, labeled samples to construct models are often limited because quality variables are difficult to be obtained. Additionally, due to the instrument of redundant sensors, the process data is high-dimensional with strong correlations. In this paper, an active learning soft sensor framework named variational adversarial active learning (VAAL) is developed to select informative unlabeled samples to enhance prediction performance. The sampling strategy of VAAL learns a latent space using a variational autoencoder (VAE) and an adversarial network trained in a way of minimax game. The VAE tries to trick the adversarial network into predicting that all samples are from the labeled pool, while the adversarial network learns how to discriminate between dissimilarities in the latent space. The Gaussian process regression model is adopted in VAAL as a base soft sensor. The prediction results of an industrial debutanizer column demonstrate the advantages of VAAL as compared to the existing active learning strategies.