Latent-Enhanced Variational Adversarial Active Learning Assisted Soft Sensor

As the acquisition of variables that measure quality is typically challenging, labeled samples for building a model for soft sensors are often inadequate. Additionally, owing to the installation of redundant sensors, high-dimensional process data with strong correlations are acquired. Therefore, in this study, a sample selection strategy for active learning (AL), referred to as latent-enhanced variational adversarial AL (LVAAL), is developed to enhance quality prediction performance with limited labeled data. The LVAAL method uses a minimax game to explore the latent representation of the original data. Specifically, a latent-enhanced variational autoencoder (LEVAE) is used to deceive the adversarial network into predicting all samples, both labeled and unlabeled, as labeled. However, the adversarial network attempts to indicate the difference between latent representations. Subsequently, Gaussian process regression (GPR) is adopted as a base model for LVAAL. With the proposed LVAAL selection strategy, informative unlabeled samples are incrementally selected, thereby enhancing the prediction performance. The application results for a numerical case and an industrial example demonstrate the advantages of LVAAL compared with existing AL strategies.