Semi-supervised generative adversarial network with guaranteed safeness for industrial quality prediction

Abstract

In process industries, due to the low sampling rate of the quality variables, there are abundant unlabeled data but limited labeled data. Most data-driven quality prediction models only use the labeled data but ignore the unlabeled data, resulting in overfitting and low generalization performance. Hence, it is necessary to extract useful information from the unlabeled process data. Due to the noise in the signal transmission process and sensors, there are unlabeled data with low confidence that will mislead the training process. To tackle this problem, we propose a semi-supervised Generative adversarial network with Co-trained Generators (GCG) that utilizes the unlabeled data safely through the co-training of generators. The optimal parameters and weight coefficients of co-trained generators guarantee the ”safeness”, i.e., the performance of GCG is not worse than generators with only labeled data. The proposed method is validated by a benchmarked industrial case and the real absorption-stabilization system in Fluid Catalytic Cracking Unit (FCCU). The results suggest that the GCG method improves the generalization performance of the quality prediction model.