Enhancing Multirate Soft Sensing via Progressive Transfer Learning and Generative Adversarial Imputation Networks

In industrial processes, data are often collected at varying sampling rates due to factors such as heterogeneous variables and diverse sensor configurations. Notably, quality-related variables typically exhibit even lower sampling rates, posing significant challenges for constructing accurate soft sensor models. In recent years, generative adversarial imputation networks have shown potential in addressing data incompleteness. However, the substantial amount of missing parts in low-sampling-rate variables hinders effective model training, and the imputed values often lack reliability. To address these limitations, this paper proposes a novel soft sensor framework named Progressive Transfer GAIN coupled with Semi-Supervised Variational Autoencoder (PT-GAIN-SSVAE). Firstly, PT-GAIN employs progressive transfer learning strategy across multirate variables to pretrain the generator, which fully leverages the data information contained in the regularly missing data. Secondly, the PT-GAIN model is trained in an adversarial manner and subsequently imputes the unsampled parts across different process variables. Thirdly, the imputed data is used to train a semi-supervised variational autoencoder. Finally, the effectiveness of the proposed method is validated by a real ammonia synthesis process.