Stacked Dual-Guided Autoencoder: A Scalable Deep Latent Variable Model for Semi-Supervised Industrial Soft Sensing

Stacked autoencoders (SAEs) have been widely used in soft sensing of industrial process data. However, most of them extract features by minimizing the reconstruction error of the input data at each layer, leading to accumulated loss. Moreover, the extraction of quality-related features is crucial for soft sensing. To address these issues, this article first introduces a novel stacked dual-guided autoencoder (SDGAE) to better characterize intricate data patterns and learn quality-associated features. Compared to SAE, the idea of SDGAE lies in stacking a series of hierarchical dual-guided autoencoder (DGAE), aiming to allow each DGAE to accurately reconstruct the original input data. It also enables the model to simultaneously extract features that are highly correlated with the output variable by introducing a supervisor term. Furthermore, practical scenarios typically involve limited labeled data and a substantial volume of unlabeled data, which renders the network deficient in adequate generalization capabilities. To address this predicament, we further propose a deep semi-supervised SDGAE (SSDGAE), it is able to obtain useful feature representations and structural information from unlabelled data in the pretraining phase in addition to the labeled data as SDGAE is able to do. Finally, the proposed methods are verified on two real industrial cases for its effectiveness and superiority. The results indicate that our approach can simultaneously utilize both labeled and unlabeled data to enhance the predictive performance of quality variables.