A soft sensor model based on an improved semi-supervised stacked autoencoder for just-in-time updating of cement clinker production process data f-CaO

Abstract

Abstract The free calcium oxide (f-CaO) content in cement clinker serves as a critical quality indicator for cement production. However, many soft sensor models employed for predicting f-CaO content utilize a limited amount of labeled data, leading to the underutilization of a substantial volume of unlabeled data and its associated information. To tackle these challenges, this study introduces soft sensor methodology based on improved semi-supervised attention stacked autoencoders (ASS-SAE). We propose an enhanced confidence-generating pseudo-labeling technique to identify high-confidence pseudo-labeled samples from pseudo-labels within a subset of correlated samples, addressing the issue of inadequate labeled data. To fully utilize the information hidden in the unlabeled data, the proposed method incorporating the confidence attention mechanism then assigns weights to the high-confidence pseudo-labeled data and inputs them into the SAE along with labeled data from a subset of similar samples for re-fine-tuning. By conducting an illustrative analysis using authentic cement data proposed for this study, the effectiveness of the approaches employed in this research is substantiated.