A Supervised Stacked Dual-Guided Autoencoder with Deep Learning Networks for Industrial Soft Sensing

Stacked autoencoder (SAE) has been applied to the prediction of manufacturing process data quality owing to the excellent feature selection capabilities. Nevertheless, a challenge arises in the accumulation of losses from the lowest to the highest level when minimizing reconstruction losses at each layer. Additionally, many SAE models struggle to extract sophisticated features linked to quality crucial for output predictions. This article delves into a supervised stacked dual-guided autoencoder (SDAE) proposed to describe complex data patterns and learn quality-related features better. It reconstructs the same original input data for each dual-guided autoencoder (DAE) and measures the variables at the input layer related to quality. A new pretraining objective function aims to supervise the network's pre-training process, ensureing that the extracted features include more quality prediction information. Lastly, the efficiency of the introduced SDAE is validated on the CO
2
absorption column process.