Novel Stacked Maximal Information Coefficient-Weighted Autoencoder With Attribute Expansion and Its Application to Industrial Soft Sensors

In recent years, data-driven soft sensors, especially deep learning soft sensors show great potential for application in the process industry. As a typical deep network, stacked autoencoder (SAE) has an outstanding modeling capability in soft sensors due to its ability to extract deep features. However, SAE ignores the expanded representation of the input data and uses an unsupervised approach in the pre-training phase, which may have resulted in the extraction of some features irrelevant to the key output variables. In this paper, a novel stacked maximal information coefficient-weighted autoencoder with attribute expansion (SMWAE) is proposed to establish a soft sensor. Firstly, a novel input attribute expansion strategy is designed based on the function of smooth trend and periodic trend to enhance the expression of input variables. Additionally, considering the correlation between latent features and outputs, the maximum information coefficient (MIC) is integrated into the autoencoder (AE) loss function to further refine the features to improve the sensitivity of the model to key output variables. Then, each feature layer of the AE is stacked to form a depth structure to fully explore the deep information in the data, thus effectively improving the prediction accuracy of the soft sensor model. Finally, the proposed SMWAE model is utilized to predict key variables in the gas turbine process and the purified terephthalic acid (PTA) solvent process, and the results of the comparative experiments demonstrate that SMWAE successfully achieves excellent prediction performance.