Information Complementary Fusion Stacked Autoencoders for Soft Sensor Applications in Multimode Industrial Processes

Abstract

Multimode characteristics within process data are common due to switching operating conditions or load changes, which pose challenges for soft sensor modeling. The key challenges include multiple data patterns and complex nonlinear relationships between process variables and quality variables. To address these issues, an information complementary fusion stacked autoencoders (ICF-SAE) model is proposed in this article. In ICF-SAE, a gating module is designed. Since the gating module can regulate the information flowing path, the proposed model can represent multiple patterns in the data. Moreover, the bottom hidden representations of SAE contain a lot of noise. Using them directly to predict quality variables may deteriorate the prediction accuracy of the model. Taking this issue into account, a layerwise fusion module including bottom-top and top-bottom fusion channels is designed. This module can filter noise progressively by sequential layer fusions. Detailed evaluations of the proposed ICF-SAE model are conducted using numerical data and real-world industrial data. The results show that the proposed ICF-SAE model achieves better prediction accuracy than the existing methods.