Ensemble independent component regression models and soft sensing application

• Three ensemble forms of the independent component regression model are developed.
• Soft sensors are built based on different ensemble regression models.
• Detailed comparative studies of different soft sensors are carried out.
• Both the robustness and the stability of the regression model are improved.

Independent component regression (ICR) model is able to extract underlying components, while simultaneously models high order statistics from the non-Gaussian process data. Based on different ensemble strategies, this paper aims to develop various ensemble forms of the ICR model. Specifically, by re-sampling of data samples, a bagging ICR model is developed; based on the independent component decomposition, a subspace ICR model is constructed through each direction of independent components; a further bi-dimensional ensemble ICR model is then constructed by combining bagging and subspace ICR models through two ensemble directions. For online measurements of key variables in industrial processes, various soft sensors are built based on different ensemble ICR models. Performance evaluations and detailed comparative studies of the developed soft sensors are illustrated through an industrial process.