Online quality estimation in chemical processes with random subspace deep partial least squares model

While partial least squares (PLS) has been widely used for industrial quality estimation along past decades, it was recently extended to the deep learning form, which has obtained significant achievements in both prediction performance and overcoming several critical issues of traditional deep learning methods, such as high computational burden, reliance on large-scale training data, and too many tuning parameters. In this paper, the deep PLS model is further extended to the ensemble learning form, with the strategy of random subspace. Due to the lightweight nature of deep PLS, it is of high efficient to construct the random subspace induced ensemble deep PLS model for online quality prediction. A two-scale feature ensemble strategy is formulated to enhance the performance of random subspace deep PLS model, which are termed as inside-model-ensemble and betweenmodel-ensemble. In addition, the robustness of the proposed model to variable selection has been greatly improved by the random subspace ensemble strategy, which it is critical to both PLS and deep PLS models. A real industrial example is provided for detailed quality prediction performance evaluation.