Layer-wise feature extraction approaches with deep PLS for quality prediction in industrial process

Partial least squares (PLS) has been widely applied for quality prediction in industrial processes. However, PLS can only extract one-layer linear quality-relevant features for regression tasks. Besides, shallow PLS suffers from information loss in its residual subspace. To alleviate these problems, a deep PLS (DPLS) framework and its enhanced version are proposed in this paper. DPLS consists of multi-layer PLS and uses the extracted features as layer connection. To improve model nonlinearity, nonlinear functions are introduced between two adjunct layers. With layer-wise nonlinear mappings and PLS, more high-level and quality-related features can be mined and utilized for soft sensor modeling. On this basis, an enhanced DPLS (EDPLS) method is further developed to make full use of information in residual PLS by considering it in the next layer. Finally, the effectiveness of the proposed methods is validated on an industrial hydrocracking process.