2-D regularized locality preserving projection algorithms for temporospatial feature reduction and its application in industrial data regression

In design of dynamic soft sensors for measuring product quality in complex industrial processes, the input variables of a regression model are composed of temporospatial data. Information redundancies caused by correlation nature among process variables and data samples may result in a poor generalization performance of the soft sensor. Therefore, it is useful to develop effective feature extraction techniques for soft sensor design. This paper proposes a two-dimensional regularized locality preserving projection (2DRLPP) algorithm for feature extraction, which combines locality preserving projection (LPP) method with data roughness regularization. An extension of our proposed 2DRLPP is given and termed as bidirectional 2DRLPP, denoted by (2D)2RLPP. A case study on soft sensor design for prediction of the cement raw material decomposition rate is carried out to illustrate the effectiveness of the proposed feature extraction techniques in this paper. Experimental results demonstrate that the proposed algorithm improves the prediction performance of ε-SVR-based soft sensors.