Nonlinear Soft Sensor Modeling Method Based on Multimode Kernel Partial Least Squares Assisted by Improved KFCM Clustering

Abstract

Kernel partial least squares (KPLS) method has gained successful applications in the field of nonlinear soft sensor modeling. However, a single global KPLS model may perform unsatisfactorily in some complicated processes, where there exist multiple nonlinear relationships. To handle this problem, this paper proposes a multimode KPLS based soft sensor modeling method assisted by the improved kernel fuzzy C-means (]CFCM) clustering. The proposed MKPLS method applies the “divide and rule” strategy, which partitions the training data into many clusters and builds the local KPLS model for each cluster. Different to the traditional KFCM clustering method, which divides the process data based on the spatial position similarity, this paper designs an improved KFCM method by concentrating on the functional relationships of the samples. Based on the improved KFCM clustering method, the data with the same nonlinear relationships are clustered together and the corresponding KPLS model is developed. Two case studies including one numerical system and one continuous stirred tank reactor (CSTR) system are used to validate the proposed method, and the results demonstrates the effectiveness of the proposed method.