Entropy-clustering and K-means based kernel partial least squares soft-sensing method

Kernel partial least squares(KPLS) is widely adopted for soft-sensing in nonlinear industrial process. For KPLS method, the determination of central nodes and kernel width in the kernel function will affects generalization ability and predictiability. This paper proposes an entropy-clustering and K-means based KPLS regression method. First of all, it divides the original data into several clusters by entropy clustering method and obtains the initial clustering centers. Secondly, K-means algorithm is applied on these initial clustering centers to get the final central nodes of the kernel functions. Finally, the width of kernel function is determined according to the Euclidean distance between the central node of the kernel function and its adjacent central node. The proposed method is verified based on the process data from a chemical enterprise. The experiment results show that the proposed algorithm greatly reduces the measurement error compared to the traditional KPLS regression method.