Modified Independent Component Regression Method Without Prewhitening

Abstract

Independent component analysis (ICA) can reveal the essential underlying structure of data, and independent component regression (ICR) methods usually obtain better performance than other regression methods such as principal component regression. However, when existing ICR methods separate or extract independent components using prewhitened data, the backward propagation of inevitable prewhitened errors deteriorates the final linear prediction accuracy. To overcome this weakness, first, we proposed using weighted orthogonal constraint condition to replace the prewhitening of the data in ICA. Next, the statistical independence of ICs and the close relationship between ICs and quality variables are considered at the same time. Then, by combining the merits of improved ICR and ensemble ICR algorithm which solved the problem of selecting an appropriate nonquadratic function in ICA iteration procedure, a modified independent component regression (MICR) method that directly used the measured process data was proposed. Finally, three experimental results were used to validate excellent performance of modified algorithm.