Partial Least Squares Method Based on Double Mutual Information and its Application

Abstract

As we all know, PCA algorithm provides a linear transformation matrix between high and low dimensions, which is processed with dimensionality reduction by means of projection, so as to intercept the main information useful to people. However, the PCA algorithm does not take into account that its residuals may also contain process variables closely related to quality. Therefore, no matter what is done with the extracted ingredients, it is possible to lose important information. PLS considers the deficiencies of PCA and is widely used in quality control systems. However, since there will be non-gaussian interference in the system, the covariance method of the original PLS does not adapt to system interference. Therefore, we propose a PLS method based on double mutual information (DMIPLS) to solve the above problems. This method first use mutual information to calculate the weighted mutual information between process variables X and quality variables Y, select a set of process variables X, which are closely related to the quality variables, to ensure that the following PLS have valid process variables. Then, in the second to calculate the mutual information between residual matrix, to ensure that the residual error matrix of the process variables related to quality is not missing. The problem of non-gaussian interference is solved by using the feature of mutual information. Finally, the effectiveness of the method is verified by the simulation of TE system.