Bayesian neural networks for nonlinear multivariate manufacturing process monitoring

Abstract

As a linear method, PCA is not accurate for complicated processes control when nonlinear correlations are involved in the multivariate measurement variables. As one appealing nonlinear PCA method, principal curves generalized PCA to nonlinear domain and provide a better way to nonlinear feature extraction and dimension reduction. A multivariate process monitoring method based on Bayesian neural networks is proposed in this paper, which involves a projection network and a reconstruction network to represent the nonlinearities and helps avoid the overfitting problem in the weight parameter learning. Experimental study has illustrated the potential applicability of this method for nonlinear feature extraction and multivariate process monitoring.