Robust soft sensor development based on Dirichlet process mixture of regression model for multimode processes

Abstract

Industrial processes often exhibit multimode characteristics due to factors like load variations, equipment changes, and feedstock fluctuations. This paper introduces a Dirichlet Process-based Twofold-Robust Mixture Regression Model (DPR²MRM) for multimode processes. As a Bayesian nonparametric model, it automatically determines the number of mixture components from observed data using Dirichlet process mixture techniques, avoiding underfitting and overfitting. The model employs a Student's-t mixture model for input space learning, leveraging its long-tail properties for robust mode identification. For each mode, a regression model is built to capture the relationship between inputs and outputs, incorporating Student's-t noise to ensure robustness against output space outliers. The optimal posteriors of the model parameters are inferenced within a full Bayesian framework, and an analytical posterior predictive distribution is derived. The effectiveness of the DPR²MRM is demonstrated through a numerical example and two industrial applications.