Robust probabilistic principal component regression with switching mixture Gaussian noise for soft sensing

Abstract

In the era, that data collection is not as challenging as before, data-driven process modeling for prediction of unmeasurable or expensive-to-measure variables is gaining popularity. Probabilistic principal component analysis has powerful features for modeling such as considering uncertainty and dealing with high-dimensional process data. Although data collection is more attainable these days, low quality of data still diminishes model performance. High-fidelity modeling requires high-quality data. The focus of this work is to deal with outlying observations by developing a Robust Probabilistic Principal Component Regression (RPPCR). Here, we have investigated a scenario of mixture Gaussian switching measurement noise to mimic certain type of outliers in a forward-looking approach that extends our previous work. A rigorous modeling approach that can handle switching noise and the solution methodology are discussed in detail. Two case studies, a numerical illustrative example and a real industrial counterpart, are considered to verify the robustness of proposed model.