Large Data Soft-Sensor Based on Mixture of Gaussian Process Regression Model with Lasso Criterion

Soft sensors are used to enhance process monitoring and control of difficult to measure process variables. Moreover, modern industrial processes exhibit multi-mode behaviour due to different demands from the market. As a result, the usual assumption about uni-modal Gaussian distribution of the data is not valid. One approach to resolve this is by clustering the input data using Gaussian mixture model. However, like other clustering algorithms, experience severe performance degradation in the high-dimensional setting. To overcome this shrinkage methods have been shown to yield more stable and interpretable models. The application of least absolute shrinkage and selection operator (LASSO) criterion to mixture of Gaussian process regression model (MGPR) is proposed in this work to handle multi-mode nonlinear process with high dimension of variable. It enables simultaneous identification of significant variables and determination of important clusters. The clustering penalty penalizes the distance of the regression coefficients between clusters and thus similar data are combined into common cluster, while the regression penalty selects significant regression variables within each cluster. Case studies presented show the capability of the proposed method and its applicability to an industrial melt index process.