Regularized error-in-variable estimation for big data modeling and process analytics

Abstract

This article addresses estimating the uncertainty in operational data by introducing a regularized modeling technique. Existing work (i) requires knowing the true dimension of the operational data, (ii) relies on a maximum likelihood estimation that is compromised by a stringent restriction for this true dimension and (iii) is computationally expensive. In contrast, the presented regularized error-in-variable technique (i) allows determining the true data dimension through hypothesis testing, (ii) is not limited by the restriction of existing methods, and (iii) has an objective function that can be solved efficiently. Based on a simulation example and the analysis of two industrial datasets, the paper highlights that the regularized estimation technique outperforms existing work and shows how to embed this technique within an advanced process analytics framework for advanced process control, optimization and general process diagnostics.