Quality-Related Fault Detection for Processes with Time-Varying Measurements.

Bahador Rashidi,Qing Zhao
DOI: https://doi.org/10.23919/ecc.2019.8796220
2019-01-01
Abstract:Partial least-squares techniques have been widely used for quality-related fault detection purposes when there is a linear relation between process variables and quality-outputs. However, those methods are suitable when the process variables and quality-outputs have constant mean values. In this paper, a novel least-squares-based scheme is proposed for quality-related fault detection of the dynamic linear processes, in which the measured variables and quality-outputs may have time-variant mean values. The proposed method is built upon a cascade modeling framework including complete orthogonal projection of the process variables onto quality-related and quality-unrelated subspaces and finding the principal manifolds for the projected components which represent their underlying auto-regressive moving average models. Moreover, a new residual index is proposed, which is insensitive to the mean changes of the process variables. The proposed method and dynamic improved partial least-squares (DIPLS [1]) technique are finally applied to a numerical case study and a non-isothermal continuous stirred tanks reactor (CSTR) benchmark, and the simulation results demonstrate the effectiveness of the proposed scheme.
What problem does this paper attempt to address?