Variance correction principal component analysis—gross error detection with serially correlated data

Abstract

Abstract Process industries have complex measurement systems set up to measure process variables for control, dynamic optimization, online diagnostics, and real-time monitoring. Errors in measured process data are typically classified as random or fixed (gross) errors. Data reconciliation (DR) addresses random errors, whereas gross error detection and reconstruction (GEDR) addresses gross errors. In most GEDR techniques, data collected from sources are considered independently and identically distributed (i.i.d.). Most GEDR approaches consider data acquired from sources to be dispersed independently and identically (i.i.d.). Data acquired from multiple sources do not have to be independent in industrial manufacturing practice. They may have serial correlation due to control loops, process dynamics, feedback networks, etc. This paper proposes a new technique, variance correction (VC) principal component analysis (VCPCA) based measurement test, to identify gross errors in serially correlated data. The proposed technique uses a VC approach to estimate the variance of serially correlated data and a PCA-based estimator to calculate the residuals. The advantages of the proposed technique are assessed by comparing its performance with other existing methods like VC and pre-whitening approaches. The results demonstrated the superior performance of VCPCA and produced 99% success in all simulation trials in identifying gross errors and only 1% false identification.