Constrained Multi-Output Gaussian Process Regression for Data Reconciliation

Abstract

Reliable process data are required for process monitoring, and sensors in the chemical process industries are often exposed to harsh conditions resulting in poor data quality. Data reconciliation methods are employed to improve sensor measurements and remove gross errors. This work proposes a Bayesian approach to the steady-state data reconciliation problem via the development of a novel constrained multi-output Gaussian process regression model which incorporates high-confidence process models based on steady-state conservation laws. The performance of this model is compared with steady-state data reconciliation and single-output Gaussian process regression in a simulated case study. Simulation results show an improvement upon traditional steady-state data reconciliation methods during normal operation while performance deteriorated when large process perturbations were introduced. The model shows promise for future work in process monitoring applications that use Gaussian process regression.