Improvement of virtual in-situ calibration in air handling unit using data preprocessing based on Gaussian mixture model

Abstract

Fault diagnosis of sensors used in air conditioning systems can significantly reduce the energy consumption and improve the indoor comfort. Previously, a virtual in-situ calibration method based on the Bayesian inference and Monte Carlo method was proposed to diagnose sensor faults without installing redundant sensors. However, air conditioning systems operate in continuously changing operating conditions because of the changes in the outdoor environment and indoor load. This dynamic nature of the operating conditions poses difficulty in the selection of the existing historical datasets and results in uncertainty in calibration results. To address this problem, this study proposes Gaussian mixture model (GMM) to preprocess historical data to obtain steady-state measurements under various operating conditions. A variational Bayes expectation maximization (VBEM) algorithm was applied to the GMM to increase its clustering accuracy and remove human intervention. Next, the expectation maximization (EM) algorithm was used to solve the GMM in order to illustrate that unsuccessful clustering will lead to fluctuation of calibration results and verify the accuracy of VBEM. In this case, the number of GMMs was determined manually based on attempt step by step. Finally, the clustering results of the two algorithms were incorporated into a virtual in-situ calibration method. Based on comparison of calibration results, it can be concluded that the calibration results deviated significantly from the true value when the number of GMMs was unreasonable after the EM algorithm. However, the accuracy of all sensors was improved by 75% after calibration, and convergence occurred when the datasets were two to four times the number of sensors after the VBEM algorithm.