In-situ sensor calibration in an operational air-handling unit coupling autoencoder and Bayesian inference

Abstract

Sensor errors have a considerable influence on the system operation and energy usage in an air handling unit. Sensor fault detection and diagnosis (SFDD) has been widely studied to handle the impacts of sensor errors on an air-handling unit (AHU). Beyond the SFDD, in-situ calibration can correct the faulty sensor (especially for systematic errors) automatically in field, thereby reducing the energy waste. In this study, we propose an advanced in-situ sensor calibration named virtual in-situ calibration in an operational AHU. The suggested method is intended to overcome the challenges of previous in-situ calibration methods by coupling the Bayesian inference and autoencoder. In a given sensor calibration domain of an AHU, based on the unsupervised learning neural network feature trained to duplicate their input variables, the autoencoder-coupled calibration can produce system and sensor models effectively without additional sensors and assumptions, which are the main limitations of the earlier methods. It improves the calibration performance and applicability in the AHU. In addition, a three-step strategy to construct autoencoder input variables and a new distance function to achieve successful calibration under various faulty conditions is proposed. In a case study, where the error in the cooling coil supply temperature (+2 °C) caused a total energy increase of 38%, the present method is shown to eliminate the sensor error and the energy waste completely. These results show the capabilities and potentials of the suggested method in the self-repair, diagnostics, and automation of a building energy sector.