jaue2021-079: The fault detection and diagnosis (FDD) for deep belief network (DBN) - based AHUs using data feature selection

Abstract

The fault detection and diagnosis (FDD) for air handling units (AHUs) often impact greatly on building energy
efficiency and indoor environmental quality. To better deal with the complexity of FDD for AHU, a deep learning
data-driven technology is indispensable. In this paper, a deep belief network (DBN) - based AHU FDD model is
developed. The model training and testing are implemented using the preprocessed feature set from the ASHRAERP-
1312 summer experimental data. The FDD accuracy of up to 99.89% shows the high effectiveness of the given
DNB. Then, to reduce the number of features while still maintain their characteristics, two data feature selection
methods, Max-relevance and min-redundancy (MRMR) and Maximal information coefficient (MIC) are employed
respectively to provide feature subsets for the DBN model. The results indicate that both MIC and MRMR have
excellent capabilities for attribute filtering, yet MRMR is more prominent for considering the redundancies among
features. In addition, compared with the traditional SVM, the proposed DBN model performs better for the higher
accuracy, especially for the fast calculation speed.