Ensemble learning with member optimization for fault diagnosis of a building energy system

For better service and energy savings, improved fault detection and diagnosis (FDD) of building energy systems is of great importance. To achieve this aim, ensemble learning is investigated and introduced in this study. Three types of methods like K-nearest neighbor (KNN), support vector machine (SVM), and random forest (RF) are carefully selected, optimized, and integrated into an ensemble diagnostic model (EDM) by the majority voting method. Experimental data for seven typical gradual faults in a centrifugal building chiller are used for model validation and evaluation. The results show that the diagnostic accuracy of the EDM (99.88%) is higher than that of the individual methods, with significant improvements for normal operation and refrigerant leakage, and no false alarms reported. Models based on ensemble learning, EDM and RF (homogenous ensemble), exhibit better performance for global faults, which are difficult to diagnose. In addition, five different feature sets are selected from the literature for further tests. It is found that the diagnostic performance depends not only on the principle of diagnosis, but also on the fault category and the characteristics of the feature set such as the indicative degree to corresponding faults, number of features, correlation degree between features, and information redundancy, and feature selection is proved to be more important than algorithm selection in fault diagnosis practice. Ensemble learning is proved to be a promising candidate for the fault diagnosis of building energy systems, except for the Zhou-8 feature set (eight temperature features), for which KNN is a better choice. (C) 2020 Elsevier B.V. All rights reserved.