Feature selection based on Null Importance for Fault Diagnosis of Chiller

Fault detection and diagnosis of chillers have a significant impact on equipment lifetime, indoor comfort, energy efficiency, and performance of building management systems. The feature selection of chiller variables is the most important step in the process of FDD of chillers, and the selection of important features not only reduces the variable dimension and the use of field sensors, but also improves the detection accuracy. In this paper, we apply the Null Importance method based on the LightGBM(Light Gradient Boosting Machine) model to the feature selection of chiller variables and use ASHRAE project 1043-RP for experimental validation. This set of features and all features are then used to train common models for fault detection and diagnosis separately to verify the validity of feature selection. These common models include Support Vector Machine, Decision Tree, Random Forest, and K-Nearest Neighbor. The results of the experiments show that the Null Importance feature selection method based on the LightGBM model is effective for chiller unit fault diagnosis, and the selected 16 features can achieve better fault diagnosis and detection results than using all features on the Support Vector Machine, Decision Tree, and K-Nearest Neighbor.