An enhanced principal component analysis method with Savitzky–Golay filter and clustering algorithm for sensor fault detection and diagnosis

Abstract

Sensors are critical components of heating, ventilation, and air-conditioning systems. Sensor faults can impact control regulations, resulting in an uncomfortable indoor environment and energy wastage. To detect and identify sensor faults quickly, this study proposes an enhanced principal component analysis (PCA) method using the Savitzky–Golay (SG) filter and density-based spatial clustering of applications with noise (DBSCAN) algorithm. First, the DBSCAN algorithm is used to automatically divide the dataset into sub-datasets with different working conditions to reduce the interference information and concentrate the information of each training set. Then, each sub-dataset is smoothed using the SG algorithm to reduce the effects of data fluctuations. The processed dataset is used to build a sub-PCA model that ultimately identifies and locates faults. The proposed strategy is validated using field operating data for 20 air-handling unit (AHU) systems, as obtained from a large commercial building. The fault detection performances of multiple strategies are compared and analysed under different degrees of bias in single AHU and multiple AHU systems. The verification results show that the proposed DBSCAN-SG-PCA model offers significant improvements in fault detection accuracy and fault identification sensitivity over the conventional PCA method. Compared with the SG-PCA model, the proposed model reduces the amount of data required for fault detection by an average of 13.7%, and the Youden index is increased by an average of 0.21. Furthermore, the fault detection accuracy of the proposed model is ±0.7 °C.