The impact of improved PCA method based on anomaly detection on chiller sensor fault detection

Abstract

With the widespread use of building automation systems (BAS), a large amount of chiller operating data is often readily available, which provides a good basis for optimizing the control of unit sensors. However, these data are highly coupled and contain a large number of outliers which would significantly affect sensor fault detection. Compared with the traditional PCA method, the improved kernel PCA can effectively handle nonlinear data. Meanwhile, anomaly detection has shown excellent ability in rejecting outliers, so an anomaly detection step before fault detection is also necessary. Therefore, to address the problem of data presence coupling and outliers, this study attempts to investigate the impact of different anomaly detection methods on PCA and KPCA based chiller sensor fault detection. Based on the optimal combination of features screened by importance and correlation analysis to construct a fault detection model, the impact of two anomaly detection methods (Isolation Forest and K-means) on KPCA fault detection was analyzed and compared. To obtain general results, an experimental dataset of RP-1043 is used to validate the study. The results show that both anomaly detection methods can improve the overall quality of the original training data. Compared with the traditional PCA method, KPCA improves the detection efficiency by 1.7%–28.29% at all fault magnitudes. The IF-KPCA and Kmeans-KPCA improve the fault detection efficiency by 3.66%–14.63% and 3.17%-20% on top of KPCA. This study would provide reference for the optimal feature combination selection and anomaly detection method selection for practical engineering.