PCA-based method of soft fault detection and identification for the ongoing commissioning of chillers

This paper presents a new PCA-based method of soft fault detection and identification for the ongoing commissioning of chillers, which is composed of the three main phases: threshold model training, outliers detection, and variables identification. The method is applied to Building Automation System (BAS) trend data from a case study building. The first case used one week of data for training the threshold model, and then applied the model to two months of application data, both data sets being collected during the summer of 2009. The threshold model detected 139 outliers of normal operation, and identified the variables that caused those abnormal operation outcomes. In the second case, the measurements over the full summer of 2009 were used for training the threshold model, which was then applied to measurements of summer seasons of 2010-2015. The 2009 data set was collected from the first year of operation, and partially under commissioning. All subsequent years (2010-2015) present less than 10% of detected outliers over the entire data set for all combinations of retained Principal Components. The reduction of number of detected outliers can be due to the corrections along the first year of operation. There are still a few detected outliers due to sensors and components degradation which usually occurs in HVAC systems after initial commissioning.