Using thermal load matching strategy to locate historical benchmark data for moving-window PCA based fault detection in air handling units

Abstract

Principal component analysis (PCA) methods have been reported to successfully detect some faults. Unfortunately, for the actual systems, there are few literatures on how to retrieve the historical fault-free operating information as training data which is the benchmark to characterize system performance. This study proposes a dynamic thermal load matching strategy to locate historical candidate information based on mass and thermal balance. Seven parameters related to thermal load are deduced to replace some variables such as solar flux, heat gain, and heat dissipation which are usually unavailable in most real systems. Combining with the moving-window PCA fault detection method, the strategy is validated to detect the fault symptom in 52 fault-free and 54 fault days of air conditioning systems from ASHRAE 1312-RP. The data sizes of a time-series data window, moving speed, and historical candidate pool, are defined as 60, 10 and 900 data points, respectively. The detection results of fault symptom show that the proposed strategy exhibits higher percentages of fault symptom than those reported in the published literatures. On the other hand, the fault detection effects highly depend on the severity level of fault symptom, but present slightly tiny negative correlation with PCA similarity factors.