A sensor fault detection strategy for air handling units using cluster analysis

Sensors are an essential component in the control systems of air handling units (AHUs). A biased sensor reading could result in inappropriate control and thereby increased energy consumption or unsatisfied indoor thermal comfort. This paper presents an unsupervised learning based strategy using cluster analysis for AHU sensor fault detection. The historical data recorded from sensors is first pre-processed to reduce the dimensions using principal component analysis (PCA). The clustering algorithm Ordering Points to Identify the Clustering Structure (OPTICS) is then employed to identify the spatial separated data groups (i.e. clusters), which possibly indicate the occurrence of sensor faults. The data points in different clusters are then checked for temporal separation in order to confirm the occurrence of sensor faults. The proposed sensor fault detection strategy is tested and evaluated with the data collected from a simulation system. The results showed that this strategy can detect single and non-simultaneously occurred multiple sensor faults in AHUs. The fault detection results were not strongly affected by the selection of the user defined input parameters required in OPTICS.