A Probability Extension of PCA to Detect and Diagnose Sensor Faults in Air Handling Units

Abstract

Due to sensor faults, it is a challenge to successfully detect and diagnose component faults in HVAC systems. The Principal Component Analysis (PCA) method has become a popular method to tackle this problem in recent years, but PCA is not capable of isolating sensor faults, such as sensor bias or sensor noise. The intention of this paper is to take sensor noise into account. This is accomplished by including sensor noise and sensor drift into a Bayesian probability calculation framework. In this approach, both of these potential faults are associated with a probability score once the system detects a fault. Component faults are not taken into account because we assume the PCA method is only the first step in detecting and diagnosing faults in HVAC systems. The sensor location effect has already been eliminated in the training process, so it is not considered either. This paper firstly discusses the drawbacks of applying traditional PCA method. For instance, we show that a sensor drift fault diagnosed by this method could actually be caused by sensor noise instead. The second part of the paper shows that by applying Bayesian probability calculations within the PCA analysis process, the false alarms caused by sensor noise can be partially excluded.