Fault Detection, Diagnosis and Prognosis in HVAC Air Handling Systems

As key sub-systems of HVACs, air handling systems are used to condition air to satisfy human thermal comfort and air quality requirements. Fault diagnosis is critical since it allows system operators to know which faults have occurred, how critical they are, and improve the system availability. Additionally, fault prognosis is critical since it allows system operators to know Remaining Useful lives of systems and their components, and prevents unexpected breakdowns. However, fault diagnosis of known and new fault types and fault prognosis are complex since 1) fault propagation across components is hard to capture; 2) measurement noise cause many false alarms; 3) impacts of changing environments are hard to be captured in Hidden Markov Models (HMMs); 4) normal conditions or known fault types may be identified as new fault types falsely; 5) Hidden Semi-Markov Models (HSMMs) perform well in fault prognosis but are time-consuming; and 6) HSMMs capturing impacts of concurrent failure modes are hard to establish. In this thesis, to capture fault propagation in an efficient manner, a new coupled HMM is developed. To filter out false alarms, “coupled statistical process control” techniques are developed. To adapt to changing environments, a new online learning algorithm is developed. To identify new fault types with low false identification rates, a robust statistical method is developed. To estimate states of HSMMs with low computational effort, a statistical method is developed to focus on potential state transition points. To reflect accumulation of fault impacts, a statistical method is developed based on Monte-Carlo simulation.