Fault diagnosis of HVAC: Air delivery and terminal systems

Abstract

Supply/return fans and Variable Air Volume (VAV) boxes are key components of Heating Ventilation and Air Conditioning (HVAC) systems. Fans deliver conditioned air to rooms and VAV boxes control airflow rates to satisfy human thermal comfort and ventilation requirements. Faults in these components and their sensors may lead to high energy consumption and poor thermal comfort. Identifying failure modes and their severities is thus critical in guiding maintenance crew to know what, where and how severe the faults are. Diagnosing faults of components and sensors is difficult because (1) component faults and sensor faults may have similar effects and thus hard to distinguish; (2) capturing both failure modes and fault severities may generate many system states, leading to high computational requirements; and (3) capturing variable loads in models leads to additional decision logic complexity. A component fault may cause multiple measured variables to change. A sensor fault only causes one variable to change, leading to violations of relationship among variables. Thus, the two kinds of faults are distinguished. To reduce computational requirements, failure modes are identified first and fault severities are then estimated based on the identified failure mode. This also has the beneficial effect improving the condition number of the fault severity estimation problem. To estimate states while tracking variable loads, a new online learning algorithm for estimating the hidden Markov model parameters is developed. Experimental results show that failure modes and fault severities are identified with high accuracy as quantified by the F-measure that integrates the precision and recall.