Fault Diagnosis of Components and Sensors in HVAC Air Handling Systems With New Types of Faults

Air handling systems are the key sub-systems of heating ventilation and air conditioning (HVAC) systems. They condition and deliver air to satisfy human thermal comfort requirements and provide acceptable indoor air quality. Faults in their components and sensors may lead to high-energy consumption, poor thermal comfort, and unacceptable indoor air quality. Additionally, new types of faults may falsely be identified as known types. Identifying failure modes and their severities with low false identification rates is thus critical to know what faults occur and how severe they are. However, this is challenging, since 1) classifying both failure modes and fault severities generates many categories of failures, leading to high computational requirements; 2) updating model parameters to adapt to changing environments requires accurate recursive equations that are hard to obtain; and 3) model errors and measurement noise may cause high false identification rates in detecting new types of faults. In this paper, failure modes are identified by hidden Markov models (HMMs) and fault severities are estimated by filtering methods, leading to a decrease in the number of HMM states and low computational requirements. To adapt to changing environments, a new online learning algorithm is developed. In this algorithm, HMM parameters are obtained based on their posterior distributions given new observations, thereby avoiding the need for accurate recurrence equations. To identify new fault types with low false identification rates, a robust statistical method is developed to compare current HMM observations with those expected from existing states to obtain potential new types, and then confirm new types by checking whether observations have a significant change. Physical knowledge is then used to find the reason for the new fault type. Experimental results show that failure modes and fault severities of both known and new types of faults are identified with high accuracy.