Predictive maintenance for residential air conditioning systems with smart thermostat data using modified Mann-Kendall tests

Abstract

Predictive maintenance through fault detection and diagnosis (FDD) is an effective approach to correct soft faults in residential air conditioners before complete failure. In particular, gradual degradation of heating or cooling capacity is the most common soft fault often caused by refrigerant leakage and goes largely unnoticed by occupants. Traditional FDD methods rely on extracting features from sensor measurements of the refrigeration cycle and need labeled fault-free or faulty data to establish models and rules. These methods are commonly used for large commercial systems. For residential systems, however, installing additional sensors in the refrigeration cycle and collecting labeled data from lab experiments are cost-prohibitive for manufactures. In contrast, smart thermostats are widely adopted by residential homeowners with data streamed to the cloud, enabling powerful FDD methods with limited sensor information. This paper presents two methods, namely the hourly and daily analysis, for extracting key data features from unlabeled smart thermostat data and then applying modified Mann-Kendall statistical tests to identify significant trends in cooling capacity. The effectiveness of these two methods are first evaluated by simulated data. After that, they are applied to approximately 10,000 residential air conditioners for historical trend detection and real-time condition monitoring, with case studies selected from a few verified faulty systems to validate the approach. The methods would allow technicians to identify and prioritize residential systems with gradual degradation for repair prior to catastrophic failure.