Machine Learning-Based Fault Detection in a R-290 Residential Heat Pump Unit
Anomaly and fault detection techniques are gaining increasing importance in the HVAC sector due to their ability to improve system reliability. While historically used in industrial applications, where costs were related to the plants, recent advances in digital integration have enabled their deployment even in smaller HVAC systems, such as residential heat pump units. Detecting early signs of inefficient operation can help predict faults, thereby reducing costs and user discomfort due to system failures, and preserving long-term performance. In this work, a comparative analysis is presented of various supervised fault detection algorithms applied to a typical R-290 Air-to-Water (A/W) heat pump application. A Matlab/Simscape-based virtual simulation environment of a R-290 heat pump unit is used to replicate both normal and faulty conditions. The simulation model allows for the generation of consistent and repeatable datasets, overcoming the limitations of time-consuming laboratory data collection. The algorithms under investigation include Support Vector Machines (SVM), Histogram-based Gradient Boosting (HistGB) and Neural Networks (NN). Results provide a performance overview of each method in simulated fault scenarios, offering insight into their suitability for real-world deployment and potential integration into field-ready HVAC monitoring systems.