Adaptive Density-Based Machine Learning for Fault Detection in Heat Pump Systems

This paper presents a proof-of-concept case study of a modular fault detection and diagnosis approach applied to a German residential hybrid heating system. We utilize volume flow data from external measurement equipment to analyze stable operational and miscurrent modes within the heating circuit. Subsequently, we train several custom implementations of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm using the system’s supply and return temperature data. The models demonstrate high classification accuracy in distinguishing between operational modes; however, they struggle to effectively manage unlabeled transient modes. We then reduce the input space of the models and discuss transfer options to different heating systems with partially overlapping topologies. Although the reduced models are unable to sufficiently differentiate between individual faulty operational modes, they reliably separate them from other nominal operational modes.