Hybrid Detection for Heat Pump Contamination Using Physics-Informed Machine Learning

Abstract

The deployment of residential heat pump systems is a key enabler of the decarbonization of the heating sector. However, their long-term reliability remains a barrier to sustained performance and user acceptance. A major degradation driver is water contamination within the hydraulic circuit, which leads to fouling, scaling, and corrosion of components such as plate heat exchangers – ultimately reducing efficiency and shortening system lifetime. Although installation procedures and operational filtration measures, including magnetite filtration, aim to reduce particle accumulation, continuous condition-based monitoring of component degradation remains limited. To address the scarcity of real-world failure data for training predictive models, this paper proposes a physics-informed, data-prior approach that combines physical knowledge with machine learning. Instead of embedding physics into the model architecture or loss functions, the approach incorporates it at the data level by generating labeled healthy and faulty scenarios through a physics-based laboratory setup. This enables the model to learn degradation patterns grounded in physical behavior, supporting early fault detection and producing outputs that remain interpretable and plausible for domain experts. The approach is demonstrated on a plate heat exchanger contamination use case. A design-of-experiments campaign in a climate chamber generated labeled data representing healthy, moderately contaminated, and severely contaminated states. A Random Forest classifier achieved consistent cross-validation performance (AUC ≈ 0.98) with low variance across folds. Precision–recall analysis revealed robust early fault detection, with average precision values of approximately 0.96 for moderate contamination and 0.97 for severe contamination. Cumulative gain and lift analyses indicated that inspecting the top 20–40 % of systems ranked by model risk can identify 80–100 % of the contaminated cases, supporting efficient maintenance prioritization. Model-derived feature importance was assessed using Gini importance and subsequently validated through expert review, enabling interpretable failure logic for condition-based maintenance strategies. The results demonstrate that combining physically grounded data, supervised machine learning, and explainable diagnostics provides a transferable hybrid approach for interpretable reliability assessment and condition-based monitoring beyond the specific case.