Decoupling-parallel fault diagnosis and model interpretation for solar-assisted air source heat pump systems

Solar-assisted air source heat pump (SAASHP) systems are highly efficient renewable energy technologies widely adopted for residential heating. However, their complex structure, strong inter-subsystem coupling, stochastic solar radiation, and dynamic user demands often lead to various operational faults, including solar collector (SC) fouling, air source heat pump (ASHP) performance degradation, sensor deviations, and flow leakage. Conventional data-driven fault diagnosis (FD) methods perform poorly in such strongly coupled systems due to mutual interference among subsystems and the lack of interpretability in deep learning models. To overcome these challenges, this paper proposes a novel decoupling-parallel (DP) diagnostic strategy, which systematically decouples the SAASHP system into SC and ASHP subsystems along temporal and physical dimensions. For the two subsystems, two dedicated convolutional neural network (CNN) models are developed in parallel and fused via decision-level integration. Based on energy conservation principle and low-cost sensor measurements, the two models achieve system-wide FD through decision fusion. A validated simulation SAASHP model is used to evaluate the DP strategy under eight common faults with four severity-levels in both parallel and series SAASHP configurations. Results show that DP achieves average accuracy of 0.945, with average misdiagnosis and missed diagnosis rates of 0.05 and 0.06, respectively. It outperforms traditional machine learning algorithms by over 0.2 in accuracy and reduces mis- and missed diagnosis rates by nearly 0.66 and 0.12 compared to a non-decoupling CNN (NDC) diagnostic strategy. The diagnostic performance and model robustness is further validated by analyzing influencing factors: varying training sizes, steady-state conditions, six geographic locations (six Chinese cities), and future climate scenarios (2030, 2060). Layer-wise relevance propagation (LRP) is employed for model interpretation, which indicates that DP effectively identifies critical fault-discriminative features with localization accuracy of 90%. This reveals that DP is highly reliable and provides an interpretable modeling framework for FD in multi-energy coupled building systems.