Fault detection and diagnosis (FDD) are essential for enhancing the performance of heating, ventilation, and air conditioning (HVAC) systems, preventing energy waste, and ensuring indoor comfort. However, popular data-driven FDD approaches encounter challenges, such as the lack of high-quality labeled data, poor generalization, and the black-box nature of the models, which hinder their adoption in the market. Moreover, most existing studies only developed and validated their FDD models in simulation environments or laboratory settings, overlooking practical challenges of operational HVAC systems. First, this study focuses on implementing Diagnosis Bayesian networks (DBNs) in real-world settings, specifically, air handling units with heat recovery wheels in a campus building in the Netherlands. DBNs have been proven to be promising solutions with advantages in interpretability, robustness to uncertainties, and flexibility. Second, a Kafka-based framework is introduced for real-time monitoring in HVAC systems, enabling continuous and scalable data processing. Third, a comprehensive diagnosis analysis is conducted using both historical operational data and experimental fault data. The results reveal significant discrepancies between design documents and the actual operation of the HVAC system, and the DBN successfully identifies eight out of nine injected faults during experimentation. Additionally, the results uncover issues such as false positives due to DBN's limitations, inherent system faults, and unexpected HVAC system behaviors triggered by the simulated faults, offering critical insights into the operational challenges and diagnostic potential of DBNs in real-world HVAC systems. These contributions can advance the practical deployment of interpretable and robust FDD tools in building energy systems.