Fault detection and diagnosis for heat recovery ventilation using 4S3F method

Abstract

Fault detection and diagnosis (FDD) are crucial to improving the efficiency of heating, ventilation, and air conditioning (HVAC) systems, reducing energy waste, and maintaining indoor comfort. Diagnostic Bayesian Networks (DBNs) present a compelling approach, offering robustness to uncertainty, adaptability to different sensor configurations, and interpretable inference. Existing FDD studies for air handling units (AHUs), however, are often limited to simulation or laboratory settings, seldom consider AHUs with heat recovery wheel (HRW) in operation, and rarely analyze how diagnostic performance changes under diverse sensor configurations. This study defined three practical sensor configurations (Sensor-Rich, Standard, and Limited) based on international guidelines and a practical survey, developed a corresponding DBN framework, and evaluated its performance on seventeen common faults using real-world data from an AHU in a Dutch office building. Existing FDD studies are often limited to simulation or specific Air Handling Unit (AHU) types with fixed sensor configurations, rarely investigating AHUs with heat recovery wheels, which are common in Europe. This study addresses these gaps by first defining three sensor configurations (Sensor-Rich, Standard, and Limited) based on international guidelines and a practitioner survey. A DBN-based FDD model was then developed for these configurations using historical data, expert knowledge and subsequently evaluated for its ability to diagnose seventeen common faults in an operational AHU with heat recovery wheel.The DBN correctly diagnosed fifteen, nine, and four faults for these configurations, respectively. The results show that increasing sensor availability improves overall diagnostic performance. However, certain cases demonstrate that additional measurements can also introduce conflicting evidence and reduce diagnostic accuracy. The study suggests that sensor selection must be combined with effective DBN modeling strategies to achieve robust diagnosis. Taken together, the analysis of key sensors and DBN modeling practices provides practical guidance for designing and implementing DBN-based FDD in common European AHU systems under diverse sensor configurations.The results indicate that increasing sensor quantity alone does not improve FDD performance; strategic sensor selection, placement, and effective data processing are also crucial.