Experimental analysis of fault impacts in AHU and development of a fault detection and diagnosis data-driven process

The building sector accounts for up to 40% of the global energy demand and contributes 15 to 33% of greenhouse gas emissions globally, while HVAC systems comprise almost half of the buildings’ energy consumption and 10–20% of total energy consumption. Furthermore, it is estimated that HVAC systems could increase the consumption of these sectors by 15% ~ 30%. Ensuring fault-free operation of HVAC systems is difficult due to the complexity of the systems, complex interactions between HVAC systems, buildings and occupants, lack of proper maintenance, failure of components, or incorrect installation. In this context, Fault Detection and Diagnosis (FDD) approaches allow for the identification of faults or anomalous operational patterns in HVAC system components and the diagnosis of their root causes. FDD is then crucial for promptly identifying and correcting faults that in complex systems such as AHUs, can lead to significant energy waste, shorter equipment life, occupant discomfort, poor indoor air quality (IAQ), and increased operating costs. In this regard, several methods of FDD in building energy systems have been proposed since the late 1980s in order to reduce the consequences of HVAC faults while the majority of them require at least one of the laboratory data, simulated data, or real data to be developed and consequently validated. Thus, providing reliable ground truth data of HVAC systems and AHUs with different technical characteristics is of great importance for advances in HVAC FDD methods. Thus, acquisition of reliable ground truth data and development of more efficient FDD tools are the most significant goals of current scientific literature. This thesis faces the above-mentioned challenges in two main parts. The primary objective of the first part is to experimentally examine the operational behavior of a typical single duct dual-fan constant air volume AHU in both faulty and fault free conditions. The investigation encompasses a series of experiments conducted under Mediterranean climatic conditions in southern Italy. This part investigates the performance of the AHU by artificially inducing seven distinct typical faults which is followed by fault impact assessment. The results obtained in the first part are used as the dataset to validate different Bayes theory related AFDD approaches which are proposed in the second part. This part of the thesis aims to propose AFDD methods for AHU by developing two continuous Bayesian Networks (BNs) and a discrete BN. Conditional Gaussian Network (CGNBN) and Kernel Density Estimation Bayesian Network (KDEBN) consist the continuous BNs while Tree Augmented Naïve Bayes classifier (TAN) is the BN based on discrete assumption of the input data. Furthermore, a cost-sensitive variant of each model is proposed to reduce the False Alarm Rate (FAR). Lastly, the results of the mentioned models are analyzed and compared the results and discussion section.