Experimental Performance Evaluation of Cost-Sensitive Bayesian Networks for Fault Detection and Diagnosis in HVAC Systems

Faulty HVAC operation can result in up to 30% higher energy consumption of buildings, reduced equipment life, occupant discomfort, poor indoor air quality, and increased operational costs. Fault Detection and Diagnosis (FDD) methods are crucial for identifying and rectifying these issues. This paper proposes and validates various Bayesian-based FDD methods using experimental data from a real AHU with artificially implemented faults. The proposed methods include a Conditional Gaussian Network (CGN) and a Tree Augmented Naïve Bayes classifier (TAN). A sensitivity analysis is performed to determine the optimal number of input variables for the best trade-off between model complexity and fault diagnosis performance. Additionally, a cost-sensitive process is implemented for both BN models to reduce the False Alarm Rate (FAR). The performance and effectiveness of these models were then analyzed and compared against a baseline ML algorithm (i.e., Random Forest), demonstrating their potential to enhance FDD approaches in building energy systems.