Cost-Sensitive Bayesian Networks for FDD in HVAC Systems: A Fault Impact-Driven Approach

This study proposes a cost-sensitive FDD framework based on Tree Augmented Naïve Bayes (TAN) BNs, aimed at improving fault prioritization in multiclass classification tasks. The framework was trained and tested on 34 operational conditions—33 faulty and one normal state—using separate models for summer and winter. Misclassification costs, derived from a fault impact analysis based on energy and comfort related KPIs, were used to guide the learning process. The cost sensitive TAN model achieved detection accuracies of 93% (winter) and 99% (summer), and diagnosis accuracies of 76% and 82%, respectively. Precision and recall analyses showed significantly better performance for the 17 most impactful faults compared to the least impactful ones, confirming the model ability to prioritize the faults that matter most while maintaining high overall accuracy.