CH-24-C070 - A Framework for Automated Fault Detection in Light Commercial Buildings HVAC System

Automated fault detection and diagnosis (AFDD) in HVAC systems can provide significant energy savings as demonstrated in many previous studies. However, one of the main challenges facing AFDD field implementation is the lack of standardized data structures in most commercial buildings especially without any labeling of faulty data, which makes it challenging to develop customized AFDD algorithms for specific buildings. Most previous studies investigated different machine learning methods on labeled data prepared specifically for training AFDD algorithms, which is not practical for any large-scale field deployment. To this end, the goal of this paper is to present a framework for automated identification and labeling of faulty data in typical building automation systems (BAS) used in light commercial buildings. The proposed method is applied in a light commercial building with a single packaged rooftop unit located in Montreal, Canada. The relative simplicity of HVAC systems in these buildings, and subsequently their replicability, can enable the deployment of scalable AFDD algorithms. Rule mining can be considered an initial step to find the Basic rules in an HVAC system to categorize normal and abnormal operating conditions. Accordingly, in this study, a large dataset without any labeled faults collected from the building’s BAS was used. After data cleaning, Principal component analysis (PCA) was employed to determine the principal directions. Afterward, the density-based clustering technique DBSCAN was applied, most specifically to identify the noise and potential faults by finding outliers in the low-density region of PC space. In the next step, the original data frame has been labeled based on the clusters detected by DBSCAN. Finally, a rule-mining method has been applied to find the rules that lead to abnormal operating clusters and outliers. Ultimately, the combination of antecedents that cause potential faults has been studied and discussed. The identified sequences of events serve as a basis for further investigations of unlabeled data sets in AFDD of light commercial building HVAC systems.