Automated Fault Detection and Diagnosis in Light Commercial Building’s HVAC systems

Fault detection and diagnosis (FDD) in commercial buildings' HVAC systems can significantly reduce energy consumption. Faults in these systems occur due to various issues such as aging and inadequate maintenance. Commercial buildings in Canada covered an area of 709,029,612 m² in 2019, consuming 948,216,746 GJ of energy. HVAC systems are responsible for 25%-50% of this energy consumption. In the United States, faults in HVAC systems contribute to an additional energy consumption of 103 to 500 terawatt-hours (TWh) in the building sector. Detecting and diagnosing faults in HVAC systems can reduce energy consumption by 20% to 30%. Light commercial buildings, defined as commercial buildings with fewer than six stories and less than 2500 square feet, include bank branches, offices, and small industrial facilities. These buildings have similarities in the configuration and size of HVAC components, making it feasible to develop an FDD tailored for this class of buildings, that can be easily scaled up. This goal can be achieved using data-driven methods, which have gained popularity over the past decades by installing various sensors and collecting data integrated with Building Energy Management (BEM) systems. In this thesis, three different FDD methods have been developed and validated on light commercial buildings. The first method is a semi-supervised method that includes various techniques to handle the unlabeled raw data from BEMs, resulting in a final supervised Automatic Fault Detection (AFDD) system. The second method is a fully unsupervised novel AFDD method based on PCA time series fault detection. The third method is primarily based on the inverse model of the Air Handling Unit (AHU) of the HVAC systems. A typical light commercial building in Montreal, Canada, was used for all methods. Additionally, to validate the generalizability of the unsupervised method, another light commercial building, a small industrial facility in Ireland, was used as well. The first method successfully resulted in an AFDD that can detect and diagnose faults with almost 90% accuracy, performing better in condition-based faults than control faults. The unsupervised method showed very good results in terms of generalizability. It was able to detect faults and report the problematic inputs and locations to the HVAC operators. Although the unsupervised method cannot completely diagnose condition-based faults, it provides very good information based on the system's behavior, enabling operators to diagnose the faults. Finally, the inverse modeling revealed that a physics-based neural network can outperform neural networks and genetic algorithms in modeling the system inversely and detecting anomalies mostly related to energy consumption.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.