Most heating, ventilation, and air-conditioning (HVAC) systems operate with one or more faults that result in increased energy consumption and that could lead to system failure over time. Today, most building owners are performing reactive maintenance only and may be less concerned or less able to assess the health of the system until catastrophic failure occurs. This is mainly because the building owners do not previously have good tools to detect and diagnose these faults, determine their impact, and act on findings. Commercially available fault detection and diagnostics (FDD) tools have been developed to address this issue and have the potential to reduce equipment downtime, energy costs, maintenance costs, and improve occupant comfort and system reliability. However, many of these tools require an in-depth knowledge of system behavior and thermodynamic principles to interpret the results. In this paper, supervised and semi-supervised machine learning (ML) approaches are applied to datasets collected from an operating system in the field to develop new FDD methods and to help building owners see the value proposition of performing proactive maintenance. The study data was collected from one packaged rooftop unit (RTU) HVAC system running under normal operating conditions at an industrial facility in Connecticut. This paper compares three different approaches for fault classification for a real-time operating RTU using semi-supervised learning, achieving accuracies as high as 95.7% using few-shot learning.

Feature selection is a crucial problem in efficient machine learning, and it also greatly contributes to the explainability of machine-driven decisions. Methods, like decision trees and Least Absolute Shrinkage and Selection Operator (LASSO), can select features during training. However, these embedded approaches can only be applied to a small subset of machine learning models. Wrapper based methods can select features independently from machine learning models but they often suffer from a high computational cost. To enhance their efficiency, many randomized algorithms have been designed. In this paper, we propose automatic breadth searching and attention searching adjustment approaches to further speedup randomized wrapper based feature selection. We conduct theoretical computational complexity analysis and further explain our algorithms’ generic parallelizability. We conduct experiments on both synthetic and real datasets with different machine learning base models. Results show that, compared with existing approaches, our proposed techniques can locate a more meaningful set of features with a high efficiency.