Machine learning-based fault detection and preliminary diagnosis for terminal air-handling units

Abstract: With the advent of Artificial Intelligence (AI) powered classification techniques, data-driven Fault Detection and Diagnosis (FDD) methods have become increasingly prominent in smart building implementation. Of these, cluster analysis is particularly promising for Building management system (BMS) data. This paper presents an unsupervised learning-based strategy for detecting faults in terminal air handling units as well as the systems serving them. Historical sensor data is pre-processed with PCA to reduce dimensions, followed by OPTICS clustering, which is compared with k-means. OPTICS outperformed the latter, readily identifying noise and had high accuracy across all seasons.