Automated fault diagnosis detection of air handling units using real operational labelled data and transformer-based methods at 24-hour operation hospital

Abstract

Automated Fault Detection and Diagnosis (AFDD) in Air Handling Units (AHUs) is essential for maintaining indoor air quality and extending the lifespan of HVAC systems. However, previous research has frequently been constrained by limited access to real operational data, primarily due to difficulties in data collection and complexities associated with accurate fault annotation. Additionally, transformer-based methods remain underutilized in AFDD applications despite their proven effectiveness in related domains. In this study, AHUs equipped with Constant Air Volume (CAV) systems operating continuously in a 24-hour hospital environment were specifically investigated. Data were collected over a one-year period using nine different sensors installed across eight AHUs. Four operational conditions were identified: normal operation and three distinct types of faults. Three transformer-based models—TFT, Informer, and Autoformer—were proposed and optimized through comprehensive hyperparameter tuning, resulting in the evaluation of a total of 792 models. Additionally, 1076 models based on seven traditional machine learning methods were optimized and evaluated. A detailed comparative analysis revealed that the Autoformer model outperformed all other evaluated methods, achieving an F1 score of 96.21 % and an accuracy of 96.02 %. Moreover, the Autoformer demonstrated efficient performance, capable of processing approximately 37.88 instances per second. The potential practical applications and implications of these findings for real-world operational conditions are further discussed in this research.