Implementing automated fault detection and diagnosis (AFDD) for air handling units (AHUs) is crucial for maintaining optimal indoor air quality and extending the operational life of equipment. However, previous studies often encountered challenges arising from limited real-world operational data and difficulties in accurately labeling fault conditions. Additionally, tabular-based methods, despite exhibiting robust performance in various applications, have been relatively underexplored in AFDD research. To address these research gaps, this study focuses on constant air volume (CAV) AHUs that operated continuously in a large-scale office building for 1 year. Data were collected from 18 sensors installed across 20 AHUs and categorized into four operational states: normal operation and three distinct fault conditions. Comprehensive hyperparameter optimization was performed, resulting in the evaluation of 2,592 deep tabular-based models (684 configurations of TabTransformer and 1,944 configurations of TabNet) alongside 2,968 traditional machine learning models for comparison. The TabNet-based method proposed in this study demonstrated superior performance, achieving an average F1 score of 96.34% and an accuracy of 96.23%. Additionally, the approach showed significant computational effectiveness, processing roughly 46 data samples per second. These results underscore the practical value and effectiveness of using tabular-based approaches to enhance AFDD reliability and efficiency in operational AHUs.
Seunghyeon Wang (Tue,) studied this question.