A hybrid SMOTE and Trans-CWGAN for data imbalance in real operational AHU AFDD: A case study of an auditorium building

Abstract

Class imbalance remains a significant challenge in Automated Fault Detection and Diagnosis (AFDD) for Air Handling Units (AHUs), as normal operating conditions significantly outnumber rare fault events. Prior studies mainly relied on simulated or laboratory-generated datasets, limiting their applicability to real-world scenarios due to insufficient operational complexity. This study introduces a novel hybrid data augmentation method combining the Synthetic Minority Over-sampling Technique (SMOTE) with Transfer Conditional Wasserstein Generative Adversarial Network (Trans-CWGAN), applied to real operational data collected over one year from an auditorium building equipped with 13 AHUs. Through hyperparameter optimization, a total of 1212 distinct datasets were generated across augmentation strategies. Among these strategies, the SMOTE-based Trans-CWGAN approach consistently delivered superior results. Specifically, TabNet achieved the highest performance, with a mean F1 score of 98.68 % and accuracy of 98.98 %, followed by RNN-LSTM (F1: 96.56 %, accuracy: 95.84 %). Even the DT model significantly improved from its initial baseline F1 score of 73.53 %. These findings underscore the effectiveness of integrating SMOTE and Trans-CWGAN to mitigate class imbalance, highlighting its strong potential for practical deployment in real-world HVAC monitoring systems.