End-to-end residual learning embedded ACWGAN for AHU FDD with limited fault data
Bi, Jian, Yan, Ke, and Du, Yang (2025) End-to-end residual learning embedded ACWGAN for AHU FDD with limited fault data. Building and Environment, 270. 112529.
![[img]](https://researchonline.jcu.edu.au/style/images/fileicons/application_pdf.png) |
PDF (Published Version)
- Published Version
Restricted to Repository staff only |
Abstract
Air handling unit (AHU) fault detection and diagnosis (FDD) is essential for indoor environment regulation and building energy conservation. Traditional data-driven FDD methods typically require ample and balanced input data. However, in real-world scenarios, fault data from AHU systems are significantly scarcer than non-fault data. The existing solutions to this problem often utilize generative adversarial network (GAN) for fault data augmentation and then train a classifier for AHU FDD, which requires significant resources for training and testing the model. Additionally, the quality of synthetic data cannot be refined in real-time. This paper proposes a novel end-to-end GAN to achieve data augmentation and reliable AHU FDD in real-time. The proposed residual learning embedded auxiliary classifier Wasserstein GAN with gradient penalty is specifically designed for AHU FDD with imbalanced datasets. First, Wasserstein distance and gradient penalty are introduced into the loss function to stabilize the training of auxiliary classifier GAN and prevent mode collapse. Next, residual learning is employed to build robust deep neural networks, enhancing the generator and discriminator's performance to produce reliable synthetic fault data. Finally, performance evaluation experiments are conducted using controlled fault samples on two AHU datasets. Experimental results demonstrate that the proposed method achieves the best performance on both datasets, with F1 scores improved by at least 4.83 % and 1.33 % over traditional supervised FDD methods when there are 50 real fault samples. This work addresses the challenge of training data imbalance in traditional supervised data-driven FDD methods and refines existing unsupervised GAN-based FDD approaches.
| Item ID: | 88303 |
|---|---|
| Item Type: | Article (Research - C1) |
| ISSN: | 1873-684X |
| Keywords: | Air handling unit, Deep learning, Fault detection and diagnosis, Generative adversarial network, Limited fault data |
| Copyright Information: | © 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies. |
| Date Deposited: | 07 Apr 2026 02:00 |
| FoR Codes: | 33 BUILT ENVIRONMENT AND DESIGN > 3302 Building > 330201 Automation and technology in building and construction @ 50% 46 INFORMATION AND COMPUTING SCIENCES > 4611 Machine learning > 461103 Deep learning @ 50% |
| SEO Codes: | 28 EXPANDING KNOWLEDGE > 2801 Expanding knowledge > 280104 Expanding knowledge in built environment and design @ 50% 28 EXPANDING KNOWLEDGE > 2801 Expanding knowledge > 280115 Expanding knowledge in the information and computing sciences @ 50% |
| More Statistics |
Actions (Repository Staff Only)
 |
Item Control Page |