| dc.description.abstract | This study addresses the fault diagnosis problem of Air Handling Unit (AHU) systems by proposing an unsupervised learning-based domain adaptation method. It aims to resolve the issue of model performance degradation caused by the data distribution differences between various types of AHUs. The research background highlights the need for improved energy efficiency and the fact that HVAC systems account for 58.62% of electricity consumption in office spaces. To tackle the data distribution inconsistency between Dual Duct Air Handling Units (DDAHU) and Single Duct Air Handling Units (SDAHU), this study utilizes the dual duct system as the source domain and the single duct system as the target domain, enabling label-free cross-domain fault diagnosis.
The proposed method is centered on domain adaptation (DA), employing time-frequency feature encoders to extract data features and leveraging statistical alignment techniques such as Sinkhorn Divergence and Maximum Mean Discrepancy (MMD) to reduce distribution shifts between the domains. During the feature correction stage, an auxiliary decoder is used to reconstruct the original data, thereby optimizing feature representation and improving model accuracy for unlabeled data. Additionally, this study integrates feature importance analysis using Random Forest to identify key features, enhancing model performance while reducing redundant computations.
Experimental data are derived from simulated AHU operation datasets by Lawrence Berkeley National Laboratory, encompassing multi-dimensional operational parameters of dual-duct and single-duct systems. The experimental results demonstrate that our proposal algorithm significantly outperforms traditional domain adaptation methods such as DANN in the target domain. For specific fault types (e.g., stuck cooling coil valve), our proposal algorithm achieved a recall score of 0.9260, surpassing DANN and other comparative methods. Moreover, by incorporating both time and frequency features, this study significantly improves predictive accuracy in the target domain, underscoring the importance of frequency features in domain adaptation.
This approach offers multiple advantages, including a substantial reduction in reliance on labeled data, enhanced prediction capability for unlabeled data, and potential energy savings. In the future, this system could be applied to high-energy-consumption office buildings and factories domestically and internationally, achieving improved HVAC energy efficiency and intelligent operations while contributing to the energy saving and carbon reduction target. | en_US |