Generative modeling-aided data-driven fault detection and diagnosis for heating, ventilation and air-conditioning systems of next-generation intelligent buildings
Publication Date
July 18, 2024
Creator
Abstract
Heating, ventilation and air-conditioning (HVAC) systems are responsible for 40-50% of the total building energy usage. The occurrences of HVAC faults lead to an additional 20% or more energy waste for HVAC systems. The automated fault detection and diagnosis (FDD) has the potential to reduce HVAC energy consumption by 10−40%. Recently, data-driven FDD methods have become the mainstream, due to the advancement of data science techniques and the increasing data availability in buildings. But there are still major challenges for achieving high data-driven FDD performances in practice, such as the class imbalance between fault data and normal data. Previous studies have evaluated the impact of this issue on FDD performances, and also proposed solutions to this practical issue. However, the influence of another data imbalance issue, the fault data intra-class imbalance under different climate conditions, has not yet been explored and addressed. The absence of fault data under certain climatic condition(s) may also affect FDD performance under these climatic condition(s).
This thesis assesses the impacts of HVAC single and multiple faults under different climatic conditions, and also evaluates HVAC FDD performances under different fault data seasonal intra-class imbalanced scenarios. To conduct the fault impact analyses under different climatic conditions, a series of typical variable air volume (VAV) system faults were modelled using Python scripts and a widely-used building performance simulation tool, EnergyPlus. The fault impact simulations were performed by introducing developed fault models into a verified prototype office building model. Based on fault simulations, the fault impacts under different seasons and climate periods were analysed. The results show that the fault impacts are affected by climatic condition variations and also present seasonal patterns.
Based on the insights acquired from fault impact analyses, the FDD performance for a combined convolutional and recurrent neural network classifier was assessed under fault data seasonal intra-class imbalanced scenarios, and an ideal fault data scenario covering climatic conditions from all the seasons. The developed FDD model was trained using four simulated data scenarios representing fault data seasonal intra-class imbalanced scenarios, and an ideal scenario. The trained model under each scenario was tested using each season’s dataset. Moreover, further experiments based on real fault datasets were used to validate the findings
from the experiments based on simulated fault datasets. As a result, the fault data experiments revealed two key findings: (1) The FDD architectures trained on sufficient fault data under a certain season(s) show poor generalization ability to detect and identify the faults under other unseen seasons; (2) The coverage of fault data under different seasons is more crucial in enhancing the FDD performances than the amount of fault data under a certain season(s).
To address this practical issue leading to FDD performance degradation under FDD model’s unseen seasons, a multi-label conditional Wasserstein generative adversarial network (ML- CWGAN) by incorporating both fault and season labels into its framework, was developed. The ML-CWGAN was integrated with the baseline FDD classifier to evaluate the quality of synthetic fault data. The FDD performances led by different GAN variants, i.e., multi-label conditional GAN (ML-CGAN), single-label CWGAN and the proposed ML-CWGAN, were evaluated and compared under 14 fault data seasonal imbalanced scenarios and 2 seasonal balanced but insufficient scenarios. The proposed ML-CWGAN demonstrated its ability to adapt to a wider range of seasonal imbalanced scenarios and was more robust in generating high quality fault samples. The results indicated that the ML-CWGAN considerably improved the baseline FDD performances by up to 5% in terms of F1 score, and was able to enhance FDD performances under all imbalanced scenarios. Compared with the SL-CWGAN, the ML-CWGAN-aided FDD outperformed the SL-CWGAN-aided FDD by an average F1 score of 1% and the SL-CWGAN performed poorly in improving FDD performances under some scenarios.
The future directions can be explored in two aspects, i.e., fault modeling and impact analysis aspect and FDD aspect. Future studies can explore HVAC fault impacts under different climate zones, such as hot climates, to compare the differences in fault impacts. The fault impacts from other types of HVAC systems, apart from VAV systems, are worth exploring. Besides, evolving fault intensity and fault occurrence frequency can be considered in fault modeling for simulation-based fault impact analyses. As for the FDD aspect, evaluating the impact of the fault data intra-class imbalance among different working conditions on HVAC FDD performances can be considered as a future direction. The integration of a high-quality synthetic fault sample selection protocol into the proposed framework is required to filter out high-quality fault samples. Moreover, Other variants of GANs, such as auxiliary classifier GAN, can be attempted to generate synthetic fault samples in future.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Subjects (LC)
Associated Schools / Departments
Department of Architecture and Built Environment (UK)
eprints ID
78249
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Zhong_Fangliang_16518687_Final_Submission.pdf
Type
Full-text
Description
Examined. Corrected thesis
Size
12.71 MB
Format
Adobe PDF
Checksum (MD5)
71adcfea0e1254cd9e992d056fb2e526