A semi-supervised approach to fault detection and diagnosis for building HVAC systems based on the modified generative adversarial network

Developing efficient fault detection and diagnosis (FDD) techniques for building HVAC systems is important for improving buildings' reliability and energy efficiency. The existing FDD methods can achieve satisfying results only if there are sufficient labeled training data. However, labelling the data is often costly and laborious, and most data collected in practice are unlabeled. Most of the existing FDD methods cannot leverage the unlabeled dataset which contains much information beneficial to fault classification, and this will impede the improvement of the FDD performance. To deal with this problem, a semi-supervised FDD approach is proposed for the building HVAC system based on the modified generative adversarial network (modified GAN). The binary discriminator in the original GAN is replaced with the multiclass classifier. After the modification, both the unlabeled and labeled datasets can be utilized simultaneously: the modified GAN can learn the data distribution information present in unlabeled samples and then combine this information with the limited number of labeled data to accomplish a supervised learning task. Additionally, a novel self-training scheme is proposed for the modified GAN to correct the class imbalance in both labeled and unlabeled data. With the self-training scheme, the modified GAN can still efficiently exploit the information contained in unlabeled data to enhance the FDD performance even if the class distribution is highly imbalanced. Experimental results demonstrate the effectiveness of the proposed modified GAN-based approach and the self-training scheme. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

This study proposes a semi-supervised FDD approach based on a modified GAN, which can simultaneously utilize both unlabeled and labeled datasets to improve the performance of building HVAC systems. With a self-training scheme, the modified GAN can efficiently utilize the information contained in unlabeled data, enhancing FDD performance even in the presence of imbalanced class distribution.