Domain Adaptation Training of a Transformer
Junwon Lee, Kyoungho Hwang, Minsuk Kwak, Minho Lee · 2022 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia) · 2022
Domain adaptation is a powerful technology that solves the problem of data distribution mismatch among different domains, and improves generalization performance in various deep learning application fields. However, this technique is not easily applicable to home appliances such as air conditioners that are important in daily life. In this study, we propose a Domain Adaptation Transformer(DAT) to predict the amount of refrigerant in air conditioners using deep learning based on a transformer encoder and the domain-adversarial training of neural networks (DANN). The proposed DAT is a novel deep-learning-based refrigerant prediction model that is not constrained by the specific types of air conditioners, and can be used to other types of air conditioners. We construct a novel dataset to develop and test our model on different types of air conditioners and experiments on those datasets. Experimental results demonstrate that our approach outperforms heuristic methods based on conventional physics phenomena and achieves excellent performance in domain adaptation tests.