A Study of Source Domain Data Selection for Successful Regressive Domain Adaptation

Considering that collecting the labeled data is an expensive and time-consuming job, an approach for utilizing scarce labeled data effectively is required. One such approach is domain adaptation. The domain adaptation tries to utilize the labeled data (i.e., source domain data) to build a model for the data without labels (i.e., target domain data). In this paper, we first propose a domain adaptation model using a spatiotemporal transformer as a feature extractor. We also discuss two metrics to be used for the selection of source domain dataset among multiple source domain datasets for a given target domain dataset. Please note that the domain adaptation model for each target domain dataset achieves the best performance with different source domain datasets. As the metrics, we examine the feasibility of prediction accuracy of the source domain dataset and the distance between features of the source domain dataset and the target domain dataset. The feature distance is calculated using the maximum mean discrepancy (MMD). From the experiments using the New C-MAPSS (N-CMAPSS) dataset, we first verify the feasibility of the proposed domain adaptation model. For eight datasets of the N-CMAPSS, our domain adaptation model reduces RUL prediction error (in RMSE) by 23.48% on average. Then, we show that the MMD distance is more promising as the selection metric. With the MMD distance, among eight datasets, the selection of the best source domain dataset is successful in five datasets.