Prediction quality, domain adaptation and robustness of machine learning methods: a comparison

Abstract

Domain or database shift causes performance degradation in machine learning models encountering real-life scenarios. However, it is not clear how and to what extent this degradation can be prevented, and which methods are more robust against that. In this paper, we compare a workflow based on conventional machine learning methods and deep neural networks for condition monitoring with emphasis on domain shift. It is shown that possible domain shifts can be detected using visualization techniques at feature level. Also, the conventional method shows superior results in the domain shift scenario compared with the deep learning model. Finally, domain adaptation is used to improve the models’ performance.