2019 年 139 巻 12 号 p. 1501-1508
Machine Learning (ML) techniques need a tremendous volume of training data. However, in operation and maintenance of industrial facilities, it is difficult to get such a volume of data due to the lack of sensors or infrequency of target phenomena. A promising approach to solve this problem is so-called data augmentation, which generates training data by using a prior knowledge or adding noise (perturbation) to original data. For this, Gaussian noise is generally used because of its simplicity. However, when the distribution of original data is not isotropic, the Gaussian-based augmentation breaks its shape, which causes so-called over regularization. In this paper, we propose a novel perturbation-induced data augmentation method, which does not require any prior knowledge and makes it easy to control the magnitude of perturbation. The novelty of the proposed method is to keep certain characteristics of shape of original distribution. The perturbation for this is generated by combined use of generative adversarial networks and newly proposed objective functions. We experimentally show that the proposed method enables to keep the gap between peaks of a mixed normal distribution. The effectiveness of the proposed method is also demonstrated in the case of an image classification task.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/