不平衡小样本基于局部域对抗适应网络的发动机振动预测模型

PDF下载HTML

Title:

Partial domain adversarial adaptation networks for imbalanced small samples in aeroengine vibration prediction

作者:

季友昌¹, 袁伟伟¹, 毛善斌², 任春红², 关东海¹

1. 南京航空航天大学 计算机科学与技术学院, 江苏 南京 211106;
2. 北京动力机械研究所, 北京 100074

Author(s):

JI Youchang¹, YUAN Weiwei¹, MAO Shanbin², REN Chunhong², GUAN Donghai¹

1. College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;
2. Beijing Power Machinery Institute, Beijing 100074, China

关键词:

不平衡数据; 域适应; 对抗学习; 振动预测; 异构迁移学习; 目标域; 小样本学习; 特征空间

Keywords:

imbalanced data; domain adaptation; adversarial learning; vibration prediction; heterogeneous transfer learning; target domain; few-shot learning; feature space

分类号:

TP391.41；TP18

DOI:

10.11992/tis.202210030

摘要:

在发动机振动预测中，实际装配数据样本量小且类别不平衡，难以直接建立有效的预测模型。迁移学习方法能够通过迁移源域知识来提高目标域模型性能，为此，本文提出了基于局部域对抗适应网络的发动机振动预测模型。将领域按标签分为多个局部域，通过多个局部域对抗适应网络将目标域样本映射到源域，保证小样本中的少数类得到正确迁移。并通过伪标签来解决目标样本的域转换，使用源域分类器给出可靠的预测结果。本文在多个真实数据集上验证了所提方法的有效性和泛化性，与其他方法相比，振动预测准确率能够平均提升15%左右。

Abstract:

In aeroengine vibration prediction, due to the imbalanced small samples of the actual assembly data, directly establishing an effective prediction model is difficult. Transfer learning can improve model performance in the target domain by transferring the knowledge of the source domain. Therefore, this study proposes an aeroengine vibration prediction model based on a partial domain adversarial adaptation network. Each domain is divided into multiple local domains according to labels. Through multiple local domain adversarial adaptation networks, the samples in the target domain can be mapped onto the source domain so that the minority class samples can be transferred correctly. The pseudo label is used to solve the domain transformation of the samples in the target domain, and the classifier of the source domain is used to provide a reliable prediction result. In this study, the validity and generalization of the proposed method are verified on several real datasets. Compared with other methods, the area under the curve and F1 of the vibration prediction model can be improved by approximately 15% on average.