【作者】 刘学志； 王振雷； 王昕；

【Author】 XueZhi Liu;ZhenLei Wang;Xin Wang;School of Information and Science Engineering, East China University of Science and Technology;Engineering Research Center of Process System Engineering,Ministry of Education;

【机构】 华东理工大学能源化工过程智能制造教育部重点实验室； 过程系统工程教育部工程研究中心；

【摘要】 近年来,深度迁移学习的域适应方法被广泛应用于跨域轴承故障诊断问题,以减小不同工况下数据分布的差异。但是,由于深度学习模型通常考虑的是在高维空间的域适应,数据的低维特征不能很好地适应。因此,本文提出了一种基于双层对抗网络的故障诊断模型,旨在进一步增强模型的自适应能力。首先,通过一维卷积构建特征提取网络;然后,引入对抗并构建域混淆损失,将其反向传播到特征提取网络来匹配源域和目标域分布;其次,在低维输出空间中也引入对抗学习来帮助模型实现更好的自适应;最后,本文使用两个公开数据集(CWRU和PU数据集)进行验证。实验结果表明,所提方法相比于常见的域适应方法具有更高的准确度,验证了双层对抗网络在不同特征层次上执行输出空间的域适应的有效性。更多还原

【Abstract】 In recent years, the domain adaptation method of deep transfer learning has been widely used in cross-domain bearing fault diagnosis to reduce the difference of data distribution under different working conditions. However, since deep learning models usually consider domain adaptation in high-dimensional Spaces, low-dimensional features of data do not fit well.Therefore, a fault diagnosis model based on two-layer adversarial network is proposed in this paper to further enhance the adaptive capability of the model. First, the feature extraction network is constructed by one-dimensional convolution. Then, the antagonism is introduced and domain confusion loss is constructed, which is backpropagated to the feature extraction network to match the distribution of source domain and target domain. Secondly, confrontation is also introduced in the low-dimensional output space to help the model achieve better self-adaptation. Finally, the paper uses two publicly available datasets(CWRU and PU) for validation. The experimental results show that the proposed method has higher accuracy than the common domain adaptation methods, and verifies the effectiveness of the two-layer adversarial network to perform the domain adaptation of the output space at different feature levels.更多还原