基于伪标签的弱监督迁移学习模型

Weak Supervised Transfer Learning Model Based on Pseudo-label

针对目标域标记数据少导致迁移模型泛化能力差的问题,提出基于伪标签的半监督迁移学习模型WSTLPL。卷积神经网络用于学习原始振动数据的可迁移特征,用源域数据预训练网络;利用该网络预测目标域数据类别,将分类概率最大的类标签作为数据的伪标签。根据域自适应和伪标签学习的正则化项,对神经网络的参数施加约束,以减少学习到的可迁移特征的分布差异。结果表明:与现有诊断模型相比,该迁移模型的准确率更高。

A semi-supervised transfer learning model WSTLPL(weakly supervised transfer learning based on pseudo-label)was proposed to solve the problem of poor generalization ability of transfer model due to the lack of target domain label data.The convolu⁃tional neural network was used to learn the transferable characteristics of the original vibration data,and the network was pre-trained with the source domain data;the trained network was used to predict the target domain data category,and the class label with the high⁃est classification probability was taken as the pseudo-label of data.According to the regularization term of domain adaptation and pseu⁃do-label learning,the parameters of neural network were constrained to reduce the distribution differences of the learned transferable features.The results show that the proposed transfer model is more accurate than the existing diagnostic model.