Abstract: Aiming at the insufficient labeled fault data in real cases, a method was proposed based on deep learning and subdomain adaptation. The domain-shared one dimensional CNN was first used to extract transferable features from the fault data. Then, the multi-kernel local maximum mean discrepancy was used to measure the distribution discrepancy of the learned transferable features relevant subdomains, and the measured distribution discrepancy was added to the objective function for training. Finally, the trained model was used to identify the health conditions of the target domain. The results show that the proposed method may achieve high accuracy in the case of target domain data without label.

Key words: gear fault diagnosis, convolutional neural network(CNN), subdomain adaptation, local maximum mean discrepancy

摘要： 针对生产实际中标注故障数据不足的问题，提出了基于深度学习与子域适配的齿轮故障诊断方法。首先构建域共享的一维卷积神经网络，从故障数据中提取可迁移特征；然后采用多核局部最大均值差异来测量可迁移特征相关子域的分布差异，并将测得的分布差异加入目标函数中训练；最后将训练完成的模型用于目标域健康状态的识别。实验结果表明，所提方法能在无标签目标域数据的情况下得到较高的准确率。

关键词: 齿轮故障诊断, 卷积神经网络, 子域适配, 局部最大均值差异