Abstract: In view of the complex and changeable conditions of industrial sites, it is prone to cause data missing or imbalance, resulting in low accuracy of the model. In this paper, a bearing fault diagnosis method with improved knowledge distillation and transfer learning is proposed. Firstly, the teacher-student model adopts dense convolutional neural network and artificial neural network to adaptively and randomly extract key feature information of source domain samples and target domain samples. Secondly, the soft label loss and hard label loss of the corresponding domain are obtained, and hierarchical transfer learning is introduced to improve the difference of conditional distribution of domain samples to obtain the final distillation loss function, and the distilled “dark knowledge” is feedback to update the student model. Finally, the semi-supervised fault transference decision of the intelligences is implemented using the target domain test samples. The experimental results show that the student model is able to learn various properties from the teacher model and improve the diagnostic accuracy of the simple model, and that the method has superior accuracy and robustness compared to other methods.

Key words: bearing, knowledge distillation, transfer leaning, fault diagnosis

摘要： 针对工业现场工况复杂多变，易造成样本缺失或不平衡，导致模型诊断准确率低等问题。提出改进知识蒸馏与迁移学习的轴承故障诊断方法。教师-学生模型分别采用稠密卷积神经网络和人工神经网络，自适应随机提取源域样本和目标域样本的关键特征信息；获取对应领域的软标签损失和硬标签损失，引入分层迁移学习改善领域样本的条件分布差异，获取最终蒸馏损失函数，并将蒸馏后的“暗知识”反馈更新学生模型；利用目标域测试样本实现智能体的半监督故障迁移决策。实验结果表明，学生模型能够从教师模型学习到各项性能，提升简单模型的诊断精度，相较于其他方法，该方法具备较优异的准确性和鲁棒性。

关键词: 轴承, 知识蒸馏, 迁移学习, 故障诊断