一种变工况下风电机组行星齿轮箱的故障诊断方法

Fault diagnosis method for wind turbine planetary gearbox under variable working conditions

针对现有的大多数深度迁移学习方法只能在目标转速下工作,而且在模型的训练中总是需要目标域样本的问题,研究风电机组行星齿轮箱在变工况下的故障诊断方法,设计了应用于变工况下行星齿轮箱故障诊断的深度残差半监督域泛化网络,将诊断模型推广到未知转速的故障诊断任务中。首先对振动信号进行Fast Kurtogram时频变换,生成图像并构造样本集;其次模拟实际情况,以含标签源域样本集和无标签源域样本集为输入,使用深度残差网络提取深层故障特征,并引入对抗博弈机制和基于伪标签的半监督学习方法对网络进行训练;最后根据训练后的网络搭建了域泛化故障诊断模型,利用行星齿轮箱故障诊断实验进行评估。实验结果表明,所设计的网络可以有效利用定速样本实现对未知转速样本和变速样本的故障识别,对目标域的平均识别率达到95.24%。

For most existing deep transfer learning methods can only work under the target speed, and they always need target domain samples during the training process, in order to study the fault diagnosis methods of wind turbines planetary gearbox under the variable condition, the deep residual semi-supervised domain generalization network was designed for fault diagnosis of planetary gearboxes with variable speed. This network can generalize the diagnosis model to the fault diagnosis task with unseen speed.Firstly, Fast Kurtogram transformation was carried out on vibration signals to generate images and construct sample sets. Then, in order to simulate the actual situation, the deep residual network was used to extract the deep fault features with a labeled source domain sample was set and an unlabeled source domain sample set as the input, and the antagonistic game mechanism and semi-supervised learning method based on pseudo labels were introduced to the training process. Finally, the domain generalized fault diagnosis model was established according to the trained network. The experimental results show that the designed network can effectively use the fixed speed samples to realize the fault identification of unseen speed samples and variable speed samples, and the average recognition rate of target domain is 95.24%.