小样本下基于递归图和迁移学习的轴承故障诊断

Bearing Fault Diagnosis Based on Recursive Graph and Transfer Learning in Small Sample

针对实际工程中故障振动信号数据分布不同、数据量小的问题,提出一种基于卷积神经网络进行迁移学习的滚动轴承诊断方法。利用递归图对滚动轴承的一维时序数据进行图像转换,得到二维图像下的源域数据和目标域数据;将源域数据输入到添加ECA注意力机制的ResNet网络中进行预训练,得到预训练权重;将预训练权重迁移至模型当中,用少量样本进行训练,以验证集准确率为基准,获取此时的训练权重,并保存至目标模型中,最后将测试集数据输入到此时的模型进行验证。结果表明:所提方法能够在目标域仅有少量训练样本的情况下,达到较高的故障识别准确率,且具有较强的鲁棒性能和泛化性能。

To address the problems of different data distributions and small data size of fault vibration signals in practical engineering,a transfer learning method based on convolutional neural network was proposed for rolling bearing diagnosis.The 1D time series data of rolling bearings were transformed into images by using recursive graphs,and the source domain data and target domain data were obtained in the 2D image domain.Then,the source domain data were input into the ResNet network with ECA attention mechanism for pre-training,and the pre-trained weights were obtained.The pre-trained weights were transferred to the model,and a small number of samples were used for training.The validation accuracy was used as the criterion to obtain the training weights at this time,and they were saved to the target model.Finally,the test set data were input into the model at this time for validation.The results show that the proposed method can achieve high fault identification accuracy in the target domain with only a small number of training samples,and it has strong robustness and generalization performance.