基于多源域异构模型迁移的滚动轴承故障诊断方法

A fault diagnosis method of rolling bearings based on multi-source domain heterogeneous model transfer

针对实际工业生产中不同规格和工况下滚动轴承振动数据分布差异大,多个相似数据集资源利用不充分,导致诊断模型准确率不高的问题,提出一种基于多源域异构模型迁移的滚动轴承故障诊断方法。该方法利用短时傅里叶变换获取滚动轴承振动信号的时频谱图;选择多种不同规格和工况下已知标签数据作为多源域,其他规格和工况下少量已知标签数据作为目标域;使用多个源域数据训练多个ResNet-34深度网络,并提出利用基于进化策略的与模型无关元学习改进异构模型参数迁移策略,使其能够自适应决定迁移到目标域的知识层级及内容;提出将源域知识迁移到VGG-16深度网络得到多个目标域模型后,将其提取的特征首尾相接输入同一个极限学习机中实现特征融合和分类,最终建立滚动轴承故障诊断模型。经试验验证,所提方法可实现不同规格和工况下滚动轴承间的迁移诊断问题,并具有较高的准确率。

For the problems of large distribution differences in rolling bearing vibration data under different specifications and working conditions in actual industrial production,and the resources of multiple similar data sets are not fully utilized,which lead to low accuracy of diagnosis models,a rolling bearing fault diagnosis method based on heterogeneous model transfer in multi-source domain were proposed.The Fourier transform was used to obtain the time-frequency spectra of the rolling bearing vibration signals.A variety of labeled data under different specifications and conditions was used as the multi-source domain,and a small amount of labeled data in other specifications and conditions was used as the target domain.The multiple source domains data were used to train ResNet-34 deep networks,the parameter transfer strategy of heterogeneous model was improved by using evolution strategies model agnostic meta learning,so that it can adaptively determine the level and content of knowledge transferred to the target domain.It was proposed to transfer the source domain knowledge to the VGG-16 deep networks to obtain multiple target domain models,the extracted features were input into the same extreme learning machine head to tail to achieve feature fusion and classification,and finally rolling bearing fault diagnosis model could be established.Experimental results show that the proposed method can achieve the transfer diagnosis between rolling bearings under different specifications and working conditions,and has higher accuracy.