基于两种改进RedNet的滚动轴承故障诊断方法研究

Research on the Rolling Bearing Fault Diagnosis Based on Two Improved RedNet

RedNet网络自带的余弦退火算法易使学习率陷入局部极小值,出现拟合现象,导致精度过低。针对此问题,对RedNet进行改进处理,提出了两种MicroNet-RedNet和MobileNetV3-RedNet新型网络。基于RedNet的Involution核思想,用MicroNet网络的微分解卷积和Dynamic Shift-Max动态激活函数对RedNet网络进行改进处理,提出了MicroNet-RedNet新型网络;利用MobileNetV3网络的h-swish激活函数和Squeeze-and-Excitation模块对RedNet进行改进处理,提出MobileNetV3-RedNet新型网络。通过对滚动轴承的实测内圈、外圈和滚动体3种故障的诊断分析可知:所提MicroNet-RedNet和所提MobileNetV3-RedNet可有效地诊断上述故障,诊断精度分别高达98.57%和93.81%,且较传统CNN和原算法RedNet的诊断精度提高很多。

The cosine annealing algorithm of RedNet network is easy to make the learning rate fall into the local minimum and the over-fitting phenomenon occurs,which leads to the low accuracy.In view of the problems,RedNet was improved and two new networks of MicroNet-RedNet and MobileNetV3-RedNet were proposed.Based on the Involution kernel idea of RedNet,the micro-factorized con⁃volution and Dynamic Shift-Max activation function of MicroNet were used to improve RedNet,and thus a new network MicroNet-Red⁃Net was proposed.The h-swish activation function and Squeeze-and-Excitation module of MobileNetV3 were applied to improve Red⁃Net,and thus a new network MobileNetV3-RedNet was proposed.Based on the measured inner ring fault,outer ring fault and rolling ele⁃ment fault of the rolling bearing,it can be seen that the above faults can be diagnosed by the two proposed networks of MicroNet-Red⁃Net and MobileNetV3-RedNet effectively.The accuracies are as high as 98.57%and 93.81%respectively,which are much higher than those got by traditional CNN and the original RedNet.