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.
维普
万方数据