基于深度残差网络算法的减速器故障诊断研究

Research on Retarder Fault Diagnosis Based on Deep Residual Network Algorithm

为了进一步提高对机床不同故障的分类准确率,设计一种深度残差网络。通过对机床振动测试试验台信号预处理分析,优化网络结果并进行故障诊断对比分析。研究结果表明:训练集处理可以使准确率收敛至100%,表明模型没有发生欠拟合情况;测试准确率达到了98.2%以上,表现出非常有益的泛化效果。当行数比列数更少时,随着两者差异的增加,模型的分类准确率显著降低;行数超过列数后,模型达到了更高的分类准确率并保持相对稳定的状态。泛用性验证表明,采用残差网络模型进行滚动轴承信号分类时也可以获得99.51%的准确率。CNN网络比浅层模型表现出了更强的识别性能。ShortCut结构具备明显优越性,有助于网络具备更强识别能力。

In order to improve the classification accuracy of different faults of machine tools,a deep residual network is designed.Through the signal preprocessing analysis of the machine tool vibration test bench,the network results are optimized and the fault diagnosis is compared and analyzed.The results show that the training set processing can make the accuracy converge to 100%,indicating that the model does not underfit.The test accuracy reaches more than 98.2%,showing a very beneficial generalization effect.When the number of rows is smaller than the of columns,the classification accuracy of the model decreases significantly with the increase of the difference between them.When the number of rows exceeds the one of columns,the model achieves higher classification accuracy and remains relatively stable.The universality verification shows that the residual network model can also achieve 99.51%accuracy in the classification of rolling bearing signals.The CNN network shows stronger recognition performance than the shallow model ResNet.ShortCut structure has obvious advantages,which helps the network to have stronger identification capability.