基于NGO-VMD和改进GoogLeNet的齿轮箱故障诊断方法

Gearbox Fault Diagnosis Method Based on NGO-VMD and Improved GoogLeNet

目前的齿轮箱故障诊断方法,在多转速工况及噪声干扰下,存在过拟合及诊断效果不佳的问题。针对此问题,提出一种北方苍鹰(NGO)算法优化变分模态分解(VMD)结合改进GoogLeNet的齿轮箱故障诊断方法。使用NGO对VMD进行参数寻优,利用优化后的VMD去除故障信号中的噪声;对原始GoogLeNet的结构进行合理删减,并利用延迟丢弃法、可训练的ReLU函数(TReLU)对其改进;最后,将去噪后的故障信号转换为二维图作为改进GoogLeNet的输入数据进行网络的训练及分类,得到故障诊断结果。实验结果表明:与其他降噪方法相比,NGO-VMD方法的降噪效果明显,能显著提高故障诊断的准确率;与常见的卷积神经网络相比,提出的改进GoogLeNet能进一步提高故障诊断的准确率,达到了97.2%。

The current fault diagnosis methods of gear box have the problems of overfitting and poor diagnosis effect under multi-speed conditions and noise interference.To solve this problem,a northern goshawk optimization(NGO) algorithm optimized variational mode decomposition(VMD) combined an improved GoogLeNet gearbox fault diagnosis method was proposed.NGO was used to optimize VMD parameters,and the optimized VMD was used to remove noise from fault signals.The structure of the original GoogLeNet was deleted reasonably and improved with delayed dropout and trainable ReLU(TReLU).Finally,the denoised fault signals were converted into 2D graphs as input data of improved GoogLeNet for network training and classification,and fault diagnosis results were obtained.The experimental results show that compared with other noise reduction methods,NGO-VMD method has obvious noise reduction effect and can significantly improve the accuracy of fault diagnosis.Compared with the common convolutional neural network,the improved GoogLeNet can further improve the accuracy of fault diagnosis,reaching 97.2%.