基于改进的SqueezeNet直升机滚动轴承故障诊断  

Fault diagnosis of helicopter rolling bearing based on improved SqueezeNet

作　　者：余志锋 熊邦书[1] 李新民[2] 欧巧凤[1] YU Zhifeng;XIONG Bangshu;LI Xinmin;OU Qiaofeng(Provincial Key Laboratory of Image Processing and Pattern Recognition,Nanchang Hangkong University,Nanchang 330063,China;Science and Technology on Rotorcraft Aeromechanics Laboratory,China Helicopter Research and Development Institute,Aviation Industry Corporation of China,Limited,Jingdezhen Jiangxi 333001,China)

机构地区：[1]南昌航空大学图像处理与模式识别省重点实验室,南昌330063 [2]中国航空工业集团有限公司中国直升机设计研究所直升机旋翼动力学重点实验室,江西景德镇333001

出　　处：《航空动力学报》2022年第6期1162-1170,共9页Journal of Aerospace Power

基　　金：国家自然科学基金(61866027);航空科学基金(2016ZD56008;20185756006)。

摘　　要：针对现有基于卷积神经网络的故障诊断方法存在模型参数量和计算量大的问题,提出一种改进的SqueezeNet模型应用于直升机滚动轴承故障诊断。该模型借鉴VGG16模型的思想,在经典的SqueezeNet基础上,采用3个3×3卷积核代替1个7×7卷积核,实现了在相同感知野条件下增加网络容量、增强非线性、减少网络参数量,采用卷积层、池化层和Fire模块、池化层两大结构交替的方式组成模型特征提取层,在保障特征提取能力的情况下,进一步减少了网络参数量。通过轴承数据开展模型故障诊断实验,结果表明该模型诊断精度达到99.65%,与传统卷积神经网络及经典的SqueezeNet模型相比诊断精度相当,而计算量与参数量最大缩减约6倍和36倍。In order to solve the problem of large amount of model parameters and calculations in the existing fault diagnosis methods based on convolutional neural networks, an improved SqueezeNet model was proposed to be applied to the fault diagnosis of helicopter rolling bearings. By drawing on the idea of the VGG16 model based on the classic SqueezeNet,the model used three sizes of 3 × 3 convolution kernels instead of one size of 7 × 7 convolution kernel,and realizes the increase of network capacity,enhancement of nonlinearity,reduction of network parameters amount under the same perceptual field conditions. To further reduce the amount of network parameters, the convolutional layer, pooling layer, Fire module and pooling layer were alternated to form the model feature extraction layer. While ensuring the feature extraction capability, the amount of network parameters was further reduced. The model faults diagnosis experiment was carried out through the bearing data onto the research group. The results showed that the diagnosis accuracy of the model reached 99. 65%,which was comparable to the traditional convolutional neural network and the classic SqueezeNet model. The calculation amount and the parameter amount were reduced by about 6 times and 36 times.

关 键 词：直升机滚动轴承 卷积神经网络 VGG16模型 轻量化 SqueezeNet模型 

分 类 号：V2335[航空宇航科学与技术—航空宇航推进理论与工程] TH133.33[机械工程—机械制造及自动化]