摘要
针对旋转机械故障特征需要人工提取、复杂故障识别困难和诊断模型鲁棒性差的问题,在经典卷积神经网络Alex Net基础上,提出基于一维深度卷积神经网络的故障诊断模型,模型采用改进的一维卷积核和池化层以适应一维时域信号。相比传统智能诊断模型的人工特征提取和故障分类两阶段模式,该模型将两者合二为一:首先利用多个交替的卷积层和池化层完成原始信号自适应特征学习,然后结合全连接层实现故障诊断。通过轴承和齿轮箱健康状态监测实验表明,提出了模型可以实现高精度、稳定和快速的故障诊断,并与BP神经网络、SVM、一维Le Net5模型和经典Alex Net模型对比,证明了提出模型的优势,最后通过PCA可视化分析说明模型在特征提取上的有效性。
Aiming at problems of rotating machinery’s fault features needing to be extracted manually,complex fault recognition being difficult and diagnosis model’s poor robustness,a novel1D depth convolutional neural network-based rotating machinery fault diagnosis model was proposed based on the classical convolution neural network model AlexNet.This new model adopted the modified1D convolutional kernel and pool layers to adapt1D time domain signals.The traditional intelligent diagnosis model included two distinct modules of manual feature extraction and classification,and the proposed model combined these two modules into one.With the proposed model,multiple alternate convolution and pool layers were used to complete learning the original signal’s self-adaptive features and then all connected layers were combined to realize fault diagnosis.Bearings and gearboxes health monitoring tests showed that the proposed model can realize accurate,stable and fast fault diagnosis;compared to BP neural network,SVM,1D-LeNet5model and the classical AlexNet model,this new model is the best;the feature extraction effectiveness of the proposed model is verified with the PCA visualized analysis.
作者
周奇才
刘星辰
赵炯
沈鹤鸿
熊肖磊
ZHOU Qicai;LIU Xingchen;ZHAO Jiong;SHEN Hehong;XIONG Xiaolei(School of Mechanical Engineering,Tongji University,Shanghai201804,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2018年第23期31-37,共7页
Journal of Vibration and Shock
关键词
深度学习
卷积神经网络
特征学习
智能诊断
旋转机械
deeply learning
convolutional neural network
feature learning
intelligent diagnosis
rotating machinery
作者简介
第一作者,周奇才,男,教授,博士生导师,1962年生。
