摘要
考虑到卷积神经网络在滚动轴承故障诊断中存在网络结构难以确定、训练次数过多、时间过长等问题,设计了一种贝叶斯优化改进LeNet-5算法,以及采用该算法构建的轴承故障诊断模型。采用贝叶斯优化训练过程中学习率等超参数,多种故障轴承的振动信号直接作为改进LeNet-5网络的输入,对池化输出采用批归一化处理和改进池化层激活函数防止过拟合,利用全局平均池化层替代全连接层提高改进LeNet-5网络的泛化能力,用Softmax分类器实现滚动轴承故障的分类。通过轴承数据库开展实验,实验表明,该算法构建的轴承故障诊断模型在训练集上准确率为99.94%,验证集上的准确率为99.89%,测试集准确率也达到99.65%,与一维卷积神经网络和二维卷积神经网络对比分析,基于贝叶斯优化改进LeNet-5算法构建的轴承故障诊断模型在滚动轴承的故障诊断模型具有更高的准确率,更少的训练次数和训练时间。
Considering that the network structure is difficult to determine,too many training times and too long training time exist in rolling bearing fault diagnosis by convolutional neural network(CNN),a Bayesian optimization algorithm for improving LeNet-5 is designed,and a bearing fault diagnosis model constructed by this algorithm is presented.The learning rate and other hyperparameters in the Bayesian optimization training process are adopted,the original vibration signals of various fault bearings are directly used as the input of the improved LeNet-5 network,batch normalization is adopted for the pooled output and the activation function of the improved pool layer is adopted to prevent over-fitting,and the global average pool layer is used to replace the full connection layer to improve the generalization ability of the improved LeNet-5 network,the fault classification of rolling bearing is realized by softmax classifier.The experimental results show the accuracy of bearing fault diagnosis model constructed by this algorithm training set is 99.94%,the accuracy of verification set is 99.89%,and the accuracy of test set is 99.65%,compared with 1D-CNN and 2D-CNN,bearing fault diagnosis model based on Bayesian optimization and improved LeNet-5 algorithm has higher accuracy,less training times and training time in the fault diagnosis model of rolling bearings.
作者
汤亮
凡焱峰
徐适斐
蔡凯翼
TANG Liang;FAN Yan-feng;XU Shi-fei;CAI Kai-yi(School of Mechanical Engineering,Hubei University of Technology,Wuhan,Hubei 430068,China;Hubei Engineering Research Center for Manufacturing Innovation Method,Wuhan,Hubei 430068,China)
出处
《计量学报》
CSCD
北大核心
2022年第7期913-919,共7页
Acta Metrologica Sinica
基金
国家自然科学基金(61976083)。