基于卷积神经网络的滚动轴承故障诊断研究综述

Review of rolling bearing fault diagnosis based on convolutional neural network

随着机器学习技术的兴起,深度学习被用于故障诊断领域并得到迅速发展,其中,卷积神经网络是具有出色特征提取能力的深度学习模型,因其适用于处理图像数据和高维数据而成为故障诊断研究的热点。针对传统故障诊断方法难以解决轴承振动信号存在的特征提取困难和信号噪声污染的问题,为高效、准确地完成滚动轴承故障诊断工作,首先,对卷积神经网络的结构进行了简单介绍,并研究了近年来经典卷积神经网络模型用于滚动轴承故障诊断的重要进展;然后,从深度特征提取、超参数调整和网络结构优化等角度,对各种优化卷积神经网络的方法原理进行了简单介绍,详细探讨了将卷积神经网络应用于滚动轴承故障诊断的优化途径和已经取得的研究进展;最后,对几种典型优化方法的优势与不足进行了比较,并对不同角度优化卷积神经网络的途径进行了总结。研究结果表明:基于卷积神经网络的滚动轴承故障诊断方法还需要解决数据不平衡、模型特征提取能力不足和泛化性不强的问题,后续研究工作应聚焦于多源数据融合、模型性能优化以及多方技术结合等方向。

With the rise of machine learning technology,deep learning(DL)was utilized in the field of fault diagnosis and underwent rapid development.Among these developments,the convolutional neural network was a deep learning model with excellent feature extraction ability,which had become a hot spot in fault diagnosis research because it was suitable for processing image data and high-dimensional data.Aiming at the problem of the traditional fault diagnosis methods in addressing the difficulties of feature extraction from bearing vibration signals and the contamination of signals by noise,to efficiently and accurately accomplish the fault diagnosis of rolling bearings,firstly,the structure of convolutional neural network(CNN)was briefly introduced,and the important progress of classical convolutional neural network model for rolling bearing fault diagnosis in recent years was studied.Then,from the perspectives of deep feature extraction,hyperparameter adjustment and network structure optimization,various methods for optimizing convolutional neural networks were briefly introduced,and the optimization methods of applying convolutional neural networks to rolling bearing fault diagnosis and the research progress made were discussed in detail.Finally,the advantages and disadvantages of several typical optimization methods were compared,and the ways to optimize convolutional neural networks from different angles were summarized.The results show that the rolling bearing fault diagnosis method based on convolutional neural network also needs to solve the problems of data imbalance,insufficient model feature extraction ability and weak generalization,and the follow-up research work should focus on multi-source data fusion,model performance optimization and multi-party technology combination.