摘要
提出了一种将连续小波变换-对抗训练-卷积神经网络(CWT-AT-CNN)相结合的轴承故障诊断方法,以提高航空轴承故障诊断的准确性和可靠性。基于连续小波变换将轴承原始信号转换为时频图,全面捕捉轴承振动信号的时、频域特征;针对卷积神经网络模型鲁棒性差的问题,使用基于DeepFool算法的对抗训练提高CNN模型的鲁棒性,使其在面对外界攻击时仍能保持较高的准确率。试验结果表明对抗训练可以有效提高模型的可靠性:受到攻击后,CNN模型的故障分类准确率由92.0%降至18.5%,而经过对抗训练的CWT-AT-CNN模型在受到攻击时的故障分类准确率仍可达到85.5%,远高于LeNet-5,CNN-BiLSTM和Involution-CNN等模型。
A bearing fault diagnosis method combining continuous wavelet transform(CWT),adversarial training(AT)and convolutional neural network(CNN)is proposed to improve the accuracy and reliability of fault diagnosis for aviation bearings.The CWT is utilized to transform the raw bearing signals into time-frequency spectrograms,capturing both time and frequency domain features of bearing vibration signals in a comprehensive manner.To address the poor robustness of CNN model,the adversarial training based on DeepFool algorithm is applied to enhance the robustness of CNN model,enabling it to maintain high accuracy when subjected to external attacks.The experimental results demonstrate that the adversarial training significantly improves the reliability of models:the fault classification accuracy of CNN model drops from 92.0%to 18.5%after being attacked,while the adversarially trained CWT-AT-CNN model maintains a fault classification accuracy of 85.5%when attacked,significantly outperforming other models such as LeNet-5,CNN-BiLSTM and Involution-CNN.
作者
方东亮
卓识
战利伟
白晓峰
申立群
FANG Dongliang;ZHUO Shi;ZHAN Liwei;BAI Xiaofeng;SHEN Liqun(AECC Harbin Bearing Co.,Ltd.,Harbin 150025,China;School of Instrumentation Science and Engineering,Harbin Institute of Technology,Harbin 150001,China)
出处
《轴承》
北大核心
2025年第9期101-108,共8页
Bearing
基金
国家科技重大专项资助项目(J2019-IV-0004-0071)
中国航发自主创新专项资助项目(ZZCX-2018-048)。
关键词
滚动轴承
故障诊断
小波变换
深度学习
对抗攻击
rolling bearing
fault diagnosis
wavelet transform
deep learning
adversarial attack
作者简介
卓识(1995-),男,硕士研究生,主要研究方向为航空轴承试验测试技术,E-mail:zhuoshi8125@163.com;通信作者:战利伟(1984-),男,博士研究生,主要研究方向为航空轴承试验测试技术,E-mail:zhanliwei333@163.com。
