摘要
针对基于深度学习模型的滚动轴承故障诊断方法易受环境噪声干扰的问题,提出了一种基于拉普拉斯小波滤波(LWF)和自注意力机制-动态选择-卷积神经网络(SA-DS-CNN)的滚动轴承故障诊断模型。首先,提出一种拉普拉斯小波阻尼参数自适应选取策略,使用拉普拉斯小波对采集的滚动轴承振动信号进行相关滤波并进行功率谱变换;其次,基于卷积神经网络框架,引入自注意力机制和动态选择机制,构造SA-DS-CNN;最后,利用SA-DS-CNN提取功率谱特征,根据轴承的不同故障状态定位相关特征信息,实现故障特征的提取和诊断。对N205EM圆柱滚子轴承的故障诊断试验表明:LWF降噪效果较好,能为SA-DS-CNN模型提供优秀的训练样本;SA-DS-CNN模型能抑制无用通道信息,增强网络特征学习能力;LWF和SA-DS-CNN组合模型的故障诊断准确率达到99.65%,优于其他组合模型。
The fault diagnosis methods for rolling bearings based on deep learning model are vulnerable to environmental noise, a fault diagnosis model for rolling bearings is proposed based on Laplace wavelet filtering(LWF)and self attention-dynamic selection-convolution neural network(SA-DS-CNN). Firstly, an adaptive selection strategy of damping parameters for Laplace wavelet is proposed, the collected vibration signals of rolling bearings are denoised by Laplace wavelet correlation filtering method and transformed to power spectrum. Secondly, the convolution neural network(CNN) framework is constructed, and self-attention(SA) and dynamic selection(DS) are introduced to construct SA-DS-CNN. Finally, SA-DS-CNN is imployed to extract the features of power spectrum, and the relevant feature information is located according to different fault states of the bearings to achieve extraction and diagnosis of fault features. Taking N205EM cylindrical roller bearing as an example, the fault diagnosis tests show that: LWF has good denoising effect, providing excellent training samples for SA-DS-CNN model;SA-DS-CNN model suppresses useless channel information and enhances the learning ability of network features;the fault diagnosis accuracy of LWF and SA-DS-CNN combined model is 99.65%, superior to other combined models.
作者
魏亚辉
郭计元
郜帆
WEI Yahui;GUO Jiyuan;GAO Fan(School of Mechanical and Electronic Engineering,Zhumadian Vocational and Teachnical College,Zhumadian 463000,China;Chongqing University,Chongqing 400030,China;State Key Laboratory of Mechanical Transmission,Chongqing 400030,China;Chongqing Wasu Robot Co.,Ltd.,Chongqing 400714,China)
出处
《轴承》
北大核心
2023年第2期89-96,共8页
Bearing
基金
国家自然科学基金资助项目(51775065)。
