基于频率通道注意力机制和DenseNet45模型的齿轮故障诊断方法  

Gear Fault Diagnosis Method Based on Frequency Channel Attention Mechanism and DenseNet45 Model

作　　者：陶迎雪 杜艳平 窦水海 王兆华 白慧娟 孙兆永 TAO Ying-xue;DU Yan-ping;DOU Shui-hai;WANG Zhao-hua;BAI Hui-juan;SUN Zhao-yong(School of Mechanical and Electrical Engineering,Beijing Institute of Graphic Communication,Beijing 102600,China;School of Basic Education,Beijing Institute of Graphic Communication,Beijing 102600,China)

机构地区：[1]北京印刷学院机电工程学院,北京102600 [2]北京印刷学院基础教育学院,北京102600

出　　处：《印刷与数字媒体技术研究》2023年第6期38-48,共11页Printing and Digital Media Technology Study

基　　金：北京市教育委员会科技/社科计划项目--面向印品质量的高端印刷装备复杂传动系统动态特性研究及优化设计(No.KZ202210015019)。

摘　　要：针对齿轮信号具有较强非平稳性、易被强烈噪声干扰及不同故障类型间易混淆的问题,本研究提出了融入频率通道注意力(Frequency Channel Attention,FCA)机制和DenseNet45模型的齿轮故障诊断方法,实现了对故障类型的精确识别和分类。首先,对齿轮信号进行短时傅里叶变换(Short-time Fourier Transform,STFT),得到二维时频谱图作为样本,将样本按照4:1划分为训练集和测试集;然后调用更少的密集连接块构造DenseNet45模型,同时将FCA模块融入DenseNet45模型的卷积层中,并将训练集输入到模型中进行学习,通过改变学习率下降的速度来提升模型性能;最后,将学习好参数的网络模型应用于测试集进行验证,输出测试集故障识别的准确率。结果表明,所提方法在两种工况下的准确率分别达到99.875%和99.75%,能有效将不同类型的故障进行识别并分类,与其他典型卷积神经网络模型相比,所提方法准确率更高,收敛速度更快。Aiming at the problems that the gear signal has strong non-stationarity,is easy to be disturbed by strong noise and easily confused between different fault types,in this study,a gear fault diagnosis method was proposed based on DenseNet45 model incorporating Frequency Channel Attention(FCA)mechanism,which realized accurate identification and classification of fault types.Firstly,the Short-time Fourier Transform(STFT)on the gear signal was performed and the two-dimensional time-frequency spectrum map was obtained as a sample.The sample was divided into a training set and a test set according to 4:1.Then,the DenseNet45 model was constructed by calling fewer dense connected blocks.At the same time,FCA module was integrated into the convolution layer of the DenseNet45 model,and the training set was taken as the input of the model for learning.The model performance was improved by changing the learning rate decline speed.Finally,the network model with learned parameter was applied to the test set to output the accuracy of the test set fault recognition.The results showed that the accuracy of the proposed method can reach 99.875%and 99.75%respectively under two working conditions,and it can effectively identify and classify different types of faults.Compared with other convolutional neural network models,the proposed method has higher accuracy and faster convergence.

关 键 词：齿轮 FCA模块 STFT 故障诊断 DenseNet模型 

分 类 号：TS803.9[轻工技术与工程] TH17[机械工程—机械制造及自动化]