多域信息融合结合改进残差密集网络的轴承故障诊断  

Bearing fault diagnosis based on information fusion and improved residual dense networks

作　　者：袁彩艳 孙洁娣[1,2] 温江涛[3] 时培明[3] 闫盛楠[1,2] YUAN Caiyan;SUN Jiedi;WEN Jiangtao;SHI Peiming;YAN Shengnan(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China;Hebei Key Laboratory of Information Transmission and Signal Processing,Yanshan University,Qinhuangdao 066004,China;Key Laboratory of Measurement Technology and Instrumentation of Hebei Province,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]燕山大学河北省信息传输与信号处理重点实验室,河北秦皇岛066004 [3]燕山大学河北省测试计量技术及仪器重点实验室,河北秦皇岛066004

出　　处：《振动与冲击》2022年第4期200-208,252,共10页Journal of Vibration and Shock

基　　金：国家自然科学基金(61973262,62073282);河北省自然科学基金(E2020203061);河北省引进留学人员项目资助(C201827);河北省高等学校科学技术研究项目(QN2019133)。

摘　　要：针对滚动轴承原始时域信号信息单一、深度卷积神经网络提取的特征对信息的传递存在差异等问题,该研究提出了一种多域信息融合与改进残差密集网络的轴承故障诊断方法。为了获取故障的多方面信息,先对原始数据进行多域变换,再将融合信息输入经卷积注意力改进的残差密集网络进行深度学习。经注意力机制改进的网络能够实现对提取特征的重要性区分,提高网络的训练速度、改善识别准确率。试验结果及对比分析表明该算法可以提取较为全面的特征,较传统方法具有更好的识别效果。Aiming at the problem that the original time signal of a rolling bearing is relatively simple and the features extracted by convolution neural network are different during useful information transmission,a bearing fault diagnosis method was proposed based on multi-domain information fusion and an improved residual dense network.In order to obtain the multifaceted information of fault signal,the original data are transformed in multiple domains.Then the multi-domain information is input into the residual dense network improved by convolution attention mechanism for deep learning.Discrimination is realized according to the importance of the extracted features.The training speed and efficiency of the neural network were improved.Experimental results and analysis show that the proposed method can extract more comprehensive features and has higher recognition accuracy than traditional methods.

关 键 词：轴承故障诊断 残差网络 多变换域处理 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TH865[自动化与计算机技术—计算机科学与技术]