基于多尺度特征融合残差神经网络的旋转机械故障诊断  

Fault diagnosis of rotating machinery based on residual neural network with multi-scale feature fusion

作　　者：邓飞跃[1,2,3] 丁浩 郝如江 DENG Feiyue;DING Hao;HAO Rujiang(State Key Laboratory of Mechanical Behavior in Traffic Engineering Structure and System Safety,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;Hebei Province Key Laboratory of Mechanical Power and Transmission Control,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;School of Mechanical Engineering,Shijiazhuang Tiedao University,Shijiazhuang 050043,China)

机构地区：[1]石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,石家庄050043 [2]石家庄铁道大学河北省工程机械动力与传动控制重点实验室,石家庄050043 [3]石家庄铁道大学机械工程学院,石家庄050043

出　　处：《振动与冲击》2021年第24期22-28,35,共8页Journal of Vibration and Shock

基　　金：国家自然科学基金(11802184,11790282);河北省自然科学基金(E2019210049);北京市重点实验室研究基金资助课题(PGU2020K009);河北省‘三三三人才工程’资助项目(A202101017)。

摘　　要：轴承、齿轮等旋转部件常在复杂工况下运行,环境噪声干扰大,导致故障特征微弱而难以准确诊断。基于此,该研究提出一种新的多尺度特征融合残差块(multi-scale feature fusion residual block,MSFFRB)设计方法,基于此构建了一维残差神经网络用于旋转机械故障诊断。该模型能够将不同尺度的网络卷积层级联在一起提取多尺度特征信息,在残差块内部实现了多尺度特征信息的有效融合,兼顾了残差网络跨层恒等映射与多尺度特征提取的优势,克服了传统卷积操作只能提取单一尺度特征信息的缺点。所构建的残差神经网络可以直接输入样本数据,不需要进行任何数据预处理,而且模型结构具有较高的灵活性,易于扩展。试验分析表明,所提网络可有效用于旋转机械的故障诊断,相比传统CNNs、ResNets、1D-LeNets、1D-AlexNets、MC-CNNs等5种当前常用网络,具有更好的抗噪性能,故障分类准确率更高,这为旋转机械故障诊断提供了一种新的途径。Aiming at the problem that weak fault feature of rotating machinery,such as rolling bearing and gear,is difficult to detect under strong background noise,a novel design method of multi-scale feature fusion residual block(MSFFRB)was proposed,and a one-dimensional residual neural network was developed to diagnose fault of rotating machinery.In the proposed MSFFRB,some convolutional layers with different scales are cascaded together to extract multi-scale feature information,and the multi-scale feature information is effectively fused.It simultaneously takes into account the advantages of crossing layer identity mapping and the multi-scale feature extraction,and overcomes the disadvantage that traditional convolution layer with fixed scale only extracts single scale feature information.The proposed network can input data directly without any data preprocessing.Moreover,the architecture of the network has high flexibility and is easy to further expand.Experimental results show that the method can be effectively used for fault diagnosis of rotating machinery.Compared to the traditional CNNs,ResNets,1D-LeNets,1D-AlexNets,and MC-CNNs,the proposed method has better anti-noise performance and higher classification accuracy,which provides a new solution for rotating machinery fault diagnosis.

关 键 词：旋转机械 故障诊断 残差神经网络 多尺度特征融合 

分 类 号：TH17[机械工程—机械制造及自动化]