基于形态帽积卷积自编码器模型的齿轮箱故障诊断  

Gearbox Fault Diagnosis Research Based on Morphological Hat Product Convolutional Autoencoder

作　　者：黄佳欢 余建波[1,2] HUANG Jiahuan;YU Jianbo(School of Mechanical Engineering,Tongji University,Shanghai 201804;Longmen Laboratory,Luogang 471000)

机构地区：[1]同济大学机械与能源工程学院,上海201804 [2]龙门实验室,洛阳471000

出　　处：《机械工程学报》2023年第23期132-145,共14页Journal of Mechanical Engineering

基　　金：龙门实验室前沿探索课题基金(LMQYTSKT028);国家自然科学基金(71771173);上海市科委“科技创新行动计划”(22N21900100);中央高校基本业务经费(22120220575)资助项目。

摘　　要：在旋转机械设备故障早期,振动信号掺杂着各种噪声以及无效信息难以提取出微弱的故障信号特征。为了解决此问题,提出一个新的深度神经网络模型,形态帽积卷积自编码器模型(Morphological hat product-convolutional auto encoder,MHP-CAE),可有效用于机械设备的故障特征信息提取与识别。首先提出了一个形态滤波帽积层,无缝地嵌入卷积自编码器提取振动信号故障特征。为了克服使用单尺度形态学分析处理信号不全面、不准确的缺陷,采用多尺度形态帽积运算提取故障信号中的脉冲特征。采用基于峭度的方法来融合不同尺度的形态算子提取的特征所包含脉冲成分。最后,进一步利用残差学习连接编码器和解码器,获得良好的振动信号故障特征学习性能。通过单工况和多工况的齿轮箱故障诊断实验表明,MHP-CAE能够以无监督学习的方式对振动信号进行降噪和特征学习,具有良好的去噪和特征学习性能,其特征提取效果要优于一些最新的深度神经网络。In the early stage of equipment fault,the fault signal is mixed with various noises and invalid information,and it is difficult to extract weak fault signal characteristics.In order to solve this problem,a Morphological Hat Product-Convolutional Auto Encoder(MHP-CAE)model is proposed for fault feature information extraction and recognition of mechanical equipment.Firstly,the morphological filter hat product operation is embedded in the convolutional autoencoder as one of the network layers to extract and identify the fault characteristics of the gearbox.In order to overcome the defects of incomplete and inaccurate signal processing using single-scale morphological analysis,multi-scale morphological cap product operation is adopted,in which morphological cap product is mainly used to extract pulse characteristics in fault signals.The kurtosis-based method is used to fuse the pulse components contained in the features extracted by morphological operators of different scales.Then the residual learning is further used to connect the encoder and decoder to obtain good fault feature learning performance.The effectiveness of the method is verified by a gearbox fault diagnosis example.The gearbox fault diagnosis experiments in single and multiple working conditions show that MHP-CAE can perform noise reduction and feature learning of vibration signals in an unsupervised learning manner.The model has good denoising and feature learning performance.Its feature extraction effect is better than some of the latest deep neural networks.

关 键 词：齿轮箱 故障诊断 形态帽积运算 卷积自编码器 残差学习 

分 类 号：TH132[机械工程—机械制造及自动化]