【作者】 王博；

【导师】 王亚萍；

【作者基本信息】 哈尔滨理工大学 ， 机械设计及理论， 2022， 硕士

【摘要】 行星齿轮箱因其结构紧凑,可以实现大传动比、大功率的减速运动,在精密传动与重型机械等领域有着广泛的应用。当行星齿轮箱内部结构出现故障时,将会严重影响设备系统的稳定运行与安全工作,严重时将会造成重大的财产损失与人员伤亡。因此,研究行星齿轮箱故障诊断方法,准确且快速的识别故障类型,对于设备的安全运行具有重大的意义。本文以行星齿轮箱作为研究对象,针对故障诊断领域关键问题,包括信号降噪、特征降维、故障诊断三个方面展开研究。首先,针对稀疏表示类信号降噪算法自适应降噪能力有限,以及深度学习类降噪算法缺乏理论解释性,提出基于K-奇异值分解(K-Singular Value Decomposition,K-SVD)自编码器信号降噪方法。根据行星齿轮箱振动信号形态特点,利用傅里叶原子构建初始字典;改进K-SVD降噪方法网络结构,搭建K-SVD自编码器来提高算法降噪能力;对输入数据采用分批次降噪处理提高计算效率。输入行星齿轮局部故障仿真信号降噪,探究网络层数与初始化字典方式对K-SVD自编码器的影响,对比正交匹配追踪降噪方法与K-SVD降噪方法效果证明所提降噪方法有效性与优越性。其次,针对传统特征降维方法未充分挖掘数据间信息,使用拉普拉斯特征映射方法(Laplace Eigenmapping,LE)进行特征降维。提取K-SVD自编码器降噪后信号的时域频域特征;通过构建拉普拉斯特征映射目标函数,求解获取降维后特征矩阵。探究核参数对LE降维方法影响,并与主成分分析(principal component analysis,PCA)、核主成分分析(Kernel Principal Component Analysis,KPCA)降维方法进行对比,证明本文所使用LE降维方法有效性与优越性。再次,针对目前故障诊断方法需要大量利用有故障标签数据进行监督式学习,在故障标签获取上耗费大量人力物力成本问题,提出一种基于半监督-图卷积(Semi-supervised Graph Convolutional neural Network,Semi-GCN)行星齿轮箱故障诊断方法。将经过LE特征降维后数据集利用欧氏距离方法构建其无向图,根据谱图理论构建Semi-GCN层单元,并建立Semi-GCN半监督故障诊断模型。探讨Semi-GCN数据标签率与卷积核尺寸对Semi-CGN影响;与BP神经网络和卷积神经网络故障诊断模型对比诊断效果,证明Semi-GCN的有效性与优越性。最后,为验证本文所提方法对于实际数据有效性,搭建行星齿轮箱故障诊断试验台,依次对K-SVD自编码器降噪方法、LE特征降维方法与Semi-GCN故障诊断方法做实验验证,证明本文所提出方法的有效性与优越性。更多还原

【Abstract】 Planetary gearbox is widely used in precision transmission and heavy machinery due to its compact structure,which can achieve high transmission ratio and high power deceleration motion.When the internal structure of the planetary gear box fails,it will seriously affect the stable operation and safe work of the equipment system,and will cause significant property losses and casualties in serious cases.Therefore,it is of great significance to study the fault diagnosis method of planetary gearbox and identify the fault type accurately and quickly for the safe operation of equipment.This thesis takes planetary gearbox as the research object,aiming at the key problems in the field of fault diagnosis,including signal noise reduction,feature dimension reduction and fault diagnosis.Firstly,aiming at the limited adaptive denoising ability of sparse representation signal denoising algorithm and the lack of theoretical interpretation of deep learning signal denoising algorithm,an autoencoder signal denoising method based on KSingular Value Decomposition(K-SVD)is proposed.According to the morphological characteristics of vibration signals of planetary gearboxes,Fourier atoms are used to construct the initial dictionary.The network structure of K-SVD denoising method was improved,and the self-encoder of K-SVD was built to improve the denoising ability.Batch denoising is applied to the input data set to improve computational efficiency.The influence of network layer number and initialization dictionary mode on k-SVD autoencoder was investigated by inputting the simulation signal of planetary gear local fault to denoise.The effectiveness and superiority of the proposed algorithm are proved by comparing the noise reduction effect of orthogonal matching pursuit and K-SVD method.Secondly,as the traditional eigendimension reduction methods do not fully mine the information between data,the Laplace Eigenmapping(LE)is used for eigendimension reduction.The time-domain and frequency-domain characteristics of k-SVD autoencoder after denoising were extracted.By constructing the objective function of Laplacian feature mapping,the eigenmatrix after dimension reduction is obtained.Explore the influence of Kernel parameters on LE dimension reduction method and compare it with principal Component Analysis(PCA)and Kernel Principal Component Analysis(KPCA)dimension reduction methods.The effectiveness and superiority of LE dimension reduction method used in this paper are proved.Thirdly,a semi-supervised Graph Convolutional neural Network based on Convolutional neural Network is proposed to solve the problem that current fault diagnosis methods require a great deal of supervised learning with fault label data,and the acquisition of fault labels costs a great deal of manpower and material resources.Semi-GCN planetary gearbox fault diagnosis method.After dimension reduction of LE feature,undirected graph was constructed by Euclidean distance method,semiGCN layer unit was constructed according to spectral theory,and semi-GCN semisupervised fault diagnosis model was established.The influence of semi-GCN data label rate and convolution kernel size on semi-CGN is discussed.Compared with BP neural network and convolutional neural network fault diagnosis models,the effectiveness and superiority of semi-GCN is proved.Finally,in order to verify the validity of the proposed method for actual data,a planetary gearbox fault diagnosis test rig was built,and the k-SVD autoencoder noise reduction method,LE feature dimension reduction method and semi-GCN fault diagnosis method were experimentally verified,which proved the validity and the Superiority of the proposed method.更多还原