【作者】 罗鹏；

【导师】 杨宇； 黄伟；

【作者基本信息】 湖南大学 ， 机械工程（专业学位）， 2018， 硕士

【摘要】 传统的智能诊断方法一般都是基于“特征提取+分类器”模型,其核心在于特征值的提取以及分类器的设计。针对不同的诊断对象,通常需要根据先验知识提取不同的故障特征值,这必将给最终的诊断结果带来诊断误差;与此同时,传统的分类器一般使用浅层模型,这使得其难以表征信号与装备运行状况之间复杂的映射关系。在这种背景下,HINTON等提出了深度学习(Deep Learning)概念。相对于传统人工智能诊断方法建立的浅层模型,深度学习理论的核心在于模仿生物大脑的思维学习模式,建立深层次学习模型,结合海量的训练样本,有效学习样本数据中隐含的特征,从而建立表征样本信号与诊断对象健康状况之间复杂的映射关系。作为深度学习算法的典型代表,深度卷积神经网络(Deep Convolutional Neural Network,简称DCNN)以及深度信念网络(Deep Belief Network,简称DBN)通过建立深度学习模型,能够自适应提取故障特征及对设备健康状况进行智能识别,避免了信号处理及对专家经验的依赖。本论文在国家自然科学基金项目(51575168)的资助下,对深度学习方法在旋转机械故障诊断中的应用进行了研究。论文主要研究工作如下:(1)探究了变分模态分解(Variational mode decomposition,简称VMD)基本原理。针对VMD在实际应用过程中需要根据先验知识确定惩罚函数和分量分解个数这一缺陷,论文提出了一种改进算法,即广义变分模态分解(Generalized variational mode decomposition,简称GVMD)。该方法减少了人为因素对分解结果造成的主观影响,将信号分解转化为非递归、变分模态分解方式,能够有效分离频率成分相近的谐波分量,能够有效解决信号分离时可能出现的模态混叠问题。仿真信号以及实验信号的分析结果验证了GVMD方法的有效性。(2)DCNN通过建立深度学习模型,避免了对信号处理方法与专家经验的依赖。但是作为一种大数据处理工具,当训练样本数较少时,诊断精度将受到一定影响。而我们在工程实际中能够获取的带标签样本信号数是有限的,为充分利用DCNN自适应提取特征的能力,基于Fisher准则的DCNN算法(Fisher-based Deep Convolutional Neural Network,简称FDCNN)在DCNN基础上引入了类内和类间约束的能量函数模型,能够适用于小样本情况下的特征提取及故障诊断。但是该方法中的模型参数需要人为选择,这必然给诊断结果带来一定影响。针对其缺陷,论文提出了一种自适应Fisher准则的深度卷积神经网络(Adaptive Fisher-based Deep Convolutional Neural Network,简称AFDCNN)方法。该方法首先采用优化算法得到最优能量函数模型参数,继而直接从原始时域信号中自动提取故障特征,最后实现故障诊断。对齿轮故障以及轴承故障仿真和实验信号的分析结果表明,该方法可以有效地实现小样本下的旋转机械故障智能诊断。(3)作为深度学习算法典型代表之一的DBN,通过对原始样本数据逐层学习发现其数据分布式特征,从而完成样本识别。与此同时,在将获取的振动信号输入构建的DBN模型时无需信号重构,因而已受到越来越多研究者的关注。但是DBN依然存在网络结构需要人为设定的缺陷,这也限制了DBN在工程实际中的应用。为解决DBN网络结构难以确定及如何提升其在工程实际应用中的诊断效率问题,论文提出了一种新的深度信念网络,即结构自适应深度信念网络(Structure Adaptive Deep Belief Network,简称SADBN)。与DBN相比,SADBN以网络结构熵误差以及计算成本为考量指标,从而确定优化后的网络结构,有效提高诊断精度及效率。对齿轮故障以及滚动轴承故障振动信号的分析结果表明了改进网络的有效性。(4)针对旋转机械的复合故障特征提取困难这一现状,本论文将AFDCNN以及SADBN分别引入旋转机械的复合故障诊断之中,提出基于AFDCNN方法的旋转机械复合故障诊断方法以及基于SADBN方法的旋转机械复合故障诊断方法。以齿轮箱为诊断对象,分别根据AFDCNN及SADBN方法设计相对应的仿真和实验模型,通过仿真与实验分析,表明两种方法能够满足要求。更多还原

【Abstract】 The traditional intelligent diagnosis method is based on the "feature extraction add classifier" model,the core of the method are the extraction of feature value and the design of classifier.For the different diagnosis objects,it is usually necessary to extract different fault eigenvalues according to the prior knowledge,which will inevitably bring diagnostic errors to the final diagnosis results.At the same time,the traditional classifier uses the shallow models generally.This makes it difficult to characterize the complex mapping relationship between the signal and the equipment operation.In this context,Hinton and his colleagues put forward the concept of deep learning.Compared with the shallow model established by the traditional artificial intelligence diagnosis method,the deep learning theory is devoted to learning the hidden features of the data by simulating the learning process of the brain,establishing the deep learning model and combining the massive training data.Thus the complex mapping relationship between signal and health status is established.As a typical representative of the deep learning algorithm,the deep convolution neural network(DCNN)and the deep belief network break away from the dependence on the signal processing technology and the expert knowledge by establishing the deep learning model.Thus,the adaptive extraction of fault features and the intelligent diagnosis of health condition can be completed.Supported by National Natural Science Foundation(No.51575168),the deep learning method is introduced to the fault diagnosis of rotating machinery in this dissertation.The main research contents of this dissertation are listed as follows:(1)The application of VMD in fault diagnosis of rotating machinery is discussed,and the variational mode decomposition method is studied emphatically.In order to solve the problem that it is necessary to determine the number of penalty function and component decomposition according to prior knowledge in practical application,an improved algorithm,generalized variational mode decomposition(GVMD),is proposed in this dissertation.The subjective influence of human factors on the decomposition results can be reduced in this method,and the signal can be resolved into non-recursive and variational mode,which can separate harmonic components with similar frequency components effectively.And it has good robustness to the signal with low signal-to-noise ratio(SNR).The effectiveness of the proposed method verified by the results of simulation and experimental signal analysis.(2)By establishing a deep learning model,DCNN gets rid of the dependence on signal processing technology and expert knowledge.But,as a big data processing tool,the diagnostic accuracy will be affected when the number of training samples is small.However,the number of labeled sample signals we can obtain in engineering practice is limited.In order to make full use of the ability of DC NN adaptive feature extraction,on the basis of DCNN,the energy function model of intra-class and inter-class constraints are introduced by the deep convolution neural network recognition algorithm based on Fisher criterion(FDCNN).It can be applied to feature extraction and fault diagnosis in small samples.However,the model parameters in this method need to be selected artificially,which will inevitably affect the diagnosis results.In this dissertation,an adaptive Fisher criterion for deep convolution neural networks(AFDCNN)is proposed.In this method,the parameters of the optimal energy function model are obtained by using the optimization algorithm,and then the fault features are automatically extracted from the original time domain signal,and finally the fault diagnosis is realized.The analysis of gear fault signals and bearing fault signals show that this method can effectively realize the intelligent and quantitative fault diagnosis of rotating machinery with small samples.(3)As one of the typical representatives of the deep learning algorithm,the deep belief network can form abstract high-level representation by combining the lower level features,and then the distributed features of the data can be discovered.At the same time,there is no need for signal reconstruction when the obtained vibration signal is input into the DBN model.As a result,more and more researchers have paid their attention to it.However,DBN still has the defect that the network structure needs to be set artificially,which limits the application of DBN in engineering practice.In order to solve the problem of how to determine the structure of DBN network and how to improve its diagnostic efficiency in engineering practice,a new kind of depth belief network,namely structural adaptive deep belief network,is proposed in this dissertation.Compared with DBN,SADBN takes the entropy error of network structure and calculation cost as the index to determine the optimized network structure and improve the efficiency of diagnosis.The results of vibration signal analysis of gear fault and rolling bearing fault show that the improved network is effective.(4)In order to solve the problem of complex fault feature extraction of rotating machinery,the AFDCNN and SADBN methods are introduced into the complex fault diagnosis of rotating machinery in this dissertation.The method of complex fault diagnosis of rotating machinery based on AFDCNN method and the method of complex fault diagnosis of rotating machinery based on SADBN method are presented.Based on the AFDCNN and SADBN methods,the corresponding si mulation and experimental models are designed for the gearbox.The effectiveness of the two methods in the complex fault diagnosis of rotating machinery is verified by simulation and experimental analysis.更多还原