【作者】 张明亮；

【导师】 马跃；

【作者基本信息】 大连理工大学 ， 机械工程（专业学位）， 2021， 硕士

【摘要】 滚动轴承作为旋转机械中重要的组成部分,其健康状态直接决定了装备的工作效率和安全性能。由于受到运行环境、载荷冲击等复杂多变的工况影响,滚动轴承的故障频发,因此对滚动轴承的状态进行监测和诊断具有重大的意义。本文以滚动轴承为研究对象,针对当前故障诊断领域中普遍存在的关键性问题开展研究和讨论,结合深度学习理论,提出4种基于卷积神经网络范畴的故障诊断方法。论文的研究内容概括如下:(1)阐述课题的研究背景及意义。分析了滚动轴承的结构、故障类型及其振动信号的特性。对滚动轴承故障诊断的发展历程进行了总结与研究,将其划分为三个阶段,并对每个阶段常用的诊断方法进行了归纳与分析。分别针对当前故障诊断方法中存在的关键性问题提出了不同的解决方案。(2)针对目前传统故障诊断方法对人为先验知识依赖性较大、基于深度学习方法的网络参数较多问题,提出基于敏感分量与MCPG的滚动轴承故障诊断方法。首先使用经验模态分解对振动信号进行处理;而后利用离散Fréchet距离作为衡量指标选取出故障敏感分量作为表征故障状态的数据源;最终利用此数据源完成对多卷积池化组网络模型(MCPG)的训练与测试。通过实验验证,证明该方法具有较好的识别效果。(3)针对基于深度学习的故障诊断方法在小训练样本下较难实现准确识别问题,提出基于时空特征的滚动轴承故障诊断方法。所提模型具有三个主要的功能层,能够兼顾输入数据的时序特征和空间特征。实验结果表明,该方法能够在小训练样本问题上取得较好的识别效果和泛化能力。(4)针对深度学习方法中因所用模型结构单一且输入特征不全面等原因,造成不能实现高效、准确识别问题,提出基于双流CPG网络架构的滚动轴承故障诊断方法。所提模型具有两条特征学习路线,能够兼顾多种类型的样本数据特征,可根据不同需求对路线数量和特征类型进行调整。实验结果表明,该方法具有较好的识别效果、泛化能力和普适性,同时在小训练样本问题上具有较为出色的表现。(5)针对传统卷积神经网络结构较深、搭建网络和调整网络参数较为复杂以及分类器能力有限等问题,提出基于卷积神经网络与麻雀搜索算法优化的支持向量机的故障诊断方法。该方法能够有效解决传统卷积神经网络中最终输出层参数较多以及Softmax分类器能力不足的缺陷,并能够自适应调整支持向量机中的参数,提高分类器的分类能力。实验结果表明,该方法具有较高的识别准确率和较好的普适性。(6)以前文研究作为基础,开发了滚动轴承故障数据分析与智能诊断平台。该平台能够实现对采集数据的读取与分析、对模型的训练与测试等功能。更多还原

【Abstract】 Rolling bearing is an important part of rotating machinery,and its health status directly determine the working efficiency and security performance of the equipment.Due to the running environment,complex and changeable conditions,such as impact load,frequent faults of rolling bearing,so the state of rolling bearing fault monitoring and diagnosis is of great significance.Taking rolling bearings as the research object,this paper studies and discusses the key problems existing in the current fault diagnosis field,and proposes four kinds of rolling bearing fault diagnosis methods based on the category of deep convolutional neural network by combining with deep learning theory.The research content of the paper is summarized as follows:(1)This paper introduces the research background and significance of this topic.The structure,fault types and vibration signal characteristics of rolling bearing are analyzed.The development process of rolling bearing fault diagnosis is summarized and studied,which is divided into three stages,and the common diagnosis methods in each stage are summarized and analyzed.Aiming at the key problems in current fault diagnosis methods,different solutions are proposed.(2)To deal with the problems that the traditional fault diagnosis methods rely on artificial prior knowledge and network model based on the deep learning method need to set many parameters,a fault diagnosis method of rolling bearing based on sensitive component and MCPG is put forward.Firstly,the vibration signal is processed by empirical mode decomposition.Then the discrete Fréchet distance is used as the measurement index to select the fault sensitive component as the data source to represent the fault state.Finally,the training and test of the network model of multi convolutional pool group(MCPG)are completed by using this data source.Experiments prove that this method can very accurately completed for each type of fault identification.(3)In order to solve the problem that deep learning-based fault diagnosis method is difficult to achieve accurate recognition in small training samples,a fault diagnosis method of rolling bearing based on temporal-spatial feature is proposed.The model has three main functional layers,which can take into account the temporal and spatial characteristics of input data.The experimental results show that the method can achieve better recognition effect and generalization ability on small training sample problem.(4)Aiming at the problem that the deep learning method cannot realize efficient and accurate identification due to the single model structure and incomplete input features,a rolling bearing fault diagnosis method based on dual stream CPG network architecture was proposed.The proposed model has two feature learning routes,which can take into account the features of various types of sample data,and the number of routes and feature types can be adjusted according to different needs.The experimental results show that the method has good recognition effect,generalization ability and universality,and has a better performance in small sample problems.(5)In view of the problems of deep structure of traditional convolutional neural network,complex adjustment of network parameters and the limited classifier capability,a fault diagnosis method based on convolutional neural network and sparrow search algorithm optimized SVM is proposed.The method can effectively solve the shortcomings of the traditional convolutional neural network,such as the too many output layer parameters and the insufficient classification ability of the Softmax classifier.Besides it can adaptively adjust the parameters of SVM,and make the classifier has better classification ability.The experimental results show that the method can gain higher recognition accuracy,and has good universality.(6)Based on the previous research,the rolling bearing fault data analysis and intelligent diagnosis platform is developed.It can read and analyze the collected data,train and test the model.更多还原