【作者】 张越；

【导师】 曲永志；

【作者基本信息】 武汉理工大学 ， 仪器科学与技术， 2019， 硕士

【摘要】 齿轮是传动系统里面常见的关键部件,准确地了解齿轮当前运行状态及故障诊断,对保障齿轮运行可靠性和减少齿轮故障所带来的重大损失具有重大意义。特征提取是故障诊断中的关键步骤,传统的时频域分析方法可以对信号的频谱特征和与机械参数相关的物理特性进行提取,但是需要复杂的信号处理方法和对机械系统的物理特性的了解,这些都需要领域内的专业技术人员进行分析,不便于实现故障检测的自动化和智能化。而随着大数据时代的到来,智能诊断是对机械装备进行快速且准确诊断的必由之路,大部分现有的机器学习方法提取的特征都是抽象的统计特征,无实际物理意义,而监督性学习方法提取到的特征不能直接表征机器的健康和故障状态,必须通过分类器对抽象的特征进行分类,并且所训练的有监督模型很难具有通用性,因此,研究基于智能算法的故障特征分离和提取是亟待解决的重要问题。针对这一问题本文提出采用非监督性深度学习方法用于早期齿轮故障诊断,具体来说,将重点放在不同齿轮故障状态下信号之间的关系分析上,从中提取有用的齿轮故障特征。本文的主要研究内容如下:1.建立了一种基于深度神经网络的深度解纠缠信号特征提取方法。本文基于非监督性深度自编码器提出了一种DTM(Disentangled tone mining)解纠缠特征分解与提取方法:对原始时域信号进行分段傅立叶变换,利用频谱数据进行深度无监督网络的学习,将提取到的初步特征进行基于QR分解的特征选择,得到线性无关的特征组,从而有效的分离和提取故障相关的特征。为了验证DTM方法所提取到的特征是有效的,搭建了齿轮箱模拟故障实验平台并开展了相关的故障实验。同时,将DTM方法与浅层网络PCA、MDS和传统信号处理特征提取方法进行了对比,阐释了深度神经网络的优势,并利用齿轮仿真信号对这种方法的机理做了初步解释。2.本文利用一种深度嵌入聚类算法对嵌入空间特征进行优化,并提出了聚类中心点规则化的方法,间接对嵌入空间样本进行约束,使得提取的特征规律呈现更加明显。本文为了验证这种方法的有效性,开展了齿轮疲劳磨损实验,采集相关数据进行方法试验,结果表明,利用这样一种自学习算法可以起到有效优化特征的作用,使得故障特征区分度更大,反映故障的趋势更加明显。3.本文研究了用深度自编码器方法预测齿轮故障程度的能力,并且本文对基于深度学习的齿轮剩余寿命预测方法进行了初步研究,利用一种基于深度自编码器的半监督学习方法对疲劳磨损实验数据进行了剩余寿命的预测,结果表明,该方法可以对齿轮剩余寿命的趋势进行预测,由于本文中的方法不需要大量的历史训练数据,能够利用有限的历史数据进行在线学习,并进行预测,为在线齿轮故障趋势预测提供一种新的思路。更多还原

【Abstract】 Gear is the common component in transmission system,it is of great significance to control the current operation conditions accurately and carry out gear fault diagnosis for ensuring the reliability of gear operation and reducing the great losses caused by gear fault.Feature extraction is the key step in fault diagnosis.The traditional timefrequency domain analysis methods can extract the spectrum features and the physical characteristics related to mechanical parameters,however,it needs complex signal processing methods and a good understanding of physical characteristics of mechanical systems,so all of the methods require the analysis by professional staff in the field,which is bad to realize automation and intellectualization of fault detection.With the advent of big data,using intelligent method is the inevitable choice to diagnose mechanical equipment quickly and accurately.Most of the features extracted by existing machine learning methods are abstract statistical features,which have no practical physical significance,the features extracted by supervised learning method cannot represent the health or fault condition of the machine directly,but must be classified by classifier.Moreover,the supervised model is difficult to be universal to other type machine.Therefore,it is important to research on the decomposition and extracting of fault features based on intelligent algorithm.To solve this problem,the paper proposes an unsupervised deep learning method for early gear fault diagnosis,specifically,it focuses on the analysis of the relationship between signals in different gear fault states,and extracting useful gear fault features.The main contents of the paper are as follows:1.A feature extraction method for signal disentangled based on deep neural networks is proposed.In the paper,a method called DTM(Disentangled tone mining)for feature decomposition and extraction based on unsupervised deep self-encoding is proposed: firstly,the original time-domain signal is transformed by fast Fourier transform,then the spectrum data is used to learn the deep unsupervised network,and the extracted initial features are selected based on QR decomposition to obtain the linear independent feature matrix,which contains the separated and extracted faultrelated feature.In order to verify the effectiveness of DTM method,an experimental platform for simulating gearbox faults is built and relevant fault experiments are carried out.DTM method is compared with shallow network contains PCA and MDS and a traditional signal processing method,also the advantage of deep neural network is explained,and the mechanism of the method is preliminarily explained by gear simulation signal.2.In the paper,a deep embedded clustering algorithm is used to optimize the embedding features of gear fault,and a regularization of clustering centers method is proposed,which indirectly constrains the embedding samples and makes the extracted feature law more obvious.In order to verify the validity of the method,the natural wear of gear experiments are carried out,and the relevant data are collected.The result shows that the self-learning algorithm can effectively optimize the features,it makes the fault features more distinguishable and reflects the trend of the fault more obvious.3.The paper researches on the ability of predicting gear fault degree using autoencoder,and does the preliminary research on the gear remaining useful life prediction based on deep learning method.A semi-supervised learning method based on autoencoder is used to predict the remaining useful life of gear.The result shows that this method can predict the RUL trend gear system.Because the method in the paper does not need a lot of historical training data,it can use limited historical data to do online prediction,which provides a new idea for online gear fault trend prediction.更多还原