面向故障诊断的异构特征融合与在线不均衡分类研究

发布时间：2018-07-27 10:16

【摘要】：作为机械设备中最常见的零件之一,滚动轴承的工作状态直接决定了整台设备能否正常工作,甚至关系到整条生产线能否正常运行。滚动轴承诊断技术,可以及时的发现故障,避免造成重大事故,因此,进行轴承诊断的研究具有至关重要的现实意义。传统的信号处理方法常常忽略轴承信号中的重要信息,因此,利用传统故障诊断技术进行分析存在一定缺陷,出现误诊和漏诊现象比较频繁。而且随着科学技术的发展,对故障诊断的要求也越来越高,机器学习越来越多的被应用于故障诊断。本文以滚动轴承为研究对象,针对轴承数据自身所具有的特点以及目前技术存在的缺陷,以机器学习算法为基础理论,展开研究,主要工作内容如下:(1)针对使用单一特征对轴承故障进行诊断时,所含信息具有不确定性,选择的特征无法使用最终选择的算法这一问题,提出了基于异构特征融合的轴承故障检测方法。不同方法提取的异构特征具有相互补充的作用,基于异构特征融合的方法首先将多种方法提取的异构特征并成一个联合特征集,然后把所有的特征基于组特征相关性用多目标粒子群方法将这些特征实现最优分组,保证组内特征间距最小并且组间特征间距最大,最后利用wrapper算法在组的层次上对每组特征进行特征选择,将选择得到的特征作为异构融合的最终特征。该方法以支持向量机为基础算法,对异构特征进行充分合理的融合,并在组的层次上摒弃了特征之间存在的冗余相关性。最后在美国西储大学公布的轴承故障数据和全寿命轴承故障数据上进行仿真实验,证明了该方法的有效性。(2)针对轴承故障数据的在线和类别不均衡的两个特点,提出一种基于主曲线和粒划分的在线不均衡故障诊断方法。算法包括离线和在线两个阶段,在离线阶段,首先构建主曲线,将数据分布分为置信区域和非置信区域,然后通过粒划分,分别对两个区域内的样本进行不同程度的扩充少类和削减多类,在线阶段采用同样的方法处理在线贯序达到的数据块,得到重构后的均衡数据集。该算法在不改变整体数据的分布特征的前提下,有效的减少欠采样过程中多类样本信息的丢失。最终选择用相空间重构的方法提取故障特征,在来自美国西储大学的轴承故障数据和全寿命轴承故障数据上验证了该方法的优势。
[Abstract]:As one of the most common parts in mechanical equipment, the working state of rolling bearings directly determines whether the whole equipment can work normally, or even whether the whole production line can run normally. Rolling bearing diagnosis technology can find fault in time and avoid serious accident. Therefore, the research of bearing diagnosis is of vital practical significance. Traditional signal processing methods often ignore the important information in bearing signals. Therefore, there are some defects in traditional fault diagnosis techniques, and misdiagnosis and missed diagnosis appear frequently. With the development of science and technology, the requirement of fault diagnosis is more and more high, and machine learning is applied to fault diagnosis more and more. This paper takes rolling bearing as the research object, aiming at the characteristics of bearing data itself and the defects of current technology, and taking the machine learning algorithm as the basic theory, the research is carried out. The main work is as follows: (1) when using single feature to diagnose bearing fault, the information contained is uncertain, and the selected feature can not use the algorithm of final selection. A bearing fault detection method based on heterogeneous feature fusion is proposed. The heterogeneous features extracted by different methods are complementary to each other. Firstly, the heterogeneous features extracted by different methods are combined into a joint feature set based on heterogeneous feature fusion. Then all the features are grouped optimally based on the group feature correlation and the multi-objective particle swarm optimization method is used to ensure the minimum feature spacing within the group and the maximum feature spacing among the groups. Finally, the wrapper algorithm is used to select the features of each group at the group level, and the selected features are regarded as the final features of heterogeneous fusion. This method is based on support vector machine (SVM), which can fuse the heterogeneous features fully and reasonably, and abandon the redundant correlation among the features at the group level. Finally, simulation experiments on bearing fault data and life bearing fault data published by the University of Western Reserve in the United States show the effectiveness of this method. (2) aiming at the two characteristics of online and class imbalance of bearing fault data, An online fault diagnosis method based on principal curve and particle partition is proposed. The algorithm consists of offline and online phases. In the off-line phase, the main curve is first constructed, the data distribution is divided into confidence region and disbelief region, and then the distribution is partitioned by grain. The samples in the two regions are expanded and reduced to different degrees. In the online stage, the same method is used to deal with the online sequential data blocks, and the reconstructed equilibrium data sets are obtained. Without changing the distribution characteristics of the whole data, the algorithm can effectively reduce the loss of multi-class sample information in the process of under-sampling. Finally, the method of phase space reconstruction is used to extract the fault features, and the advantages of this method are verified on the bearing fault data from the University of Western Reserve and the full life bearing fault data.
【学位授予单位】：河南师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;TH133.33

【参考文献】