基于机器学习的滚动轴承智能诊断方法研究

发布时间：2018-03-27 15:10

本文选题：轴承　切入点：故障诊断　出处：《江西理工大学》2015年硕士论文

【摘要】：机械设备在现代化工业中占据着相当关键的地位,直接影响企业中工业生产的正常运行。当机械设备出现早期故障时,需要对它及时修理,以免造成更大伤害,这就要求在技术上能够诊断出早期的轻微故障。机械设备的诊断过程通常分为三步:第一步是诊断信息获取;第二步是特征提取,即从第一步获取的信息中挖掘出设备的故障信息;第三步是状态识别和故障诊断。其中特征提取是诊断过程中关键的环节,它直接影响诊断结果的准确性,关系到诊断系统能否发现设备潜在的故障。因此,如何提取最优的低维故障特征从而提高故障分类性能是目前面临的一个巨大的挑战。状态识别阶段即使用模式识别、机器学习方法识别样本,使用不同的识别方法某种程度上会影响识别精度。本论文以机器学习算法为基础,以滚动轴承为对象,研究运用不同的机器学习方法对其特征提取与诊断识别。复杂故障设备多个特征参数的参数之间存在冗余性或不相关性,不利于故障诊断,而经典的线性降维方法不能满足非线性故障数据的降维要求。针对这个问题,本文将低秩鉴别投影的特征提取方法与稀疏表示分类器结合起来构成故障诊断模型。为了寻求对应数据的最优参数,实现参数自适应,引入网格搜索算法。经实例仿真表明,低秩鉴别投影算法能够比较准确地描述数据的全局结构和判别结构,将模型应用到滚动轴承故障特征提取中是可取的。轴承早期故障特征难以提取,针对传统流形学习算法未能充分利用样本的类别信息,并且不能对新加入的样本进行快速处理的缺点,将监督的正交局部保持投影算法应用于轴承故障特征提取。运用正交局部保持投影方法通过高维训练样本数据求出正交转换矩阵,测试样本经正交转换矩阵转化后得到低维向量。利用不同故障样本的类间散度和同种故障样本的类内散度两个指标来衡量该方法的有效性,经故障实例仿真表明正交局部保持投影算法的应用效果较传统方法以及低秩鉴别投影算法更优越。通常利用时频分析方法获取机械设备工作时信号的时频谱图后,还要继续提取特征,然而提取特征比较困难,并且可能会丢失振动信号的重要信息而影响后续的故障识别。为了解决这个难题,引入机器学习领域中针对张量样本识别问题的支持张量机方法,直接对不同状态的三维时频谱进行识别。该方法的引入不仅可以简化诊断过程,而且能较大程度地提高诊断精度。
[Abstract]:Machinery and equipment occupy a very important position in modern industry and directly affect the normal operation of industrial production in enterprises. When there is an early failure of machinery and equipment, it is necessary to repair it in time so as not to cause more harm. This requires the technical ability to diagnose early minor faults. The diagnostic process of mechanical equipment is usually divided into three steps: the first step is the acquisition of diagnostic information; the second step is feature extraction. The third step is state recognition and fault diagnosis, in which feature extraction is the key link in the diagnosis process, which directly affects the accuracy of the diagnosis results. Therefore, how to extract the optimal low-dimensional fault features to improve the performance of fault classification is a huge challenge. In the stage of state recognition, pattern recognition is used. In this paper, the machine learning algorithm is based on the machine learning algorithm, and the rolling bearing is the object. In this paper, different machine learning methods are used to extract and identify the features of complex fault equipment. There is redundancy or no correlation among the parameters of multiple characteristic parameters of complex fault equipment, which is not conducive to fault diagnosis. The classical linear dimensionality reduction method can not meet the dimensionality reduction requirements of nonlinear fault data. In this paper, the feature extraction method of low rank discriminant projection is combined with sparse representation classifier to form a fault diagnosis model. In order to find the optimal parameters of corresponding data and realize parameter adaptation, a mesh search algorithm is introduced. The low rank discriminant projection algorithm can accurately describe the global structure and discriminant structure of the data. It is advisable to apply the model to the fault feature extraction of rolling bearing. The traditional manifold learning algorithm can not make full use of the class information of the samples and can not process the newly added samples quickly. The supervised orthogonal local preserving projection algorithm is applied to the bearing fault feature extraction, and the orthogonal transformation matrix is obtained by using the orthogonal local preserving projection method through the high-dimensional training sample data. The test samples are transformed by orthogonal transformation matrix to obtain low-dimensional vectors. The validity of the method is evaluated by two indexes: the inter-class divergence of different fault samples and the intra-class divergence of the same fault samples. The simulation results show that the orthogonal local preserving projection algorithm is more effective than the traditional method and the low rank discriminant projection algorithm. Usually, the time-frequency analysis method is used to obtain the time-frequency spectrum of the signal when the mechanical equipment is working. Continue to extract features, however, feature extraction is more difficult, and may lose important information of vibration signals and affect subsequent fault identification. To solve this problem, In the field of machine learning, a new machine method supporting Zhang Liang's sample recognition problem is introduced to recognize the 3D time spectrum of different states directly. This method can not only simplify the diagnosis process, but also improve the performance of machine learning. Moreover, the diagnostic accuracy can be improved to a large extent.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TH133.33;TP181

【相似文献】