基于隐Markov模型的滚动轴承故障诊断方法研究

发布时间：2018-01-01 05:29

本文关键词：基于隐Markov模型的滚动轴承故障诊断方法研究　出处：《西南交通大学》2013年硕士论文　论文类型：学位论文

【摘要】：滚动轴承广泛应用于旋转机械中,然而由于各种原因,滚动轴承很容易发生各种形式的故障,因此对滚动轴承开展故障诊断便成为保证设备正常运行的关键,具有重大的现实意义。针对传统的模式识别方法(如神经网络识别法)一直停留在静态模式识别上的不足,本文提出采用一种近年来在语音识别技术中发展较快的动态模式识别技术——隐马尔科夫模型(Hidden Markov Model, HMM)来对滚动轴承进行故障诊断。HMM建模时统计的是一个时间跨度上的动态信息,特别适合对信息量大、非平稳、特征重复性不佳的诊断信号进行分类。通常情况下的动态过程为序列行为改变的表现,滚动轴承亦如此。若一个短时信号定义为一个帧,则每一个故障类型特定帧之间的转移是不同的,因此可以用HMM来对特定帧的存在和各帧之间的转移做统计处理。此外,利用HMM进行模型训练时所用的样本较少、速度较快,并且诊断的精度较高、模式分类能力较强,因此非常适合对滚动轴承的振动信号进行故障建模和分类。根据实际的滚动轴承典型故障实验数据和HMM故障诊断原理,本文首先对从各个轴承状态下采集的振动信号进行分帧处理,然后提取每帧信号的时域、频域和小波包能量特征参数,组成特征矢量,并运用主成分分析(Principal Components Analysis,PCA)技术对特征矢量进行降维处理,在损失状态信息较少的情况下,将多个特征指标转化为几个综合的特征指标,用较少的特征参数来代表轴承状态的绝大部分信息。将PCA降维后的特征参数组合在一起形成特征矢量,大大简化了后续模型的输入。根据HMM按观察值的分类,本文研究了基于连续高斯混合密度HMM(CGHMM)的滚动轴承故障诊断和基于离散HMM (DHMM)的滚动轴承故障诊断技术,并通过实验将两种方法进行了对比分析。由于利用DHMM进行故障诊断时,需要对特征值进行量化编码,这必然会带来一定的量化误差,因此DHMM模型的识别率(90%以上)低于CGHMM(98%以上),但由于DHMM建模简单,故其训练速度快于CGHMM。可以针对诊断系统侧重点的不同,选择两种方法。为了验证HMM模型(包括CGHMM和DHMM)用于故障诊断的有效性和优势,本文又将其识别结果与模式识别方面应用最广泛的BP神经网络的识别结果进行了对比分析。实验结果表明,BP神经网络的训练速度慢于HMM,而且识别率低于HMM。由此可见,动态模式识别方法HMM比神经网络识别法在故障诊断方面更具优势,具有更广泛的应用和发展前景。
[Abstract]:Rolling bearings are widely used in rotating machinery. However, due to various reasons, rolling bearings are prone to various faults. Therefore, fault diagnosis of rolling bearings is the key to ensure the normal operation of equipment, which is of great practical significance.
In view of the traditional pattern recognition methods (such as neural network recognition method) has been stuck in the lack of static pattern recognition, this paper adopts a speech recognition technology in recent years in the development of dynamic pattern recognition technology quickly, the hidden Markoff model (Hidden Markov Model, HMM) to statistics for fault diagnosis of rolling bearing when modeling.HMM the dynamic information is a time span, particularly suitable for a large amount of information, the non-stationary signal characteristics, diagnosis of repetitive poor classification. The dynamic process is usually the case for sequences of behavior change, rolling bearing is also true. If a short-time signal is defined as a frame, the transfer between each fault type specific frame is different, so you can use HMM to transfer between the specific frame and the presence of each frame to do statistical processing. In addition, the model is trained by HMM It has fewer samples, faster speed, higher diagnosis accuracy and stronger ability of pattern classification, so it is very suitable for fault modeling and classification of rolling bearing vibration signals.
According to the data of typical faults of rolling bearing fault diagnosis experiment and HMM principle of practice, this article first frame processing of vibration signals collected from each bearing condition, then each frame signal extraction in time domain, frequency domain and wavelet packet energy feature, form feature vector, and using principal component analysis (Principal Components, Analysis, PCA) the technology of feature vector dimension, the less loss of state information under the condition of a plurality of characteristic indexes into several comprehensive indexes, most with less feature parameters to represent bearing state information. PCA will reduce the dimension of feature parameters are combined together to form a feature vector, which greatly simplifies the follow-up the input of the model.
According to the classification of HMM according to the observations, we study the continuous mixture density based on Gauss HMM (CGHMM) and the fault diagnosis of rolling bearings based on discrete HMM (DHMM) of the rolling bearing fault diagnosis technology, and through the experiment of the two methods are compared and analyzed. Due to the use of DHMM for fault diagnosis, the need for characteristic value encoding, which will bring a certain quantization error, so the recognition rate of the DHMM model (more than 90%) than CGHMM (more than 98%), but because the DHMM modeling simple, so the training speed is faster than CGHMM. for the diagnosis system of the different focus, two kinds of methods.
In order to verify the HMM model (including CGHMM and DHMM) is effective for fault diagnosis and the advantages of combining BP neural network to identify the identification results and the pattern recognition of the most widely used results were compared and analyzed. Experimental results show that the BP neural network training speed is slower than that of HMM, and thus the recognition rate is lower than HMM. dynamic, HMM pattern recognition method than neural network recognition method in fault diagnosis has more advantages, applications and broader development prospects.

【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TH133.33;TH165.3

【参考文献】