【作者】 李冬梅；

【导师】 傅攀；

【作者基本信息】 西南交通大学 ， 测试计量技术与仪器， 2008， 硕士

【摘要】 滚动轴承作为各类旋转机械中最常用的通用零部件之一,也是旋转机械易损元件,因此滚动轴承故障诊断的理论、方法和应用得到特别的重视。振动分析是进行滚动轴承状态监测与故障诊断的重要手段。本文旨在研究滚动轴承的振动信号处理和模式识别的方法。为此,首先设计并搭建了滚动轴承故障诊断实验台,进行了大量实验;采集到滚动轴承正常状态、外圈故障和内圈故障的振动信号;对振动信号进行了常规的时域、FFT、倒频谱分析以及时延相关解调分析,通过对应的谱图初步诊断轴承的故障类型;构建BP网络和B样条模糊神经网络,将特征提取到的信号进行归一化处理,分别输入这两种网络进行模式识别,最终识别出滚动轴承的故障类型;最后利用.NET平台编写了滚动轴承状态监测诊断系统,以实现更好的人机交互。滚动轴承的振动信号呈现调制特征,因此对测得的振动信号进行解调是滚动轴承故障诊断的关键,常规分析和包络解调技术对此类信号进行分析时易受噪声影响,使得滚动轴承故障特征难以凸显。故本文采用时延相关解调法用于滚动轴承的故障诊断,这一方法的实现步骤是首先对测得的振动信号进行自相关分析,再对自相关函数进行时延,然后对时延后的自相关函数进行Hilbert变换解调。实验分析结果证实了时延相关解调技术是一种良好的降噪解调技术,并且文中对时延相关解调法的时延量做了研究,通过选择不同的时延量值进行比较分析,发现时延量在某区间内可以任意选取。抽取对振动信号分析后的有效特征,将其作为神经网络的输入。本文主要采用BP神经网络和B样条模糊神经网络进行识别,BP网络在进行识别时虽然效果比较良好,但缺点也表现出来:网络训练时表现出“黑箱”的性质,收敛速度慢等。而B样条模糊神经网络内部结构连接透明,在网络进行训练时能方便地根据实际需要调节网络结构,同时B样条模糊神经网络的局部学习和存储的特性使得收敛速度快,便于实现在线监测。通过这两种网络的对比发现B样条模糊神经网络在训练速度和识别率上均优于BP网络。最后利用.NET这一良好平台,设计并编写了滚动轴承状态监测诊断系统,将C#与Matlab联合应用实现良好的人机交互界面,让整个系统变得更智能且操作更简单。更多还原

【Abstract】 As one of the most common parts of various rolling mechanical equipments, rolling bearing is very vulnerable. Therefore, great importance has been attached to the theories、methods and applications of failure diagnosis of rolling bearing.Vibration analysis is a very important means for condition monitoring and fault diagnosis. This paper aims at the research on the methods of signal processing and pattern recognition. Therefore, firstly, a experimental platform was set up for the failure diagnosis of rolling bearing, on which we have done a lot of experiments; Then the vibration signals on the condition of normal rolling bearing、rolling bearing with failure on the outer circle and rolling bearing with failure on the inner circle were collected. After Analyzing on the vibration signal based on analyzing methods of Time-domain、Cepstrum、FFT、Time-delayed correlation demodulation, we got the spectrum through which the type of the failure of rolling bearing was found. By constructing BP networks and B-spline neural networks, dealing with the character of the vibration signal through Unification and putting them into two different networks to recognize, we could finally recognize the real type of the failure of rolling bearing. By applying the .NET platform, a condition monitoring software designing for rolling bearing was completed to fulfill the task of Human-computer interactive more effectively.The time delayed correlation demodulation is established in order to suppress the noise and demodulating the signal. The auto-correlation functions of vibration signals measured on bearing cases are first computed, which will reduce the noise greatly, but not change the modulation signature of the signals. Then the auto-correlation functions are delayed for some time lags in order to decrease the affection of noise before demodulated by Hilbert Transform. The effectiveness of this method id confirmed by simulated data and experimental data. Moreover, the faults on the outer ring, inner ring and rolling element can be recognized by the time delayed correlation demodulation. Experimental vibration signals measured from the rolling element bearings verify that time-delayed correlation demodulation is better than conventional methods, such as spectrum analysis and envelope analysis. In this paper the time-delayed size of the time-delayed correlation demodulation is investigated. Different time-delayed sizes have been analyzed, and their results are almost same. Therefore, the time-delayed size can be selected freely in one revolution.The effective character of the vibration signal was got out, it could be used for the input of neural networks. BP neural networks and B-spline neural networks is mainly applied for recognition. Although the result of the recognition of BP networks is comparatively good, its backwards seems to be evident, mainly due to its "BLACK BOX" character and slow convergence, while the inner result and interconnection of B-spline neural networks is transparent and the structure of the networks can be regulated conveniently according to the practical needs when training the networks, meanwhile the local study and the character of storage of B-spline neural networks make the fast convergence, which can be useful for the On-line monitoring. By comparing the two different networks we found that the training speed and the recognition rate of B-spline neural networks are superior to BP neural networks.Finally, by applying .NET platform, designing the software of the condition monitoring system of rolling bearing, using the C# and Matlab to realize the interface of Human-computer interactive, the whole system is becoming much more intelligent and operating the system is becoming much easier.更多还原