流体动压型机械密封开启过程的声发射特征监测研究

发布时间：2018-02-08 15:16

本文关键词： 机械密封状态监测小波包声发射特征 Elman神经网络 PSO　出处：《西南交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：机械密封是旋转机械中最常用的轴封形式之一,因其可靠性高、泄漏量少、工作寿命长和适用性强等特点,被广泛地用于石油化工、航天航空和核电能源等领域。因此,机械密封件的使用性能将直接对机械生产设备的安全性、生产过程的效率和生产成本产生影响。为此,对机械密封件采取实时状态监测,获取有用的特征信息,分析机械密封在启动过程中的端面摩擦状态,将是十分必要的。可以在密封件失效前及时发现并维修,避免因过早更换密封件造成的资源浪费、成本提高,或是延迟更换造成的安全事故。通过机械密封开启过程监测实验的设计,选取电涡流和声发射法对密封端面的膜厚信息进行监测。实验采集了机械密封在开启全过程中的电涡流和声发射信号,通过对电涡流信号的分析,建立起信号的变化与端面接触状态和摩擦状态改变的对应关系。将密封的开启过程分为干摩擦、混合摩擦和流体摩擦三种摩擦状态,采集的信号在这三个状态中均有明显的特征体现。然后对开启过程中的声发射信号进行分析,声发射信号的变化情况也能与三种摩擦状态进行对应,采用小波包分析法对信号进行降噪处理。在设计的有无密封环对比实验中,分析得到中高频信号包含更多的与机械密封有关的信息的结论,选取声发射信号中的高频信息,根据机械密封端面的摩擦特性选择适合的特征指标,进行时频域特征提取,筛选得到有效的声发射信号特征。筛选得到的特征进行归一化处理后,分析特征对三种摩擦状态具有较好的可识别性。将归一化后的特征值作为Elman神经网络的输入向量,构建含反馈层的四层网络模型。利用训练样本进行训练后,对测试样本进行识别,得到较好的识别效果。之后选取不同的训练样本和测试样本,建立不同的网络模型进行模式识别,发现均能很好地将不同的摩擦状态数据进行分类。结果证明,选取的声发射特征能有效地识别机械密封开启过程中的端面情况。基于PSO算法实现对神经网络的优化,通过加惯性权重因子和矩阵化设计实现了对PSO算法的改进,提高了算法的运行效率同时提升了算法的收敛速度。对比优化前后神经网络的输出结果证实,PSO算法对神经网络的训练速度和精度、收敛速度和状态识别精度等方面,均有明显的优化效果。
[Abstract]:Mechanical seal is one of the most commonly used shaft seals in rotating machinery, because of its high reliability, less leakage, long working life and strong applicability, it is widely used in petrochemical, aerospace and nuclear energy fields. The performance of mechanical seals will have a direct impact on the safety of mechanical production equipment, the efficiency of production process and production cost. Therefore, real-time monitoring of mechanical seals is adopted to obtain useful feature information. It will be very necessary to analyze the friction state of the end face of the mechanical seal in the starting process. It can be found and repaired in time before the failure of the seal, so as to avoid the waste of resources and increase the cost caused by the premature replacement of the seal. Or a safety accident caused by a delay in replacement. The design of the monitoring experiment through the mechanical seal opening process, Eddy current and acoustic emission methods are selected to monitor the film thickness information of the seal end face. The eddy current and acoustic emission signals of the mechanical seal during the whole process of opening are collected experimentally, and the eddy current signal is analyzed. The relationship between the signal change and the change of the contact state and the friction state is established. The opening process of the seal is divided into three kinds of friction states: dry friction, mixed friction and fluid friction. The collected signals have obvious characteristics in these three states. Then the acoustic emission signals in the process of opening are analyzed, and the changes of the acoustic emission signals can also correspond to the three friction states. The wavelet packet analysis method is used to reduce the noise of the signal. In the contrast experiment with or without the seal ring, the conclusion that the middle and high frequency signal contains more information related to mechanical seal is obtained, and the high frequency information of the acoustic emission signal is selected. According to the friction characteristics of the mechanical seal face, the suitable feature index is selected, and the feature extraction in time-frequency domain is carried out, and the effective acoustic emission signal feature is obtained. After normalized processing, the selected features are normalized. The normalized eigenvalue is used as the input vector of Elman neural network, and a four-layer network model with feedback layer is constructed. After different training samples and test samples are selected, different network models are established for pattern recognition. The results show that the selected acoustic emission features can effectively identify the end face in the process of mechanical seal opening. The neural network is optimized based on PSO algorithm. The improvement of PSO algorithm is realized by adding inertia weight factor and matrix design. Compared with the output results of neural network before and after optimization, the training speed and precision, convergence speed and state recognition accuracy of PSO algorithm for neural network are proved. All of them have obvious optimization effect.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TH136

【参考文献】