基于多重分形理论的耐火材料声发射信号特征提取及损伤模式识别研究

发布时间：2018-05-27 19:50

本文选题：镁碳质耐火材料 + 声发射　；参考：《武汉科技大学》2015年硕士论文

【摘要】：耐火材料组成成分复杂，属于多孔、多相性的微观非均质材料。在损伤过程中所产生的声发射信号包含了损伤源的丰富信息。对耐火材料损伤特征进行有效的提取，并选择合适的分类器是实现其损伤模式识别的关键。本文针对耐火材料声发射信号具有多重分形性、非线性、非平稳的特性，利用多重分形理论、经验模态分解相结合的方法进行信号的特征提取，并采用支持向量机及BP神经网络两种分类方法进行损伤模式的识别，对耐火材料微观损伤的研究具有积极的意义。本文主要研究内容如下：（1）本文以镁碳质耐火材料为研究对象，通过单轴压缩试验模拟其受压应力状态下损伤状况，并采集受压过程中损伤声发射信号以进行分析。根据复合材料中不同组成成分损伤时发出的信号频率成分与其弹性模量及密度相关，分析并分选耐火材料典型损伤信号。（2）为从多重分形各项参数（Δα、Δf、K、MeanDq）中挑选出最佳损伤特征量，根据声发射信号的特点建立了一系列不同频率结构的仿真声发射信号，并通过仿真信号分析挑选出最佳特征量，最后用实验信号进行验证。分析结果表明，多重分形谱宽Δα值能够很好表征声发射信号的特征，最适合用作损伤特征量。（3）针对声发射信号非线性、非平稳的特性，通过EMD方法将信号分解为若干IMF分量，并将整个信号及各IMF分量的多重分形谱参数组成特征向量作为分类器的输入量。然后分别采用SVM及BP神经网络两种模式分类方法对损伤信号进行模式分类，两种方法的分类准确率均达到了90%以上，这也验证了采用EMD与多重分形谱参数相结合的方法对实验信号进行损伤特征提取的合理性。（4）对SVM及BP神经网络两种分类方法在不同训练样本下的分类结果进行对比分析，发现SVM能够在较小样本情况下实现更高分类准确率，比BP神经网络方法更具优势。
[Abstract]:The composition of refractories is complex and belongs to porous and heterogeneous micro heterogeneous materials. The acoustic emission signals generated during the damage process contain abundant information about the source of the damage. It is the key to identify the damage patterns of refractories to extract damage features and select suitable classifiers. Aiming at the multifractal, nonlinear and non-stationary characteristics of acoustic emission signals of refractories, the method of combining multifractal theory and empirical mode decomposition is used to extract the features of the signals. Two classification methods, support vector machine and BP neural network, are used to identify the damage patterns, which is of great significance to the study of micro-damage of refractories. The main contents of this paper are as follows: In this paper, the damage of magnesia-carbon refractories under compression stress is simulated by uniaxial compression test, and the damage acoustic emission signals are collected for analysis. The typical damage signals of refractories were analyzed and sorted according to the correlation between the frequency components and the elastic modulus and density of the different components of composite materials. 2) in order to select the best damage characteristic quantity from the multifractal parameters (螖伪, 螖 F ~ (1) K) mean DQ, a series of simulated acoustic emission signals with different frequency structures are established according to the characteristics of acoustic emission signals, and the optimum characteristic quantities are selected by analyzing the simulation signals. Finally, the experimental signals are used to verify the results. The results show that the multifractal spectrum width 螖伪 can well characterize the characteristics of acoustic emission signals, and is the most suitable for damage characteristic quantities. In view of the nonlinearity and nonstationarity of acoustic emission signal, the signal is decomposed into several IMF components by EMD method, and the multifractal spectrum parameters of the whole signal and each IMF component are composed of the eigenvector as the input of the classifier. Then SVM and BP neural network are used to classify the damage signal, and the accuracy of the two methods is over 90%. This also verifies the rationality of the method of combining EMD with multifractal spectral parameters to extract the damage features of experimental signals. (4) the classification results of SVM and BP neural network under different training samples are compared and analyzed. It is found that SVM can achieve higher classification accuracy in the case of smaller samples, and has more advantages than BP neural network method.
【学位授予单位】：武汉科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ175.1

【参考文献】