【作者】 刘国华；

【导师】 周泽魁；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2008， 博士

【摘要】 声发射(Acoustic Emission,AE)是无损检测和评估中的一种重要方法,可广泛应用在石油化工、航天航空、材料实验和交通运输等领域。AE信号处理是AE检测技术研究的核心。根据信号波形特征,可将AE信号分为突发、连续和混合三种类型,探讨不同类型AE信号的特征和传播机理,研究相应的AE信号处理和分析技术,对于准确反演AE源信息,具有非常重要的意义。本文在理论分析和实验研究的基础上,从信号处理的角度,分别研究了三种类型AE信号预处理、特征提取和融合方法。论文的主要工作和创新点如下:(1)研究了三种类型的AE信号特征及其适用的信号去噪方法。突发型信号可以分解为一个反映群速度的调制信号和一个反映相速度的指数型信号的组合,因此论文提出一种小波和MP算法的融合去噪方法对突发型信号进行去噪处理;连续信号频散特性复杂,模态个数多,论文选取了shannon熵准则的小波包方法进行了去噪处理;混合信号在不同阶段表现出不同特征,去噪处理的关键在于对不同阶段的识别,论文分别选取了小波包和小波—MP融合方法对不同阶段的声发射信号进行了去噪处理。研究表明,不同去噪方法处理后的信号SNR明显提高,RSME降低,且原始信息得到了很好保留。(2)针对AE信号概率密度函数估计困难的问题,进行了混合因子的ICA预处理,并在此基础上提出了一种基于经验特征函数的FastICA改进算法。与常用的几种FastICA算法相比,新方法具有更好的收敛效果。实验表明,对于水管泄漏产生的连续型AE信号,通过改进ICA算法分离后的定位精度显著提高,不同泄漏点的定位精度均在3%以内。(3)提出了利用Hurst指数进行混凝土的AE信号自相似特征提取方法,并在此基础上构造了基于Hurst指数均值和方差的声发射阶段识别分类器。通过C50、C60两类混凝土的Hurst指数研究表明,随着应力增加,Hurst指数表现出一种从大变小,再急剧增大的特征,且随着材料强度增大,临界点的Hurst指数突变趋势加剧。对不同阶段的AE信号识别结果表明,基于Hurst指数识别方法对于非稳定阶段的识别具有很高的精度,且受实验条件影响小。(4)提出了利用高阶谱进行混凝土AE信号的非高斯性特征提取办法,并建立了基于高阶谱均值和方差的声发射阶段识别分类器。通过提取C50、C60两类混凝土AE信号的双谱特征发现,随着荷载增加,AE信号非高斯性增大,尤其是在破坏的临界状态,双谱均值出现3个数量级的增长。对不同阶段的AE信号识别实验结果表明,在初始阶段和稳定阶段,基于高阶谱特征分类方法具有一定优势。(5)建立了贝叶斯网络的混凝土安全特性评估模型。模型采用基于熵的离散化方法对多特征参数进行了离散化,并分别采用基于网络测度和梯度下降方法进行了贝叶斯网络结构和参数构建。实验结果表明,与单参数特征评估方法比较,基于贝叶斯网络的评估具有更好的效果和较高的精度。更多还原

【Abstract】 Acoustic emission (AE) is an important method of no-destructive measurement and evaluation, which is widely used in a number of fields, such as petrochemical industry, aerospace industry, material test, and transportation. The AE signal processing plays key role in the research of the AE measurement. According to the characteristic of the waveform, the AE signal can be divided into three types, i.e., burst signal, continuous signal and mixed signal. To study the characteristics and propagation mechanisms of different types of signals and find out their corresponding analysis methods is an important task to correctly inverse the information of the AE source.Based on both theoretical analysis and experimental study, signal preprocessing, feature extraction and information fusing methods for three types of AE signals are carefully studied from the aspect of signal processing. The main works and innovations of this dissertation are listed as follows:(1) The characteristics of three types of AE signals and their corresponding de-noise methods are studied separately. The burst signal can be decomposed into a modulation signal corresponding to group velocity and an exponential signal corresponding to phase velocity. Therefore, a new method which combines two types of methods, i.e., wavelet analysis and matrix pencil (MP) analysis, is introduced to denoise the burst signal. For the continuous signal, the frequency dispersion is relatively complicated and its mode number is large. The wavelet packet method based on the Shannon entropy rule is used to denoise the continuous signal. The mixed signal has distinctly different characteristics at different stages of the signal emission process, thus the key point of de-noise processing is to identify the different signal emission stages. Different methods, such as wavelet packet analysis and wavelet-MP-combined analysis are used at different stages of the signal emission process. The results show that the signal denoised by the methods introduced in this dissertation not only nicely keeps the information of the original signal but also obviously improves the SNR and reduces the RSME.(2) Due to the difficulty in estimation of the probability density function (PDF) of the AE signal, the factor-mixed ICA preprocessing is applied. An improved FastICA algorithm based on the empirical PDF is proposed. Compared with the general FastICA algorithms, the improved method has better convergence effect. The experimental results show that for the continuous AE signal generated from the water pipe leak the improved ICA method can effectively improved the location precision, i.e., the location precision is controlled under 3% for different leak positions.(3) Based on the Hurst exponent, a self-similar feature extraction method is introduced to analyze the AE signal of the concrete. Furthermore, a classification method is proposed to identify the stages of the AE process according to the average value and the variance of the Hurst exponent. We calculate the Hurst exponent for two types of concrete, i.e.. C50 and C60, and find there exists a transition phenomenon: for small stress the Hurst exponent decreases as the stress increases; it reaches a minimum at a certain critical point; then increases rapidly if we further increase the stress. This transition phenomenon is more obviously for the concrete with higher strength. The results show that the Hurst exponent based classification method has high precision for the identification of the unstable stage and is rarely affected by the experimental condition.(4) Based on the high-order spectral analysis, a non-Gaussian feature extraction method is also introduced to analyze the AE signal of the concrete. Another classification method is further proposed to identify the stages of the AE process according to the average value and the variance of the high-order spectrum. We study the characteristic of the bispectrum for two types of concrete, i.e., C50 and C60. The deviation from the Gaussion distribution increases as the stress increases. Especially for the status close to critical state of destruction, the average value of the bi-spectrum displays an increase of three orders of magnitude. The results show that the high-order spectrum based classification method is effective for identification of the initial stage and the stable stage.(5) A safety evaluation model of concrete based on Bayesian networks is proposed. The model applies an entropy-based discretization method to discrete the characteristic parameters. Using network measurement method and gradient decrease method, the structure and the parameters of Bayesian networks are established separately. The experimental results show that the Bayesian networks based evaluation method behaves more effectively and has a higher precision than other evaluation methods with singular parameter.更多还原