【作者】 车红昆；

【导师】 项占琴； 吕福在；

【作者基本信息】 浙江大学 ， 机械制造及其自动化， 2011， 博士

【摘要】 缺陷类型识别是定量超声无损检测中重要的基础性问题。随着超声无损检测技术朝着高可靠性、高精度、高实时以及定量化方向发展,研究应用于在线超声检测的缺陷类型识别技术具有日益重要的学术意义和工程价值。虽然通过超声成像重构缺陷几何轮廓的方式能够实现对缺陷类型识别,但需要等待耗时的全局超声扫描和数据合成的过程,其实时性不能满足高速的在线检测需要。而非成像式超声检测缺陷类型识别方法直接从超声反射回波信号中提取特征参数,通过分析特征参数与缺陷类型之间的对应关系实现对缺陷类型识别,由于不需要等待全局超声扫描和数据合成,该方法具有较高的实时性,特别适用于在线超声检测的场合。在实际应用中,由于存在材料结构噪声对信号的干扰以及小样本条件下先验知识缺乏等困难,非成像式超声检测缺陷类型识别的准确性和可推广性受到了严重影响。针对目前存在的这些问题,本论文对非成像式超声检测缺陷类型识别中的关键技术,包括结构噪声消除、缺陷特征提取与类型识别进行了系统的研究,提出了基于小波包变换的时频邻域自适应消噪方法、基于SFFS搜索的时频优选特征提取算法以及两种基于支持向量机的融合决策识别方法法,并分别采用人工缺陷和石油套管自然缺陷对上述方法的可行性和有效性进行了验证。第一章,论述非成像式超声检测缺陷类型识别的重要意义,综合国内外关于非成像式超声检测缺陷类型识别关键问题的研究现状,分析当前研究中存在的问题,确定进一步研究的方法路线。第二章,研究超声反射回波信号的组成、分布特点和平稳特性,分析典型人工缺陷的超声反射回波信号在不同空间域上的信息特征,为后续的信噪分离、特征提取和类型识别工作提供理论基础。第三章,在分析结构噪声分布特点的基础上,提出基于小波包变换的时频邻域自适应消噪方法。通过仿真信号和实测信号的消噪实验,验证该方法在提高信号信噪比和抑制信号失真方面的有效性。第四章,确定超声反射回波信号的多特征提取框架,对四种相互独立的传统特征提取方法进行研究,并给出具体的实现算法。针对传统特征提取方法缺乏量化依据的问题,提出基于小波包分解、Fisher准则和SFFS搜索算法的时频优选特征提取算法,并采用可分性测度对上述特提取方法的有效性进行评价。第五章,针对小样本条件下超声检测缺陷类型识别的困难,提出两种基于支持向量机的融合决策识别方法,分别应用于缺陷类型框架已知和未知的场合。通过对人工缺陷进行类型识别,验证上述识别方法的有效性。第六章,将所提出的方法应用于石油套管自然缺陷的类型识别,研究信号消噪和特征提取对识别正确率的影响,验证多特征融合决策识别方法的识别能力和泛化能力,分析整个识别过程的时间耗费以及应用于石油套管在线超声检测的可行性。第七章,对论文的主要内容、研究结果和创新点进行总结,并对以后的工作进行展望。更多还原

【Abstract】 Defect identification is an important basic issue in quantitative ultrasonic nondestructive testing. With the development of ultrasonic testing towarding high reliability, high accuracy, real-time and quantitative analysis, it has greatly academic significance and engineering value to research on online ultrasonic defect identification. Although it is possible to achieve defect identification by ultrasound imaging method to synthesize geometric contour of the defect, but because of the time cost for the scan of whole object and the data synthesis, the imaging method can not meet the needs of high-speed online ultrasonic testing. Non-imaging ultrasonic defect identification method, which extracts features directly from ultrasonic signal and achieve defect identification by analyzing the relationship between features and defect types, does not need the time waiting for object scan and data synthesis, so, it is very suitable for online ultrasonic inspection and defect identification. However, there are still some problems in the practical application, such as the disturbance of grain noise, the shortage of priori knowledge in the small-sample situation, which can affect the accuracy and generalization of defects identification seriously. To solve these problems, some systematic studies are carried out on grain noise removal, feature extraction and pattern recognition, and the targeted methods are put forward, including a time-frequency adaptive de-noise method base on wavelet packet decomposition, a time-frequency features extraction method based on SFFS search algorithm, two fusion decision-making identification methods based on support vector machine.At last, tests are carried out with artificial defects and natural defects on oil casing pipe to verify the feasibility and effectiveness of these methods.In the first chapter, the importance of non-imaging ultrasonic defect identification is discussed, the research situation at home and abroad on the key technologies of non-imaging ultrasound defect identification is presented, and problems in current research are analyzed to guide the ways for further research work.In the second chapter, the composition, distribution and non-stationarity of ultrasonic signals are discussed, and the characteristics of ultrasonic signal achieved from typical artificial defect are analyzed at different space domains. These analysis results can provide a theoretical basis to noise removal, feature extraction and recognition in the follow-up section.In the third chapter, considering the distinctness of distribution between defect signal and noise, a time-frequency adaptive de-noise method base on wavelet packet decomposition is presented. Experiments are carried out with both simulated signals and real signals to verify the effectiveness of this method for the signal noise ratio improvement and the suppression of signal distortion.In the forth chapter, the multiple features extraction frame of ultrasonic signal is presented. Four irrelevant traditional feature extraction methods are discussed, and their implementations are presented. Considering the lack of quantitative criterion in traditional feature extraction methods, a new time-frequency feature extraction method based on wavelet decomposition, Fisher principle and SFFS selection algorithm is presented to optimize feature mining capability. The effectiveness of thees feature extraction methods are evaluated with distinguish ability criterion.In the fifth chapter, considering the small sample problem of defect identification in ultrasonic testing, two fusion decision-making identification methods are presented respectively for the applications where the defect type frame is known or unknown. A test with different artificial defects by ultrasonic inspection method is carried out to verify the effectiveness of these identification methods for the improvement of accurate and generalization.In the sixth chapter, theories and methods presented out in this paper are applied to the identification of natural defects on oil casing pipes. The influences caused by noise and feature extraction methods to the identification accurate are discussed. And the accurate and generalization of fusion decision-making identification methods are verified. At last the time costs of these methods are measured, and the feasibility for online ultrasonic testing is analyzed.更多还原