【作者】 李国厚；

【导师】 周泽魁； 黄平捷；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2011， 博士

【摘要】 现代工业生产的迅速发展对无损检测技术提出的要求越来越高。以涡流和超声为代表的无损检测方法在常规缺陷的定性检测方面取得了较好的效果。但是,由于受到各自特点的限制,它们在多层导电结构的缺陷分类与定量化检测、表面与深层微小缺陷的定性检测与定量化评估等方面的应用仍存在很多问题有待于解决,还不能满足实际应用需求。本文结合国家自然科学基金资助项目和教育部博士点基金资助项目,围绕涡流检测的参数优化及其与超声检测方法的复合应用技术展开研究,在涡流检测参数的试验优化设计、导电结构缺陷分类与定量化检测、涡流检测与超声检测的异源数据融合、检测结果的检测概率(POD)分析与不确定性评定等方面取得了成果。论文主要研究内容和创新点如下：1.将基于田口方法的稳健设计理论应用到涡流无损检测的参数优化中,研究了优化涡流检测参数的相关技术。根据田口稳健设计中参数优选与确定方法,应用信噪比分析原理,通过对试验数据的分析与比较,优化了激励频率与激励电压等关键检测参数；并基于测量的不确定度评定理论,研究了涡流检测结果的不确定度评定技术,给出了典型的评定方法与过程。2.针对常规涡流与超声检测在面向导电结构深层缺陷和表面微小缺陷检测中存在的问题,采用Dempster-Shafer证据理论以及BP网络,对涡流与超声检测的异源数据进行融合分类,提高缺陷分类的准确度。论文提出将模糊C-均值算法与Dempster-Shafer证据理论相结合的异源数据融合方法,避免在进行证据组合时对于大量试验数据或专家知识的依赖,提高了算法的实用性,便于进行缺陷的定量化评估。3.在检测参数优化与缺陷分类的基础上,研究采用径向基神经网络(RBFNN)以及模糊贝叶斯网络(FBN)对缺陷参数进行定量化评估的方法。论文采用减聚类算法对常规的RBFNN进行改进,并将模糊处理与贝叶斯网络结合起来构建模糊贝叶斯网络,通过利用涡流与超声检测的试验数据进行网络的训练以及缺陷的定量化反演,提高了反演的精度和效率。4.引入POD对无损检测结果进行评估。首先研究借助于POD对导电材料缺陷检测结果进行评估的理论,通过综合分析多种现有经验模型的特性,提出了一种简单实用的POD经验模型,在用于检测结果的评估中达到了一定的精度；然后研究了基于辅助模型的POD(MAPOD)建模与评估技术,并采用试验数据验证MAPOD模型的推广性能,弥补了经验模型的不足。更多还原

【Abstract】 With the rapid development of the modern industrial production, higher and higher demands have been put forward for the nondestructive testing technique. As representative measures of the nondestructive testing (NDT) methods, Eddy current testing (ECT) and Ultrasonic Testing (UT) techniques have gained better effects in the qualitative inspection of the conventional defects. However, due to their own limitations, lots of difficult problems remain unsolved for ECT and UT in the defects classification and quantitative inspection of the multi-layer conductive structures, the qualitative inspection and the quantitative evaluation of the minor defects, and the practical inspection demands can’t be satisfied. In this dissertation, supported by the National Natural Science Foundation of China and the Ph.D. Programs Foundation of Ministry of Education of China, investigations have been conducted on the optimization of ECT and UT and their compound application technology, much effort has been devoted to the investigation of the optimization design of ECT experiment, the defects classification and quantitative testing of the conductive structures, the data fusion from different types of sensors of ECT and UT, the probability of detection (POD) and the uncertainty of measurement research of the testing results by means of two different kinds of NDT methods, and some primary results have been obtained in the end. The main contents of the dissertation and the innovations of the research chiefly include the following:1. The robust design based on Taguchi method is introduced into the optimization design of the ECT experiment so as to determine the optimal inspection parameters of ECT conveniently. In accordance with the parameters selection and determination method in the Taguchi method, utilizing the SNR analysis principle, such key inspection parameters as the excitation frequency and voltage have been optimized by means of analyzing and comparing the experiment data. In addition, the assessing technology of uncertainty of measurement for ECT is investigated based on the theory of uncertainty of measurement, and the typical assessing scheme and process have been presented.2. In order to solve the problems in detecting deep defects and the minor surface defects of the conductive structures for ECT and UT, Dempster-Shafer evidence theory and BP networks are applied to fusing different detecting results from ECT and UT and the precision of defects classification has been improved. A data fusion method of different sources to combine the fuzzy C-means clustering algorithm with Dempster-Shafer evidence theory is proposed, avoiding depending on large amounts of experiment data or the expert knowledge in combining different evidences. The practicability of the algorithm is improved and thus it’s convenient for the quantitative evaluation of defects.3. On the basis of detection parameters optimization and the defects classification, the radial basis function neural networks (RBFNN) and the fuzzy Bayesian networks (FBN) are used to evaluate the defects parameters quantitatively. The subtractive algorithm is utilized to improve the conventional radial basis function neural network, and the fuzzy processing is combined with Bayesian networks to construct the FBN, the ECT and UT data are used to train the networks and inverse the defects quantitatively, and better effect has been achieved in the inversion precision and efficiency.4. The application of probability of detection (POD) in the evaluation of the inspection results is introduced. Firstly, the fundamentals of evaluating the inspection results of the flaws in the conductive materials are discussed by means of POD. On the basis of analyzing different existing POD empirical models synthetically, a practical three-parameter empirical model is put forward, and a certain precision has been arrived at in evaluating the inspection results; then the modeling and evaluating schemes based on model-assisted POD (MAPOD) are investigated, the experiment results validate the generalization performance of MAPOD, and the limitation of the empirical model has been offset.更多还原