【作者】 李政；

【导师】 罗飞路；

【作者基本信息】 国防科学技术大学 ， 仪器科学与技术， 2010， 博士

【摘要】 作为五大无损检测手段之一,超声无损检测在工业领域得到了广泛的应用。论文以“十一五”预研项目为依托,围绕外场条件下飞机常见结构的原位超声无损检测,结合仿真和实验,深入研究了超声无损检测系统建模实现、螺栓缺陷的快速检出、铆钉缺陷的定量检测以及起落架缺陷的分类识别等相关问题。其主要内容及创新如下：对接触式脉冲回波超声无损检测系统进行了仿真和实验研究。根据外场条件下通用超声检测系统的特点,将传统的电路模型和数学模型结合起来,在改进的电声测量模型的基础上,提出了接触式脉冲回波超声检测系统的仿真模型。在模型中,耦合剂是不可忽略的因素。对系统回波强度和耦合剂厚度之间的关系进行了研究,定量地解释了实验中的有关现象,即耦合剂所形成的薄层厚度和传递超声波的强度之间存在确定的关系。利用现有的超声探伤仪对传输导线和换能器相关参数进行测试,并代入到模型中,获得了系统超声发射部分中激励脉宽等关键参数的调节趋势和最佳设计值,进而应用于实际系统的设计和实现。研究了基于微粒群优化和匹配追踪的超声信号参数估计方法。对经典的超声信号模型进行了仿真和实验研究,在确定该模型可用性的基础上,具体分析了一些典型的超声信号参数估计方法,并对比了这些方法在参数估计精度和速度上的差异。为了在测试中兼顾参数估计精度和参数估计速度,提出一种基于改进微粒群优化和匹配追踪的超声信号参数估计方法,在保持种群微粒数不变的情况下,淘汰适应值最小的微粒并加入新的微粒,有效提高算法的收敛速度和收敛精度。从总体上看,该方法对于超声信号的随机噪声干扰不是非常敏感,并可降低参数估计实际计算量。对该方法的估计特性进行分析,分别在不同深度的缺陷信号参数估计、不同类型的缺陷信号参数估计以及飞机螺栓部件缺陷信号参数估计上进行了实验验证。研究了基于瞬时频率熵的超声缺陷检测方法。传统超声探伤仪在检测时通常通过屏幕上不同位置的回波来判断缺陷的位置,但是对于存在于界面附近的缺陷,普通的回波观察法不易获得缺陷位置的准确信息。针对此类问题,论文从时频分析的角度,根据缺陷回波(包括界面回波)和系统噪声瞬时频率存在的差别,提出一种基于瞬时频率熵的超声缺陷检测方法,结合瞬时频率熵曲线求解过程中相对极值点以及窗的宽度参数确定方法,对不同时频分析方法所计算的瞬时频率熵,分别采用仿真和实验信号进行测试精度的对比。该方法利用基于时序多相关预处理的经验模式分解实现了瞬时频率和瞬时频率熵的计算,克服了普通经验模式分解过程中产生虚假固有模式函数分量的问题,有效降低了检测中系统噪声的影响,提高了飞机铆钉缺陷位置信息的检测精度。研究了基于最小二乘支持向量机(LSSVM)的超声缺陷分类识别技术。在飞机部件的检测中,不同类型的缺陷对飞机结构有着不同的影响,需要加以有效区分。但是外场检测中可获得的信号样本有限,基于此,论文以适用于小样本机器学习的LSSVM为基础,提出一种基于保持种群多样性微粒群优化(ARPSO)的LSSVM分类识别方法,通过ARPSO算法实现LSSVM关键参数的寻优,在保持特征可分性的前提下,利用距离评价因子对超声信号特征向量进行约简,以提高LSSVM的训练速度。分别在标准试块缺陷分类识别和起落架典型缺陷分类识别上进行了核参数选择、多类分类器设计等方面的实验验证。和其他LSSVM方法相比,该方法兼顾了缺陷单类分类准确率、总体分类准确率和总体训练时间,可有效应用于外场飞机原位检测。更多还原

【Abstract】 As one of the five major nondestructive test (NDT) methods, ultrasonic NDT has been applied in the industry field extensively. Relying on the advanced project of Eleventh Five-Year Plan, we combine the simulation and experimental study on the base of the in-situ ultrasonic NDT of common structures of the aircraft. On the basis of survey on the state of the art of ultrasonic NDT technique, we launch our studies on severel key techniques such as the simulation and realization of ultrasonic NDT system, method of rapid checkout of cracks in bolts, method of quantitative measurement on flaws in rivets and classification of defects in gear in this paper. Our contributions are summarized as follows:Firstly, we carried out the simulation and experiments on the contact pulse-echo ultrasonic NDT system. Aiming at the characteristics of general ultrasonic test system in outfield, we proposed the simulation model of contact pulse-echo ultrasonic NDT system with the combination of conventional circuit model and mathematical model, which was formed on the base of improved electro-acoustic measurement (EAM) model. The couplant was considered as a factor that could not be ignored in this model. Furthermore, we studied the relationship between the intensity of echo and the thickness of the couplant, which could explain the phenomenon in the experiment that this relationship could be expressed definitely. We used the analog ultrasonic tester to measure the parameters of the transducer and the cable, and then introduced the paremeters into the model. The optimal parameters were obtained by debugging the emitting part of the model, which was applied in the design and realization of the ultrasonic NDT system.Secondly, we proposed a parameter estimation method of ultrasonic signal based on particle swarm optimization (PSO) and matching pursuit (MP). Simulations and experiments were studied on the classical model of ultrasonic signal. On the base of the availability of this model, we gave the detailed analysis on some conventional methods of parameter estimation. The difference of accuracy and speed of these parameter estimation methods was compared. Considering the precision and speed of estimation in the test, we introduced a method of parameter estimation of ultrasonic signal based on improved PSO and MP. Particle of smallest adaptive value was obsoleted and new particle was added in the swarm, which retained the number of total particles unchanged. The convergence speed and convergence accuracy of the algorithm was improved effectively. Meanwhile, this method was not so sensitive to the random noise in the ultrasonic signal, and could reduce the calculated amount in parameter estimation. As was shown in the results of experiments, this method can not only deal with the ultrasonic signals of different types, but also with those in different positions. Thirdly, we proposed a new method, entropy of instantaneous frequency (IF), to detect the position of the flaw. Generally speaking, the flaw positon is often confirmed by the observation of different echoes in the screen of ultrasonic tester. However, this method of observation can not distingwish the flaws near the interface. Therefore, from the view of time-frequency analysis (TFA), this paper presented a method based on the entropy of IF in the detection of flaw position. Using the methods for calculating the relative extreme points and determining the window width in the process of solving the entropy curve of IF, we compared the precision of entropies measured by this method and other different time-frequency methods. This method calculated the IF and entropy of IF based on empirical module decomposition (EMD) with the pre-processing of multi-correlation time sequence (MCTS), which overcame the problems of fault intrinsic mode function (IMF) created by EMD. This method reduced the influence of system noise in the test and improved the test accuracy at the same time.Finally, we studied the classification method of ultrasonic defects based on least square support vector machine (LSSVM). In the detection of aircraft parts, different defect leads to different influence. Therefore, it is necessary to distinguish them quickly and accurately. But the signal samples are limited in the outfield test. According to this, we proposed a classification method of least square support vector machine (LSSVM) based on attractive and repulsive particle swarm optimization (ARPSO), which could find the optimal parameters of LSSVM by ARPSO. Under the prerequisite conditions of feature discriminability, the features were reduced by the distance evaluation factor, which could improve the training speed. In the test of flaws in the standard samples and the gear, this method was applied respectively. The selection of kernel parameters and design of multi-class classifier were verified in the experiments. Comparing with other method of LSSVM, this method could combine the merits of classification accuracy of single class, classification accuracy of universal class and the total traing time, which could apply in the in-situ test of aircraft outfield.更多还原