【作者】 陈渊；

【导师】 马宏伟；

【作者基本信息】 西安科技大学 ， 安全技术及工程， 2010， 博士

【摘要】 液压支架是煤矿综合机械化开采的主要设备,对现代煤矿生产具有举足轻重的意义。液压缸是液压支架完成各种运动和承载顶板压力的关键元件,它的质量直接影响着液压支架的可靠性,进而影响到煤矿的安全和正常生产。针对液压缸环焊缝质量检验的难题,为了克服目前手动超声A扫描检测方式的缺点和不足,保证焊接质量,以数字超声探伤仪为基础,以计算机为核心,以虚拟仪器编程语言LabWindows/CVI为软件平台,以现有自动焊接装置为工件装夹平台,借助于计算机技术、超声无损检测与评价技术、机电一体化技术,成功地开发了一个超声自动检测系统,实现了液压缸环焊缝检测的数字化、自动化和图像化,提高了超声检测的可靠性和效率。研究了横波斜探头超声扫描成像的原理和方法,给出了斜探头超声B、C扫描成像的定义,建立了人工缺陷超声B扫描和C扫描成像的数学模型,解决了超声扫描成像软件实现的关键技术问题,实现了超声B扫描和C扫描成像功能。针对焊接缺陷超声回波信号的特点,深入研究了小波阈值去噪法的原理和方法,在软、硬阈值函数的基础上,构造了一种新的阈值处理函数,提出了基于改进阈值函数的提升小波变换去噪算法。实验结果表明,该方法改进了软、硬阈值去噪法的缺点,获得了更好的去噪性能和更高的信噪比,而且在实时信号去噪方面具有很好的应用前景。以焊接缺陷的常见类型为对象,研究了基于小波包变换的多尺度空间能量特征提取方法,对实测的超声缺陷回波信号进行了特征提取实验,并利用基于距离的类别可分性判据对提取的特征向量进行了评价。结果表明,该方法对焊接缺陷超声回波信号的特征提取是有效的。为了克服人工神经网络需要大量训练样本,且网络结构确定困难等缺点和不足,在对比分析支持向量机常用多值分类方法的基础上,提出了基于ν-SVM的二叉树多值分类方法,实验结果表明其分类正确率高,在训练时间和测试时间方面均优于其它的支持向量机多值分类方法。基于RBF神经网络与支持向量机的结构等价性,提出了基于ν-SVM-BTMC的RBF神经网络焊接缺陷识别方法,实验结果表明该方法比RBF神经网络、基于ν-SVM的二叉树多值分类方法的分类正确率更高,其训练和测试速度最快。更多还原

【Abstract】 Hydraulic support is a main equipment of the comprehensive mechanization mining, and it is of great significance for modern coal mine production. Hydraulic cylinder is a key actuator which completes various motions and bears roof pressure. Its quality directly affects the reliability of hydraulic support, and then affects safety and normal production of coal mine.Aiming at the problem of girth weld quality test for hydraulic cylinder, in order to conquer the disadvantages of manual ultrasonic A-scan testing and ensure weld quality, taking digital ultrasonic flaw detector as base, computer as core, virtual instrument programming language of Lab Windows/CVI as software platform and existing automatic welding equipment as clamping workpiece platform, an ultrasonic automatic testing system has been successfully developed by means of technologies of computer, ultrasonic NDT and NDE, integration of mechanics and electrics. This system has implemented numeralization, automatization and imaging of girth weld testing for hydraulic cylinder, and enhanced efficiency and reliability of ultrasonic testing.The principle and method of ultrasonic scan imaging for transverse wave angle probe are studied, the definitions of ultrasonic B-scan and C-scan imaging for angle probe have been given, the mathematical models of ultrasonic B-scan and C-scan imaging for artificial defect have been established, the key technique problems of software implementation of ultrasonic scan imaging have been achieved, and the functions of ultrasonic B-scan and C-scan imaging have been realized.According to the characteristics of ultrasonic echo-signals of weld flaws, the principle and algorithm of wavelet on threshold de-noising are further studied. On the basis of soft and hard threshold fuctions, a new threshold function is structured, and a lifting wavelet transform (LWT) de-noising algorithm based on the improved threshold function has been put forward. Experimental results show that this algorithm has overcome the disadvantages of soft and hard threshold de-noising methods, and obtained a better de-noising performance and a higher signal-noise ratio (SNR). Moreover, this algorithm has a good application prospect in real-time signal de-noising.Taking the common types of weld flaws as object, the method of energy feature extraction based on wavelet packet transform (WPT) are studied. The feature values of actual measured ultrasonic echo-signals are extracted by this method and evaluated by the sort separability criterion based on distance. The experimental result shows that this method is comparatively quite effective in the feature extraction of ultrasonic echo-signals for welding flaws.In order to overcome the disadvantages of artificial neural network, such as needing the massive trainings amples, the difficulty of determining network structure, and so on, through contrastively analyzing the common multi-class classification methods of support vector machine (SVM), a binary tree multi-class classification algorithm based onν-SVM (ν-SVM-BTMC) has been proposed. Experimental results show that the classification accuracy of this algorithm is high, and it is superior to other multi-class classification methods of SVM in the aspects of training and testing time.Considering the equivalence between RBF neural network (RBFNN) and SVM in structure, a new RBFNN recognition algorithm to welding flaws based onν-SVM-BTMC is proposed. From the experimental results, it is shown that the new recognition algorithm has a higher classification accuracy than RBFNN and the binary tree multi-class classification algorithm based onν-SVM, moreover, its training and testing speed are fastest.更多还原