【作者】 王卫平；

【导师】 周辉林；

【作者基本信息】 南昌大学 ， 通信与信息系统， 2009， 硕士

【摘要】 随着高速公路通车里程的不断增加和路网规模的不断扩大以及使用年限的不断增长,高速公路的养护工作变得日趋繁重和重要。探地雷达（groundpenetrating radar,GPR）作为一种快速、连续、安全、高精度、高分辨率的无损实时探测工具,越来越广泛地应用于高速公路路面浅层病害检测。但是,GPR跟光学成像设备不同,它不能直接反映目标的特征。因此,使用GPR勘查高速公路路面浅层质量时,如何由获取的GPR数据解释高速公路路面浅层质量状况成为问题的关键。本课题结合现代数字处理、信号检测技术及模式识别算法,实现了高速公路路面浅层病害的自动检测,其主要技术手段及具体研究内容为:1、探地雷达原始数据预处理方法的研究,主要内容包括运用滑动平均法对GPR原始数据进行噪声抑制、使用包络检波器和阈值检测技术自动检测路面浅层层界面、使用Savitzky-Golay滤波器进行层界面平滑,以及ROI（region ofinterest）提取。2、基于时域和小波域回波信号的特征提取。总共在时域与小波域内提取了6个特征,即时域的三个特征:信号的最大幅值MAX_s、信号幅值的平均绝对偏差MAD_s、原始信号幅值互相关XCORR_s;小波域的三个特征:合成信号在各级小波上的互相关之和XCORR_d123、第三级小波近似系数a3的最大幅值M-AX_a3、第三级小波近似系数a3的平均绝对偏差MAD_a3。3、基于支持向量机（support vector machine,SVM）的路面浅层病害检测。利用专家经验,从层界面反射信号中提取出好路面和坏路面（有病害路面）样本信号,划分成训练样本和测试样本。提取其6个时域小波域特征,用训练样本训练SVM,得到相应的支持向量网络,并把测试样本作为输入,由SVM进行特征分类,从而检测出路面浅层是否存在病害。更多还原

【Abstract】 The maintenance of highway has been getting more and more important and labor-intensive because roadway network system is extending, and most of built roads are aging. As a fast, continuous, secure, nondestructive, real-time detection tool with high precision, GPR （ground penetrating radar） has been used in highway pavement distress detection. However, GPR is different from optical imaging equipment as it is not able to reflect the feature of object directly. Thus, during GPR is used to survey the quality of highway pavement, how to interpret the acquired GPR data as the quality status of highway pavement becomes the key issue.Using modern digital processing, signal detection and pattern recognition algorithm, highway pavement distress has been detected automatically in this project. The main techniques and detailed research work are list out as follows:1. Research on GPR original data preprocessing algorithm for clutter suppression, layer interface detection, layer interface smoothing, and ROI extraction.2. Research on feature extraction algorithm in time domain and wavelet domain. Extract three features from time domain: maximum amplitudes （MAX_S）, mean absolute deviation （MAD） of amplitudes （MAD_S）, and cross-correlation of original signals amplitude （XCORR_S）. And extract another three features from wavelet: summation of cross-correlation of synthesized signals at all level of wavelet (XCORR_d123), magnitude of wavelet approximate coefficient a₃ (MAX_a3), and MAD of approximate coefficient a₃ (MAD_a3).3. Research on pavement distress detection algorithm based on SVM. Basing on expert experiences training samples and testing samples are selected from GPR reflected signals in good and deteriorative pavement. Train SVM with six extracted features of training samples obtaining support vector network correspondingly, then input testing samples into SVM and implement feature classification determining whether there is distress in pavement or not.更多还原