【作者】 赵学风；

【导师】 段晨东；

【作者基本信息】 长安大学 ， 环境工程， 2007， 硕士

【摘要】 损伤诊断是进行结构健康监测的基础。本文以结构监测与损伤诊断为目的,研究了基于小波包分析和支持向量相结合的智能诊断方法。阐述了实施土木工程结构健康监测的必要性和迫切性,介绍了结构健康监测与损伤诊断的概念,讨论系统组成、损伤诊断的方法及其研究现状;研究了小波包变换的多分辨分析及其正交性。实验证明:(1)小波包分解频带能量分布能够表征信号分量的能量变化。信号经小波包分解后其损伤特征更加明显,而且受噪声的干扰小;(2)小波支撑区间越大,正交性越好,信号分解到不同的频带中,使其各频带中的信息无冗余;(3)不同类型损伤的小波包能量分布是不同的;其次,对于同一种损伤,不同节点响应信号的小波包分解能量分布大小是不同的;(4)通过小波包变化处理,使得结构响应信号中的损伤特征更加明显,减小了噪声的干扰。因而可以作为一种理想的特征指标来表征结构健康状态。针对智能损伤诊断中损伤样本缺乏、分析处理的数据量大的问题,提出了一种基于小波包特征提取的支持向量机智能诊断方法。该方法将结构振动信号小波包分解后的频带能量作为特征,输入到多分类的支持向量机中,实现了结构多损伤的识别和定位。经过小波包特征提取后的的支持向量机分类效果明显优于未经过任何特征提取的支持向量机的分类效果。为了提高系统决策的准确性和鲁棒性,避免单一信号表征损伤信息的片面性,提出了另一种基于特征融合的支持向量机智能诊断方法。利用小波包变换提取小波包特征进行多点数据融合,构造特征矢量,将这些特征矢量输入到分类器中诊断损伤发生的位置和程度。数据融合不仅能够使损伤信息相互补充,而且减小了检测信息的不确定性;这种基于特征融合的支持向量机智能诊断方法充分利用各个数据源包含的冗余和互补信息,大大提高了分类的准确性。针对工程应用中损伤样本难以获得,并且为了实现结构健康状态的在线自动监测和诊断,提出了一种基于正常样本的支持向量数据描述单值分类新方法。该方法仅仅依靠正常运行时的数据信号,而不需要任何损伤数据。首先采用小波包分解对数据预处理,以频带能量序列为特征,然后把多测量点的能量序列融合后作为目标向量,输入到支持向量数据描述分类器进行训练,实现在线自动监测和损伤诊断。支持向量数据描述分类器较好地区分了结构正常与非正常状态,达到损伤自动诊断的目的。更多还原

【Abstract】 Damage diagnosis is the precondition of structural health monitoring (SHM). For the purpose of structure health monitoring and diagnosis, intelligent damage diagnosis approaches based on the wavelet packet analysis and support vectors (SVs) are studied in this paper.The necessity of civil engineering SHM is firstly discussed. Secondly, the concept of SHM & damage diagnosis, and the architecture of SHM system are introduced. Then, structure damage diagnosis techniques and their development are reviewed. In order to extract damage features from noisy signal, the principle of multi-resolution analysis and orthogonality of wavelet packet transform (WPT) are investigated. Experimental results shown: (1)The wavelet packet energy distribution can indicate the component energy variation in a signal. (2) The bigger supported intervals an orthogonal wavelet has, the better orthogonality the wavelet possesses. It is helpful to isolate the components into different sub frequencybands and have less redundancy in these frequencybands. (3) Signals form different kinds of damage reveal different packet energy distribution through WPT, and for a special damage the wavelet packet energy distribution is different at different measurement nodes. (4)Demage features embedded in structure response signal can be distinctly clarified with the wavelet packet energy distribution which is robust to noise. It can be used as an ideal feature index to represent the structural health condition.In order to solve problems of faulty sample shortage and proceesed data overabundance in damage diagnosis, an intelligent method using support vector machine (SVM) is proposed by means of extracting feature with WPT. According to the method, the energy sequences of different frequency bands decomposed by WPT are investigated, which are input to a multi-classified support vector machines to implement multi-damage recognition and damage localization. The classification accuracy of the proposed method is greatly improved compared with the multi-classification SVM without feature extraction. In order to enhance the accuracy and robustness of the system decision-making, and avoid the information unintegrity of only using the sole signal in damage diagnosis, another SVM diagnosis method based on the data fusion technique is put forward. Constructing damage feature vectors with extracted features from several measurement nodes using WPT, and inputing these feature vectors to a SVM classifier, degree and localization of damages are found. The diagnosis information was enriched by means of data fusion, and the uncertainty of damage detection information was also depressed. Since diagnosis information from the measurement nodes is redundant and supplementary, the diagnosis accuracy is greatly improved.Because it is difficult to obtain the damage samples in engineering application, and for the purpose of on-line automatic monitoring and diagnosis for structural health condition, a new monitoring and diagnostics method based on support vector data description (SVDD) is proposed, which only needs samples under normal condition, and needs no abnormal samples. WPT is also used to preprocess original signals, the signal energies in different frequency-bands are taken as condition feature. Then the features from different measurement nodes are fused as a target vector. A developed SVDD classifier is applied to implement structural condition monitoring by inputting the target vector. SVDD classifier was able to distinguish the normal and abnormal condition of structure ideally, and can be used as an automation approach for monitoring and diagnosis.更多还原