混凝土结构缺陷的融合识别研究

Research on Identification of Concrete Flaws by Information Fusion

混凝土结构缺陷的无损检测是一项非常困难的工作 ,尤其是小尺寸或浅层的缺陷检测非常困难。由于仅用一种检测方法往往难以给出令人信服的结论 ,因此 ,本文同时采用超声检测和脉冲回波检测两种方法对不同大小的剥离和空洞缺陷进行了探测 ,并用小波分析方法对这两种信号进行了特征抽取 ,以第三阶尺度上的极大模作为信号的特征向量 ;随后 ,用一个多层前馈神经网络进行了单种检测方法的软决策 ,定义并计算了以该软决策为基础的概率分配函数 ;最后 ,用证据理论方法进行了两种检测方法的决策级融合识别 。

Nondestructive flaw detection of concrete structures is a difficult task, especially for detection of small or shallow flaws. In this research, both ultrasonic tests and impact-echo tests are used for detection of simulated flaws of different sizes because only one kind of detecting method can not give out a believable conclusion sometimes. After signals are collected, wavelet analyses are used for feature extraction from these two kinds of signals. Maximum-modulus of the third scale component are assigned as feature vectors. Then a feed-forward multi-layer neural network is used to implement local soft classification. A fuzzy membership assignment is presented. Probability mass based on fuzzy membership is constructed. After that, Shafer-Dempster reasoning is used for decision-level identity fusion and the hard decision of flaw detection is made. The classification results showed that recognition by information fusion is better than by only any one of the two tests.