基于FLDA与BP神经网络的超声3D目标识别

Ultrasonic 3D Target Recognition Based on FLDA and BP Network

针对目前超声3D识别普遍存在的识别率低、鲁棒性差等问题,以物体内部人工标准缺陷为超声靶标,通过对超声靶标脉冲超声回波信号进行处理,提取了相对能量、相对幅值、相对频域带宽、相对峰度系数、相对离散系数、相对包络面积、相对偏度系数和相对频谱半高宽等多个特征参数,利用Fisher线性判别分析(Fisher Linear Discriminative Analysis,FLDA)对这些特征参数进行融合,形成融合特征,并采用反向传播(Back Propagation,BP)神经网络对融合特征进行训练与识别,对物体内部矩形槽、横通孔及平底孔三类超声靶标进行识别。试验结果表明:三种靶标的识别率分别高达了93.3%,93.3%,100%;对噪声有抑制能力,对测试工况不敏感,识别稳健性得到了提高,可为超声3D目标识别提供理论和技术参考。

The technique has a broad prospect of applications in nondestructive testing and evaluation. However, low recognition rate and bad robustness have been faced with at present in ultrasonic 3D target recognition. Aiming to these problems, this paper proposed an effective method. The object inner artificial discontinuities were as the ultrasonic targets, and a few characteristic parameters were extracted by means of dealing with the pulse ultrasonic echo signals of the ultrasonic targets, such as relative energy, relative amplitude, relative frequency domain bandwidth~ relative coefficient of kurtosis, relative coefficient of variation, relative envelope area, relative coefficient of skewness and relative FWHM of frequency spectrum, then, FLDA was utilized to fuse the characteristic parameters and integration features were discovered, finally, BP neural network was used to train and recognize integration features. The recognition rate of three kinds of ultrasonic targets, which were object interior rectangular groove, horizontal hole and flat-bottom hole, achieved 93. 3%, 93. 3% and 100% respectively. The experimental result shows that the method is quite effective, having inhibitory ability to noise, not sensitive to test condition and robust. It can provide theoretical and technical reference to ultrasonic 3D target recognition.