基于Kohonen网络的激光超声缺陷探测研究

在激光超声缺陷检测技术中,不同类型缺陷采样信号的准确分类至关重要.针对激光超声表面波实验采样信号高维小样本的特点,采用了一种有监督学习的Kohonen神经网络(S_Kohonen)自适应分类方法.在S_Kohonen网络自组织学习的过程中,通过改进网络的学习率提高了网络的收敛速度.通过采用一种无需邻域半径判断的自适应权值调整方式来实现竞争层神经元权值不同程度的调整,从而更有效的表征输入样本的分布特征.通过对不同类型缺陷探测样本的多次实验,验证了所述方法具有良好的分类预测效果,多次交叉验证分类正确率均能达到100%.

Defect Detection Using Laser Ultrasound Based on Kohonen Network

The accurate classification for different types of sampled defects signal is critical in defect detection technique with laser ultrasonic.According to the characteristics of small sample size and high-dimensional in the laser ultrasonic experiments,a supervised learning Kohonen network(S_Kohonen)method is employed to achieve adaptive classification.In the learning of S_Kohonen,the convergence speed of the network is promoted by improving the learning rate.A new adaptive way to update weight vectors of competitive layer neurons without any neighborhood radius judgment is adopted that can characterize the distribution of input samples more effectively.Several experiments for different types of defect detection validates that the proposed method has good classification prediction results,and the classification accuracy rate can reach 100%in multiple cross-validation experiments.