基于改进半监督FCM聚类的复合材料超声检测脱粘缺陷识别

针对径向基（RBF）神经网络在进行超声检测脱粘缺陷识别时存在参数选择不确定、网络结构鲁棒性差等问题，提出一种改进的自适应半监督模糊C均值聚类（FCM）的RBF神经网络的方法，将kw近邻估计法和半监督模糊C均值聚类方法相结合，改进了隶属度函数，自适应确定聚类数目。将改进的RBF神经网络应用于超声检测脱粘缺陷识别，实验结果表明：与传统RBF神经网络相比，本方法减弱了孤立样本对网络结构的影响，增强了网络结构的鲁棒性，提高了脱粘缺陷识别的准确率，是一种较好的超声检测脱粘缺陷识别分类方法．

Recognition of Composite Materials Un-Bond Flaws within Ultrasonic Testing Based on Improved Semi-Supervised FCM Clustering

An improved radial basis function （RBF） neural network based on adapting semi-super vised fuzzy c-means is proposed to choose the parameters and has a robust structure. The method firstly combines the kN amphictyonic estimate with semi-supervised fuzzy c-means to improve the membership degree function;then determines the threshold to make sure the clustering members. The proposed approach is applied to recognize ultrasonic signals of un-bond defects, and the ex- perimental results show that the proposed approach can lower the effect of the isolated point on the network structure and can be more robust while improving the accuracy rate. The proposed RBFNN approach is a feasible recognition algorithm for ultrasonic signals.