Abstract: For solving the problem that fixed classification model, limited training set size and difficulty to reflect the differences of different defects and polymorphism of similar defects which conventional ultrasonic weld defect identification method exists, combined with the condition of data analysis strategy under the big data environment and the ideas of the match in genetic defect identification, the principal component analysis and CURE clustering algorithm convert the defect echo signal to the object of the matching, then the current detecting defect characteristics can match with the historical data, the nearest neighbor method implements the expansion of the historical defects data set. Through the application of Smith-Waterman alignment algorithm based on the basic space penalty on R, the simulation results verify that the defect identification method is feasible and effective to identify the porosity, slag inclusion, cracks, incomplete penetration and incomplete fusion five kinds of common defects, the method has good identification accuracy.

Key words: defects identification, principal component analysis, Clustering Using Representatives（CURE） clustering algorithm, alignment algorithm

摘要： 为了解决常规超声波焊缝缺陷识别方法分类模型固定和训练集规模有限而难以体现不同缺陷的差异性和同类缺陷的多态性的问题，结合当今大数据环境下的数据分析策略和基因缺陷识别中匹配的思想，通过主成分分析和CURE聚类算法将缺陷回波信号编码转换成可进行匹配的对象，进而将当前检测缺陷特征与历史检测数据进行匹配，并利用最近邻方法实现了对缺陷历史检测数据集的扩充。通过在R上应用基于基本空位罚分的Smith-Waterman比对算法进行仿真实验验证了该缺陷识别方法是可行的，有效地识别了气孔、夹渣、裂纹、未焊透和未熔合五类常见缺陷，具有较好的识别准确率。

关键词: 缺陷识别, 主成分分析, CURE聚类算法, 比对算法