基于选择性集成学习的焊接缺陷识别研究  

Study of welding defect recognition algorithm based on selective ensemble learning

作　　者：徐桂云[1,2] 陈跃[1,3] 张晓光[1,2] 刘云楷[1] 

机构地区：[1]中国矿业大学机电工程学院,江苏徐州221116 [2]哈尔滨工业大学现代焊接生产技术国家重点实验室,黑龙江哈尔滨150001 [3]徐州工程学院机电工程学院,江苏徐州221008

出　　处：《中国矿业大学学报》2011年第6期949-953,共5页Journal of China University of Mining & Technology

基　　金：现代焊接生产技术国家重点实验室开放课题;江苏省高技术研究项目(BG2007013)

摘　　要：针对射线检测焊缝图像中缺陷识别正确率低的问题,提出一种选择性集成学习的焊接缺陷识别算法.算法中的个体学习器由稳定分类器和非稳定分类器组成,使用SVM-RFE算法移除集成学习器中的冗余个体学习器,保留子学习器预测输出加权作为集成学习器的输出,有效地增强了个体之间的差异性,进而提高了集成的泛化性能.结果表明:该算法充分利用更多的缺陷特征和样本数据集信息,继承了强集成学习的优点,有效地提高分类正确率.使用一对多的方法把二分类选择性集成学习器推广到多分类问题中,所提出的算法在训练精度为92.4%时;焊缝缺陷识别率提高到85.5%.Aiming at the unsatisfied recognition rate of welding line defect in ray detection,a welding defect recognition algorithm based on selective ensemble learning is presented.The component learners consisted of stable and unstable classifiers,and the redundant component learners were removed from the ensemble learners using a support vector machine for recursive feature elimination（SVM-RFE）,and the weighted predicted output of the retained sub-learners was used as the output of the ensemble learners,which effectively improved both the diversity of component learners and generalization performance of ensemble learners.The results show that the method made use of more defect features and sample data,inherited the advantages of strong ensemble learners,and effectively improved the classification accuracy.The two-category selective ensemble learners were extended to multi-classification problems using one-to-many method,and the welding-defect-recognition rate of the presented algorithm was improved to 85.5% when the training accuracy attained 92.4%.

关 键 词：焊接缺陷 分类 选择性集成学习 支持向量机 K-NN分类器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]