作者：杨德宸, 吴伟, 邬冠华, 陈曦

作者单位：南昌航空大学 无损检测技术教育部重点实验室，江西 南昌 330063

关键词：TC6锻件;背散射信号;微小缺陷回波;核主成分分析;集成学习

摘要：

针对TC6锻件超声检测，微小缺陷回波通常容易淹没在结构噪声和非声学噪声的背散射信号中，造成漏检的问题，提出基于集成学习的微小缺陷存在性判定方法。以含Φ0.6，Φ0.8，Φ1 mm盲孔试验件为研究对象，提取超声检测背散射信号的18个时频域特征参数构成检测数据集；利用核主成分分析（kernel principal component analysis，KPCA）对数据集进行降维，选择累计贡献率达到91.1%的前3个主成分作为信号分类特征向量；分别使用BP神经网络、支持向量机（support vector machine，SVM）、决策树、概率神经网络（probabilistic neural network, PNN）建立单分类模型，并采用集成学习方法，通过加权投票机制对4种分类模型进行融合。测试样本试验结果表明，与单分类模型相比，融合后的模型分类准确率为95.8%，能够有效识别TC6锻件中是否存在微小缺陷，准确率优于单分类模型。

Research on ultrasonic backscattering model for TC6 forging small defects based on KPCA ensemble learning
YANG Dechen, WU Wei, WU Guanhua, CHEN Xi
Key Laboratory of Nondestructive Testing Technology (Ministry of Education), Nanchang Hangkong University, Nanchang 330063, China
Abstract: For TC6 forgings ultrasonic testing, the echo of micro defects is usually submerged in the backscattered signal of structural noise and non acoustic noise, which is easy to cause the problem of missing inspection. A method for judging the existence of micro defects based on ensemble learning is proposed. Taking the test pieces with blind holes of Φ 0.6, Φ 0.8 and Φ 1 mm as the research object, 18 time-frequency characteristic parameters of ultrasonic back scatter signal are extracted to form the detection data set, and the kernel principal component analysis (KPCA) is used to analyze the ultrasonic back scatter signal The first three principal components with cumulative contribution rate of 91.1% were selected as the feature vectors of signal classification, and BP neural network and support vector machine (SVM), decision tree and probabilistic neural network (PNN) are used to build a single classification model, and the four classification models are fused by weighted voting mechanism using ensemble learning method. The test results of test samples show that, compared with the single classification model, the classification accuracy of the fusion model is 95.8%, which can effectively identify whether there are small defects in TC6 forgings, and the accuracy rate is better than that of the single classification model.
Keywords: TC6 forging;back scatter signal;micro defect echo;kernel principal component analysis;ensemble learning
2021, 47(10):95-102  收稿日期: 2020-09-16;收到修改稿日期: 2020-11-26
基金项目: 无损检测教育部重点实验室（南昌航空大学）开放基金（EW201708505）；南昌航空大学研究生创新专项基金（YC2019040）
作者简介: 杨德宸（1996-），男，河北石家庄市人，硕士研究生，专业方向为无损检测技术
参考文献
[1] 航空用钛及钛合金棒材和锻坯规范: GJB 2218A—2018[S]. 北京: 国家军用标准出版发行部, 2018.
[2] Ultrasonic inspection titanium and titanium alloy bar, billet and plate: AMS 2631E—2017(R)[S]. 2017.
[3] 柯常波, 陈铁群. 粗晶材料超声无损检测信号处理技术发展趋势[J]. 物理测试, 2006(5): 37-41
[4] ROSE J L. Ultrasonic waves in solid Media[M]. Cambridge: Cambridge University Press, 1999.
[5] PRAVEEN A, VIJAVAREKHA K, ABRAHAM S T, et al. Signal quality enhancement using higher order wavelets for ultrasonic TOFD signals from austenitic stainless steel welds[J]. Ultrasonics, 2013, 53(7): 1288-1292
[6] 句海洋, 王新华, 赵以振. 改进VMD钢质管道损伤信号提取算法研究[J]. 中国测试, 2020, 46(5): 103-110
[7] 唐国维, 巩淼, 张方舟, 等. 基于模糊神经网络的焊缝缺陷识别的研究[J]. 计算机技术与发展, 2014, 24(5): 243-247
[8] 敬人可, 李建增, 周海林. 聚类分析在缺陷分类中的应用[J]. 国外电子测量技术, 2012, 31(10): 55-58
[9] 胡宏伟, 张婕, 彭刚, 等. 基于LBP-KPCA特征提取的焊缝超声检测缺陷分类方法[J]. 焊接学报, 2019, 40(6): 34-39+162
[10] 张世荣, 程琴, 张芳芳. 基于感测线圈和KPCA的电机轴承故障检测[J]. 电机与控制应用, 2018, 45(4): 98-104
[11] 李伟, 姜智通, 张璐莹, 等. 碳纤维复合材料损伤声发射信号模式识别方法[J]. 中国测试, 2020, 46(6): 121-128
[12] 蒋芳, 张南飞, 胡艳军, 等. 基于BP神经网络的CSI无源目标分类方法[J]. 北京邮电大学学报, 2020, 43(1): 40-45
[13] 刘云轩, 吴伟, 陈曦, 等. 基于概率神经网络的钛钢爆炸复合棒材拉剪性能分类[J]. 数据采集与处理, 2020, 35(2): 307-314
[14] 李大中, 张克延, 李昉. 大数据驱动下的风电机组工作状态决策树判别方法[J]. 电力科学与工程, 2020, 6(2): 22-27