基于多维度选择性搜索的小样本缺陷识别方法

超声内检测是油气管道缺陷的主要检测方式之一，目前超声内检测在工业小样本的情况下存在缺陷识别边界定位不准的问题。本文提出了一种基于多维度选择性搜索的小样本缺陷识别方法，该方法首先对超声回波进行特征提取，其中包含使用基于孤立森林的回波特征点提取，和基于自然断点法的特征点聚类；其次提出了风险相似性度量方法，并使用梯度提升树建立波形特征和风险程度的回归模型；然后将多维度缺陷相似性信息融合在选择性搜索算法中，实现小样本缺陷识别；最后使用异常分数等区域风险度量指标实现缺陷边界的精准定位。实验结果表明，本文设计的基于多维度选择性搜索的小样本缺陷识别方法的查全率和查准率分别高达95.08%和85.46%,能有效解决超声信号缺陷识别边界定位不准的问题。

Small sample defect recognition method based on multi-dimensional selective search

Ultrasonic internal inspection is one of the main defect detection methods for the oil and gas pipeline. At present, the location of the defect boundary is inaccurate in the case of small industrial samples for ultrasonic internal inspection. This article proposes a small sample defect recognition method based on multi-dimensional selective search. Firstly, ultrasonic echo features are extracted by two steps, which are feature point extraction based on isolated forest and feature point clustering based on the natural breaks classification method. Secondly, the risk similarity measurement method is proposed. A regression model of waveform characteristics and risk degree is formulated by the boosting tree. Thirdly, multi-dimensional defect similarity is integrated information into a selective search algorithm to realize small sample defect identification. Finally, regional risk metrics such as anomaly scores are used to achieve precise positioning of defect boundaries. Experimental results show that the recall and precision of the small sample defect recognition method based on multidimensional selective search are 95. 08% and 85. 46% , which can effectively solve the problem of inaccurate positioning of the ultrasonic signal defect boundary detection.