|
| Abstract: |
| A feature extraction method was proposed to sectorial scan image of Ti-6Al-4V electron beam welding seam based on principal component analysis to solve problem of high-dimensional data resulting in time-consuming in defect recognition. Seven features were extracted from the image and represented 87.3 % information of the original data. Both the extracted features and the original data were used to train support vector machine model to assess the feature extraction performance in two aspects: recognition accuracy and training time. The results show that using the extracted features the recognition accuracy of pore, crack, lack of fusion and lack of penetration are 93%, 90.7%, 94.7% and 89.3%, respectively, which is slightly higher than those using the original data. The training time of the models using the extracted features is extremely reduced comparing with those using the original data. |
| Key words: electron beam welding phased array ultrasonic sectorial scan image feature extraction principal component analysis |
| DOI：10.11916/j.issn.1005-9113.15287 |
| Clc Number:TG115.28 |
| Fund: |
|
| Descriptions in Chinese: |
| 基于主成分分析的厚板电子束焊缝超声相控阵 扇形扫描图像特征提取 刚铁1，栾亦琳2，张弛1 (1.哈尔滨工业大学 先进焊接与连接国家重点实验室，哈尔滨 150001； 2.黑龙江科技大学 材料科学与工程学院，哈尔滨 150022) 摘要：为解决缺陷识别模型训练时高维数据引起的耗时巨大问题，提出一种基于主成分分析的特征提取算法。从Ti-6Al-4V厚板电子束焊缝超声相控阵扇形扫描图像中提取出7个特征值，代表了原始数据87.3%的信息量。将提取的特征值和原始数据均用于训练支持向量机缺陷识别模型，从缺陷识别准确性和训练时间两个方面评价特征提取算法的有效性。结果表明，采用特征值训练的缺陷识别模型，气孔、裂纹、未熔合和未焊透的识别率分别为93%、 90.7%、94.7% 和 89.3%，略高于采用原始数据训练的模型。采用特征值的模型训练时间相比于采用原始数据的模型训练时间大大降低了。 关键词：电子束焊缝；超声相控阵；扇形扫描图像；特征提取；主成分分析 |
