改进FasterR-CNN的大型铸造不锈钢机匣超声相控阵检测图像的缺陷智能识别  

Improved Faster R-CNN for the intelligent recognition of defects in ultrasonic phased array inspection images of large cast stainless steel casing

作　　者：赵玉琦 李婧 董德秀 黄鑫章 陈振华[1] 卢超[1] ZHAO Yuqi;LI Jing;DONG Dexiu;HUANG Xinzhang;CHEN Zhenhua;LU Chao(Key Laboratory of Nondestructive Testing of Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China;AECC Shenyang Liming Aero-Engine Co.,Ltd.,Shenyang 110043,China)

机构地区：[1]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [2]中国航发沈阳黎明航空发动机有限责任公司,沈阳110043

出　　处：《应用声学》2025年第2期497-504,共8页Journal of Applied Acoustics

基　　金：国家科技重大专项(J2019-VII-0002-0142);国家市场监督管理总局科技计划项目(2021Mk065)。

摘　　要：大型铸造不锈钢机匣的超声相控阵检测技术具有检测能力强、检测效率高的优势。然而,相控阵图像中显示的缺陷类型仍需检测人员判读,存在主观性强、易误判、效率低、可靠性不足等问题。据此,提出基于深度学习的机匣超声相控阵检测图像缺陷类型的自动识别方法。首先,采集机匣典型铸造缺陷的超声相控阵图像,对缺陷图像扩充并制备数据集;其次,在FasterR-CNN深度学习网络的特征提取网络、多层特征信息融合网络、感兴趣区域模块等方面进行优化改进;最后,对比分析改进前后深度学习网络模型的缺陷识别与分类准确率。结果表明:相比于原始FasterR-CNN深度学习网络,在采用深度残差网络、特征金字塔网络、区域一致性池化等优化措施后,平均准确率均值提高至95.3%,模型对缺陷图像的识别精度得到了有效的提高;改进的FasterR-CNN目标识别算法克服了超声相控阵缺陷图像人工识别与分类的问题,具有较好的工程应用价值。The ultrasonic phased array testing technology for large stainless steel machine boxes has the advantages of strong detection capability and high detection efficiency.However,the defect types displayed in the phased array images still need to be judged by the testing personnel,leading to problems such as strong subjectivity,easy misjudgment,low efficiency,and lack of reliability.In response to this,an automatic identification method for detecting image defect types of machine box ultrasonic phased array based on deep learning is proposed.Firstly,ultrasonic phased array images of typical casting defects of the machine box are collected,and the defect images are augmented to prepare the dataset.Secondly,optimization and im provement are carried out in the feature extraction network,multi-layer feature fusion network,and region of interest module of the Faster regions with convolutional neural network(R-CNN)deep learning network.Finally,the defect recognition and classification accuracy of the deep learning network model before and after the optimization are compared and analyzed.The results show that,compared with the conventional Faster R-CNN deep learning network,the average accuracy has increased to 95.3% after the use of optimization measures such as deep residual network,feature pyramid network,and region consistency pooling,and the model’s accuracy in identifying defect images has been effectively improved.The improved Faster R-CNN target recognition algorithm overcomes the problem of manual recognition and classification of phased array defect images and has good engineering application value.

关 键 词：超声相控阵检测 改进FasterR-CNN 缺陷智能识别 

分 类 号：TG115.28[金属学及工艺—物理冶金]