期刊
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
卷 66, 期 7, 页码 1883-1892出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TIM.2017.2673024
关键词
Convolutional neural network (CNN); defect identification; injurious or noninjurious; magnetic flux leakage (MFL); nondestructive testing
资金
- National Natural Science Foundation of China [61273164, 61433004, 61473069, 61673093, 61627809, 61621004]
This paper proposes an injurious or noninjurious defect identification method from magnetic flux leakage (MFL) images based on convolutional neural network. Different from previous approaches, this method is fed by the MFL images instead of the features of the MFL measurements, and thus it can skip the procedure of feature extraction. Moreover, for convenience, a normalization layer is added to the front of model. In the convolution layers, the rectified linear units are employed as the activation functions to shorten the training period and improve the performance. In addition, two local response normalization layers are also embedded into the proposed structure. We demonstrate the performance of the proposed model using real MFL data collected from experimental pipelines. Benefited from the special structure of the proposed model, this method is robust for shift, scale, and distortion variances of input MFL images. We also present a comparative result of the proposed model and other methods. The results prove that the proposed method can achieve higher accuracy than the traditional approaches.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据