Inspection of defects in weld has both safety and economic significance in industries. However, it is still a challenging problem to be studied due to the interference of the uneven surface of the weld and nonuniformity of the material. This article proposes a new flexible eddy current testing (ECT) probe with differential multimodes for weld inspection. The probe is constructed based on a flexible printed circuit board (PCB). Consequently, it can be bended according to the uneven surface of a weld sample. The differences of pairs of coils are recoded as the outputs of the probe, in which the common mode noises and background signals are canceled. Considering the different orientations of the differential pair coils, the probe has four modes. Different modes complement each other in terms of sensitivity to different kinds of defects. The effect of the excitation frequency has been studied, and 1 MHz is chosen as the operating frequency of the probe. An image preprocessing algorithm called spatial domain filtering and gradient feature edge extraction (SDF-GFEE) is proposed for suppressing the noise of the experimental images. A target detection algorithm based on deep neural network named mask region convolution neutral network (Mask-RCNN) is utilized to identify defects from the images. Experimental results demonstrate that this method is capable of identifying the groove defect with dimensions 3 mm (length) $\times0.1$ mm (width) $\times0.5$ mm (depth) and flat bottom hole with dimensions $\Phi 0.8$ mm $\times0.5$ mm in a weld sample.

中文翻译：

---

使用差分阵列ECT探针和深度学习算法检测焊缝缺陷

焊缝缺陷检测在工业上具有安全和经济双重意义。然而，由于焊缝表面不平整和材料不均匀的干扰，这仍然是一个需要研究的具有挑战性的问题。本文提出了一种用于焊缝检测的具有差分多模式的新型柔性涡流检测 (ECT) 探头。探头基于柔性印刷电路板 (PCB) 构建。因此，它可以根据焊接样品的不平整表面进行弯曲。线圈对的差异被记录为探头的输出，其中消除了共模噪声和背景信号。考虑到差分对线圈的不同方向，探头有四种模式。不同模式在对不同类型缺陷的敏感性方面相互补充。研究了激发频率的影响，选择1 MHz作为探头的工作频率。提出了一种称为空间域滤波和梯度特征边缘提取（SDF-GFEE）的图像预处理算法来抑制实验图像的噪声。一种基于深度神经网络的目标检测算法，称为掩膜区域卷积中性网络（Mask-RCNN），用于识别图像中的缺陷。实验结果表明，该方法能够识别尺寸为3 mm（长度）的凹槽缺陷 提出了一种称为空间域滤波和梯度特征边缘提取（SDF-GFEE）的图像预处理算法来抑制实验图像的噪声。一种基于深度神经网络的目标检测算法，称为掩膜区域卷积中性网络（Mask-RCNN），用于识别图像中的缺陷。实验结果表明，该方法能够识别尺寸为3 mm（长度）的凹槽缺陷 提出了一种称为空间域滤波和梯度特征边缘提取（SDF-GFEE）的图像预处理算法来抑制实验图像的噪声。一种基于深度神经网络的目标检测算法，称为掩膜区域卷积中性网络（Mask-RCNN），用于识别图像中的缺陷。实验结果表明，该方法能够识别尺寸为3 mm（长度）的凹槽缺陷 $\times0.1 $ 毫米（宽度） $ \ 次0.5 $ mm（深度）和平底孔尺寸 $ \ 披0.8 $ 毫米 $ \ 次0.5 $ mm 在焊缝样品中。