Reproduce “Depth Evaluation for Metal Surface Defects by Eddy Current Testing using Deep Residual Convolutional Neural Networks” on OpenResearch

About this paper

Tian Meng, Yang Tao, Ziqi Chen, Jorge R. Salas Avila, Qiaoye Ran, Yuchun Shao, Ruochen Huang, Yuedong Xie and 5 more

A team from The University of Manchester engineered a portable Eddy Current Testing device and assembled the open-access MDDECT dataset, which they then used to train 1-D residual Convolutional Neural Networks for automated metal surface defect depth evaluation. Their ResNeXt1D-38 model achieved 93.58% accuracy in classifying defect depths with 0.1 mm resolution and showed strong resilience to lift-off variations.

Key insights

• Deeper 1-D residual CNN architectures, particularly ResNeXt, are more effective than wider or shallower networks for accurately classifying complex ECT time-series data.
• The incorporation of real-world variances in the MDDECT dataset, such as untrained operators and varying lift-offs, is crucial for training deep learning models that are robust in practical scenarios.
• Class Activation Mapping (CAM) provides valuable interpretability, showing how the models focus on specific signal regions to discern subtle differences for precise depth classification.

Summary by alphaXiv.