Eddy Current Array for Defects Detection based on Spatiotemporal Self-attention Network

Abstract

Unlike single eddy current coil, eddy current array (ECA) which arranges multiple eddy current coils in a certain way, has the property of higher accuracy and efficiency to detect defects. The process of eddy current array collecting data own naturally spatial and temporal characteristics. In this paper, we introduce spatiotemporal self-attention mechanism to ECA Testing, and propose a spatiotemporal self-attention network for defect detection. In our framework, features from different channels are extracted separately and fused together by Downsampling Residual Attention Modules (DRAM) and Residual Attention Modules (RAM) in a pyramid manner, in which temporal attention module (TA) and spatial attention module (SA) are incorporated to capture spatiotemporally the features of defects. And the depth-wise and point-wise convolution are utilized to compute channel weights and spatial weights in TA and SA modules, respectively. Multiple channel data is taken as input from ECA, which finally leads to a classification result. The experimental results show that the proposed method not only outperforms the traditional image processing method significantly, but also is better than the state of the arts - ResNet, DenseNet in terms of F1 and accuracy.