Abstract
In this paper, we focus on eddy current array (ECA) technology for defect detection in finely grooved structures of spinning cylinders, which are significantly affected by surface texture interference, lift-off distance, and mechanical dither. Unlike a single eddy current coil, an ECA, which arranges multiple eddy current coils in a specific configuration, offers not only higher accuracy and efficiency for defect detection but also the inherent properties of space and time for signal acquisition. To efficiently detect defects in finely grooved structures, we introduce a spatiotemporal self-attention mechanism to ECA testing, enabling the detection of defects of various sizes. We propose a Multi-scale SpatioTemporal Self-Attention Network for defect detection, called MSTSA-Net. In our framework, Temporal Attention (TA) and Spatial Attention (SA) blocks are incorporated to capture the spatiotemporal features of defects. Depth-wise and point-wise convolutions are utilized to compute channel weights and spatial weights for self-attention, respectively. Multi-scale features of space and time are extracted separately in a pyramid manner and then fused to regress the bounding boxes and confidence levels of defects. Experimental results show that the proposed method significantly outperforms not only traditional image processing methods but also state-of-the-art models, such as YOLOv3-SPP and Faster R-CNN, with fewer parameters and lower FLOPs in terms of Recall and F1 score.