Multi-scale neighborhood query graph convolutional network for multi-defect location in CFRP laminates

Abstract

This paper presents a novel deep learning architecture named multi-scale neighborhood query graph convolutional network (MNQGN). In MNQGN, the spatial relationship between sensors is represented by constructing a vibration sensor distribution map. Furthermore, MNQGN enhances the feature representation by integrating a neighborhood query interaction mechanism to learn features from different scales. This approach better captures the common characteristics between multiple defects and reduces the impact of directional and anisotropic properties of composite materials. In addition, the spatio-temporal features in the collected signal data are extracted by the multi-scale spatio-temporal attention module in MNQGN, which focuses more on the time steps relevant to multi-defect localization tasks, thereby improving the accuracy and efficiency of the model. To evaluate the performance of MNQGN and demonstrate its effectiveness, experiments are conducted to test its ability to accurately identify and locate multiple defects within carbon fiber reinforced plastic laminates. The results show that MNQGN achieves a high accuracy rate of 97.90 % in multi-defect detection and localization, which is at least 1.67 % higher than the existing models.