Defects detection in metallic additive manufactured structures utilizing multi-modal laser ultrasonic imaging integrated with an improved MobileViT network

Abstract

The multi-modal coupling of the laser generated ultrasonic waves in metallic additive manufacturing poses significant challenges for defects detection when using traditional C-scan imaging methods. This paper proposes an improved MobileViT-based intelligent method for defects detection using laser ultrasonic C-scan imaging. First, the Efficient Channel Attention is integrated into the inverted residual block to enhance the prominent features in the down-sampled feature maps. Second, a Receptive-Field Attention Convolution is introduced to dynamically assign convolutional kernel weights based on the significance of image features, enhancing the model’s capability to capture global image features. When utilizing C-scan image sequences from metal additive manufactured structures, the modified MobileViT network achieves a defect recognition accuracy of 98.31%. In addition, the proposed network also shows good classification results on the public NEU-CLS dataset, surpassing the comprehensive performance of EfficientNetB1, ShuffleNetV2, et al. This result shows that the improved MobileViT network offers a promising solution for defect detection precisely in metal additive manufacturing, which has potential for online inspection applications in the future.