The primary objective of this study is to transform the status quo of laser-welding defect detection. By developing an online deep learning system, this study aims to enable the identification and measurement of surface defects in laser-welded aluminum-alloy sheets with high precision and efficiency. The specific focus is on two prevalent defects: undercuts, characterized by the insufficient melting of the base material at the weld toe, and sagging, which is the undesirable downward displacement of the material along the weld seam. The use of high-density point cloud data is key to overcoming the limitations of traditional defect detection methods and enhancing the adaptability of the system to diverse welding conditions.

A binocular-structured light sensor capable of capturing detailed point cloud data of defects in laser-welded samples is used. This sensor is strategically positioned to cover the entire welding area,which ensures the collection of comprehensive defect data. The acquired point cloud data undergo meticulous preprocessing to eliminate noise and artifacts, resulting in a clean and informative dataset. The dataset serves as the foundation for training the faster region-based convolutional neural network (Faster R-CNN) model, a deep learning architecture renowned for its object detection capabilities. The Faster R-CNN model is augmented with an area recommendation network, a critical addition to improve defect localization precision. The training process involves subjecting the model to various defect scenarios to ensure its adaptability to various welding conditions and defect types.

The trained Faster R-CNN model exhibits an outstanding recognition precision rate of 93% when is tested on high-density point cloud data. This significant improvement compared to that of the model trained on images from a traditional two-dimensional vision sensor demonstrates the efficiency of leveraging point cloud data in defect detection. The ability of the Faster R-CNN model to recognize and locate defect positions is essential for swift, accurate, real-time online detection during laser welding. A noteworthy finding of the study is the significant increase in the maximum welding speed allowed by the developed inspection system for online detection. The system demonstrates a maximum speed of 316.87 mm/s, a considerable advancement beyond typical laser-welding speeds. This achievement not only showcases the potential for high-speed online detection without compromising precision but also underscores the transformative impact of the developed system on industrial practices. The discussions extend beyond the principal results, exploring the implications of the system performance in various laser welding scenarios. Variations in the material thickness, welding parameters, and defect types are systematically analyzed to assess the robustness of the proposed model. The results show the adaptability of the model to different welding conditions, highlighting its versatility in practical applications. The robustness test also provides insights into potential optimizations and improvements, setting the stage for future developments in laser-welding defect detection. The study emphasizes the significance of defect localization in achieving precise measurements. The integration of an area recommendation network with the Faster R-CNN model significantly contributes to improved defect localization, a critical factor for enhancing defect measurement accuracy. This aspect of the model design is examined in detail, clarifying the mechanisms that contribute to its superior performance in defect detection.

The developed online detection system, powered by the Faster R-CNN model and high-density point cloud data, achieves a recognition precision rate of 93%. This demonstrates a substantial advancement in defect detection. By effectively addressing the challenges of classifying and measuring surface defects in laser welding, the system is established as a transformative technology with far-reaching implications in the manufacturing industry. The integration of high-density point cloud data provides rich information that enhances the efficiency and accuracy of defect detection. This breakthrough not only mitigates the limitations of traditional two-dimensional vision sensors but also positions the system as a pioneering solution for high-speed online detection in laser-welding processes. The study opens new avenues for research and development in smart manufacturing, paving the way for the integration of advanced technologies in industrial applications.