The role of convolutional kernels in automated welding defect detection using t-SNE and DBSCAN clustering

Welding defect detection is a critical aspect of quality control in the manufacturing industry, ensuring structural integrity and preventing failures in essential infrastructure. As the demand for higher quality standards continues to rise, ensuring the reliability and safety of welded structures has become increasingly important. Traditional methods of defect detection rely heavily on manual interpretation of radiographic images, which is time-consuming and prone to inconsistencies. Automated approaches using machine learning, particularly convolutional neural networks, have emerged as a promising solution to overcome these challenges. In this study, we analyze the changes in the categories and distributions of convolutional kernels during the training process of a welding defect detection model using a convolutional neural network. In this study, we analyze the changes in the categories and distributions of convolutional kernels during the training process of a welding defect detection model using a convolutional neural network. We systematically analyze the roles of convolutional kernels in feature extraction through a combination of dimensionality reduction using t-Distributed Stochastic Neighbor Embedding and clustering using Density-Based Spatial Clustering of Applications with Noise. Our analysis reveals that convolutional kernels within the network can be categorized into four distinct types, each contributing uniquely to feature extraction. Additionally, we quantitatively track the distribution of kernel types throughout the training process, demonstrating how the model’s feature extraction strategy evolves to enhance accuracy in welding defect detection. The insights gained from this study provide guidance for optimizing convolutional neural networks to achieve improved performance in automated non-destructive testing applications.