The study used two evaluation methods, namely K-Fold cross-validation and the training validation test, to assess the performance of the trained network. In K-Fold cross-validation, the data set was divided into ten folds, and ten models were trained and validated. The results revealed consistently high average precision values (AP50) for the different classes, ranging from 0.986 to 0.998. The Train-Validation-Test approach involved dividing the dataset into training, validation and test sets. The best-performing model on the validation set was selected, then evaluated on the test set. As with K-Fold cross-validation, the results showed remarkable AP50 values for the classes, ranging from 0.977 to 0.997. Overall precision, recall and F1 score were also impressive, with values of 0.92, 0.99 and 0.95, respectively. To further validate the performance of the entire framework, video samples were used.
The precision and recall values for each class are 0.987 and 0.982 respectively. In addition, the frame’s ability to track objects was assessed, giving a success rate of 0.994. These results collectively underline the frame’s exceptional object detection and tracking capabilities. In addition, the study examined the frame’s performance in a real factory during the production phase. A comparison was made between the framework’s predictions for the class of missed welds and actual quality control reports from the production line. This evaluation used a confusion matrix to assess the accuracy of the framework. Over a period of 17 working days, overall frame performance was reported, including predicted and reported numbers of missed welds. Accuracy was calculated at 0.888. These results provide valuable insights into the effectiveness of the frame in detecting missed welds during the production phase.
This research is making significant contributions to both academia and industry. In the academic field, it advances deep learning-based defect detection by introducing a comprehensive framework for real-time monitoring and classification of welded nodes in steel beams. This new approach not only improves detection accuracy, but also addresses practical challenges such as tracking and processing time, which are crucial for real-world applications. The development of post-processing techniques enriches academic understanding of how to improve prediction accuracy in similar applications. In industry, this research offers a game-changing solution for prefabrication quality control. By seamlessly integrating hardware, software and machine learning techniques, it provides a practical and efficient means of detecting missing welds on the factory floor. The exceptional precision achieved through advanced deep learning algorithms and fine-tuning sets a new industry standard for welding quality control. The lessons learned from implementing and validating the framework in a production environment also serve as a valuable guide for industry professionals looking to adopt similar technologies to improve their quality control processes.
