AI- Based welding Defect Detection: Training a model for the Classification of welding defects.

Sammanfattning

Welding quality remains critical for manufacturing safety and reliability, yet traditional manual inspection methods suffer from significant limitations including inconsistency, fatigue-induced errors, and low throughput. This thesis addresses these challenges by developing and validating an artificial intelligencebased system for automated welding defect detection. The research employed a ResNet18(Residual Network with 18 layers) deep learning architecture trained on a hybrid dataset combining publicly available images from Kaggle (125 images) and industry-sourced welding data (123 images). The model was configured for binary classification distinguishing between acceptable welds (GOOD WELD) and defective welds (BAD WELD) according to ISO 5817:2023 quality standards (levels B, C, D). Dataset was partitioned 70% training, 20% validation, and 10% test. Model training was conducted using Roboflow, a cloud-based computer vision platform, over 60 epochs. The trained model achieved a validation accuracy of 92%, with precision of 95.0%, recall of 83.3%, and F1score of 87.4% at an optimal confidence threshold of 43%. The confusion matrix analysis revealed zero false positives (no acceptable welds incorrectly rejected) and 100% recall for the GOOD WELD class, demonstrating the model's strength in preventing unnecessary production disruptions while detecting approximately 83% of actual defects. Comparative analysis with documented manual inspection benchmarks from literature reveals substantial performance advantages: the AI system reduces defect escape rates by 33-58% compared to manual visual inspection (83.3% vs. 25-40% escape rates) and achieves 100% prediction reproducibility compared to 62% inter-inspector agreement in traditional inspection. This research validates that AI-based welding inspection achieves production-grade performance suitable for industrial deployment in non-critical applications, with potential to enhance quality control substantially. The study contributes to bridging the gap between research demonstrations and practical manufacturing implementation while acknowledging limitations, data requirements, and opportunities for future multi-class defect classification and real-time deployment optimization.