Optimizing YOLOv8 architecture using particle swarm optimization for high-precision binary quality classification in industrial welding seams

Abstract

The structural integrity of heavy machinery fundamentally depends on precise welding quality. However, traditional manual inspections remain inconsistent, labor-intensive, and susceptible to human error. While You Only Look Once v8 (YOLOv8) architectures have become the standard for real-time object detection, their performance in accurately classifying micro defects like porosity or cracks is frequently hindered by suboptimal default hyperparameters. To overcome this limitation, this study proposes PSO YOLOv8, an intelligent framework integrating the Particle Swarm Optimization (PSO) algorithm to automatically tune YOLOv8 critical hyperparameters, specifically learning rate, batch size, and weight decay. The framework was evaluated using a specialized dataset of 2,600 high resolution welding seam images, strictly categorized into Normal and Defective classes. Utilizing validation Mean Average Precision (mAP) as the fitness function, PSO was configured to maximize accuracy over 50 iterations. Experimental results demonstrate a substantial performance enhancement. The PSO optimized model achieved an mAP@50 of 94.2%, a significant improvement over the 83.7% baseline. Furthermore, the optimized configuration attained a 96.5% Precision rate, effectively reducing false-positive detections by 38.4%. These findings validate that fusing metaheuristic algorithms with deep learning provides a robust, high precision tool for automated quality assurance in smart manufacturing.