Industry 4.0 demands intelligent quality control systems capable of real-time automated defect detection in manufacturing environments. This study presents a systematic comparative evaluation of three You Only Look Once (YOLO) architectures—YOLOv5s, YOLOv8s, and YOLO11s—for automated casting defect detection in industrial pump impeller manufacturing. Models were evaluated on 1300 real-world grayscale images across detection accuracy, inference speed, computational efficiency, and statistical reliability. YOLOv8s achieved the highest detection performance on the evaluated dataset, attaining 99.9% precision, 100% recall (95% CI: 0.962−1.000), and 99.5% mAP@0.5 at 196 FPS, corresponding to a 19.6 \(\times \) margin over the 10 FPS production-line benchmark. All three architectures operated between 179 and 213 FPS, exceeding real-time production requirements by 17–21 \(\times \) . Statistical validation via bootstrap resampling (95% CI, 1000 iterations) and permutation testing (10,000 permutations) confirmed significant inter-architecture performance differences, with large practical effect sizes (Cohen’s \(d = 0.52\) –1.89). A learning curve analysis identified performance saturation at approximately 600–700 training images, supporting the selected 1170-image training subset. All architectures achieved a 95.5% reduction in model storage requirements relative to traditional convolutional neural network approaches, supporting deployment on resource-constrained industrial edge devices. Results are reported on a single controlled dataset representing one product type; generalisation to other casting geometries requires further validation. Architecture-specific deployment guidelines for Manufacturing Execution System (MES) integration and edge computing scenarios are provided.
