Lightweight Swin Transformer for high-precision industrial defect detection in smart manufacturing

Industrial defect detection keeps quality, efficacy, and cost-effectiveness in one wing, all in the fold of modern smart manufacturing. Traditional manual inspections are highly subjective, inconsistent, ineffective, and sometimes even impractical in speeding up an automated line of production. On top of this, real-world datasets for defect detection can often be challenging to train in model perspective due to the highly imbalanced nature of the data wherein defective samples merely account for a fraction of the overall data. To fight the battle, this study proposes a Lightweight Swin Transformer-based defect detection system for Copper Filter Drier (CFD) manufacturing with a Hybrid-sampling method to enhance defect identification accuracy. Over a span of twenty days, a real-world dataset was brought together involving 31,193 welding images from a factory floor, with defective examples representing about ∼ 4 % of total data. In order to address the extreme data imbalance problem, a hybrid sampling strategy was employed, that is, oversampling of defects that are rare when their comparison is against under-used samples, which allowed equal training. Different deep learning architectures such as ResNet, EfficientNet, YOLO v8, and the Swin Transformer were employed to achieve the best performance. The Swin transformer (large) achieved an accuracy highest among all other architectures, but its argument for computational complexity (107.5G FLOPs, 203M parameters) just makes it so impractical for real-field industrial implementations. Thus, the proposed lightweight Swin Transformer model balances precision with computational efficiency and ensures that the processing weight is reduced while ensuring continued defect detection accuracy which puts it in good stead to suit smart factories and SMEs. This research brings to light the need to overcome data imbalance in any deep-learning-based defect-detection system, which showcases the efficiency of advanced sampling techniques at an operational level. Future work can look into hybrid CNN and transformer architecture, self-supervised learning, and edge AI deployment systems to further optimize real-time defect detection for industrial applications.