"Forging Stock Dimension Detection"

"In the upset forging process, real-time monitoring of forging stock dimensions is crucial for ensuring deformation quality. However, continuous shape deformation and harsh environmental conditions make dimensional inspection difficult. Traditional manual inspections and conventional vision systems often fail to provide reliable and continuous feedback, leading to inconsistent product quality and material waste. To address these issues, this study proposes a deformation-adaptive and lightweight YOLOv5s-based framework tailored to forging environments. A C3_DCS module that integrates DCN and CBAM is designed to enhance feature alignment to forging-induced deformation. To improve efficiency, GSConv is introduced to reduce model complexity and maintain real-time performance. Furthermore, the SIoU is employed to accelerate the convergence of bounding-box regression and mitigate vibration-induced localization drift. A cross-scale transfer learning strategy is also developed to transfer deformation-aware features from small forging stock to large forging stock, effectively overcoming data scarcity and improving generalization across forging stock sizes. Experiments show that the proposed model achieves a Recall of 90.2%, a Precision of 85.7%, and an mAP@0.5 of 89.6%, outperforming other mainstream advanced detectors while maintaining 24.63 FPS. Transfer learning further boosts the mAP@0.5 from 80.5% to 90.5% on large forging stock and increases inference speed to 29.15 FPS. The model is finally deployed on a mobile industrial device and validated in real forging workshops, demonstrating stable real-time performance and high measurement accuracy. These results confirm the effectiveness and practical applicability of the proposed framework for online dimensional monitoring in extreme forging environments."