Real-time object detection of robotic welding defects and adaptive parameter tuning based on YOLO

Abstract

To address the issues of offline flaw detection feedback delay and limited edge deployment of complex vision models in continuous welding scenarios using industrial robots, this paper proposes a defect perception and adaptive electrical parameter control algorithm based on lightweight multi-scale feature reconstruction. The study utilizes reparameterization technology and an asymmetric feature space attention mechanism to effectively compress the neural network parameter footprint (2.81M) while improving the accuracy of dynamic defect localization. A temporal smoothing mapping model between defect confidence features and welding electromechanical execution compensation variables is constructed to suppress control oscillations caused by nonlinear transient deviations. Closed-loop deployment tests on edge computing nodes show that the algorithm achieves an average accuracy (mAP@0.5:0.95) of 69.53% for complex welding defect detection, reduces the single-frame image tensor inference delay to 6.84ms, and controls the end-to-end total time for adaptive correction of system drive parameters within 9.30ms. The results verify that the algorithm effectively ensures the stability of the closed-loop control process and the cladding quality while meeting microsecond-level real-time response specifications.