Multi-information fusion welding defect identification combining neighborhood rough set and optimized SVM

A multi-information fusion welding defect recognition method is proposed by combining NRS with optimized SVM to address the “big data” generated during the multi-sensor information fusion welding process. A fast reduction algorithm based on NRS was constructed using feature importance as inspiration information, and the DOA was used to select the key parameters of SVM. Through welding experiments, welding quality information such as melt pool images, welding currents, and vibration signals was obtained. A simplified dataset was generated using feature fusion and NRS reduction, and then loaded into DOA-SVM for optimization training to establish a welding defect recognition model. Multiple sets of experiments were designed to compare and verify the method. The results showed that the model achieved an accuracy of 98.03% in identifying six types of weld quality categories, with short training and prediction time and strong generalization ability, which could meet the requirements of online welding quality detection.