Online monitoring of typical welding defects in robotic laser-MAG hybrid welding based on multi-source information fusion and Stacking-PSO-LightGBM

Due to the intricate mechanisms of laser-MAG hybrid welding, sensor signals from a single source exhibit inherent limitations in both comprehensiveness and reliability for quality monitoring tasks. This study employs multi-source information fusion to integrate molten pool and keyhole images, welding current, and molten pool temperature at the feature level, and further implements online monitoring of typical defects during laser-MAG hybrid welding through ensemble learning. The proposed method employs fixed regions of interest, wavelet denoising, and frame-based windowing for preprocessing multi-source signals. A Ghost-ResNet18 convolutional neural network is introduced to extract discriminative features from molten pool and keyhole images, while statistical features are computed from welding current and molten pool temperature. After feature fusion and normalization, a multi-model averaged permutation importance method is employed for feature selection, effectively reducing redundancy and improving accuracy. To achieve robust defect recognition, a Stacking-PSO-LightGBM ensemble model that combines RF, KNN, and SVM as basic learners is proposed, with hyperparameter optimization performed by particle swarm optimization. This framework successfully identifies five typical welding defects with an accuracy of 98.86 %, achieving an average processing time of 78.23 ms, thereby meeting the real-time requirements of welding quality online monitoring systems.