Abstract: An L1/L2 constrained ELM penetration identification model is proposed to solve the low accuracy problem for penetration state identification, caused by nonlinear factors during arc welding. Images of the molten pools are obtained by high speed visual sensing system. Then, feature extraction and dimensionality reduction are carried out by principal component analysis. ELM algorithm is used to train the penetration identification model for identification. To obtain generalization ability, L1 norm constraint is imposed on ELM optimization process to constraint suppresses outliers in ELM output weights. L2 norm constraint is introduced to obtain cluster features and smooth ELM output weights to improve the identification accuracy of the weld penetration. The results show that the weld penetration state recognition model based on L1/L2-ELM can quickly and effectively distinguish the three states of full penetration, partial-penetration and over-penetration.