Abstract: The number of welding tests and the welding parameters of Q235 robot obtained through the orthogonal test L9 and the 9 groups of sample data got from tension test and low temperature impact test are done by normalization processing.BP, RBF and Elman network are established by Matlab.The best structures and major parameters are confirmed to be BP 3×10×2, RBF 2.4, and Elman 25 through learning and gerneralization test, which are analyzed to find the prediction accuracy and application effect on the weld joint strength and toughness of Q235.The results show that the average relative errors of BP, RBF and Elman are less than 10% and they can be used to predict the strength and toughness of weld joint, especially for the tensile strength.In the sample condition of this paper, Elman network is more stable and its prediction accuracy of strength and toughness is higher and better than BP and RBF, which can reflect the actual change rules and tendency of the strength and toughness of weld joint.The robot welding and the ray detection method can improve the accuracy and representativeness of sample data to enhance the prediction effect of neural network.Highlights:(1)Comparative study of the predicition accuracy and effect of BP, RBF and Elman on the weld joint strength and toughness of Q235.(2)Robot welding, ray detection and orthogonal test were used to improve the accuracy and representativeness of sample data.