Abstract: Penetration is one of the most important index in the welding quality evaluation. Nonuniform penetration is easily happened in aluminum alloy welding process because of the high sensitivity of aluminum to welding parameters. In the single-side welding and double-side molding experiment clear weld pool images of three kinds of penetration status-incomplete, complete and over penetration have been obtained by near-infrared visual sensing method. The characters of weld pool image such as weld width, weld half-length, molten pool area, perimeter and parabolic coefficient which is associated to weld penetration can be extracted by a special image processing algorithm. The penetration recognition model of aluminum alloy in tandem arc welding based on BP neural network was established and the result showed that the 5-13-3 structured BP neural network model has the highest recognition accuracy which is 89.05%.