Abstract: In this paper, the correlativity is analyzed among characters of time domain and frequency domain of arc sound signal, prccess of transfer and spatter in CO₂ short circuiting welding. In different frequency bend,wavelet packet is applied to sound signal analysis to get the energy character that can be regarded as the characteristic vector to denote the spatter. The model which can denote the spatter using character vector is proposed using the neural network to predict the spatter in CO₂ short-circuiting arc welding. The simulation shows that welding spatter can be satisfactorily predicted through the characteristic vector of sound vector and that it is a new way on-line quality control.