Abstract: A network model for penetration state diagnosis based on the signal characteristics of arc sound in MIG welding is developed by recombining and improving artificial neural network, wavelet transform, and genetic algorithm.The arc sound signals, which are denoised by using wavelet transform and extracted by the frequency-band energy characteristics via wavelet packet decomposition and reconstruction, are used as the input eigenvectors of the wavelet neural network model, the genetic algorithm which has the ability of global optimization is adopted to dynamically modify the network structure and parameters and eliminate the rate tardiness of neural network training and relapse into local extremum, and then the complex nonlinear modeling and data mining are accomplished.The penetration state diagnosis result of the trained network model verifies the feasibility and validity of the modeling methods.