A methodology based on relax-type wavelet network was proposed for predicting surface roughness. After the influencing factors of roughness model were analyzed and the modified wavelet pack algorithm for signal filtering was discussed, the structure of artificial network for prediction was developed. The real-time forecast on line was achieved by the nonlinear mapping and learning mechanism in Elman algorithm based on the vibration acceleration and cutting parameters. The weights in network were optimized using genetic algorithm before back-propagation algorithm to reduce learning time.The validation of this methodology is carried out for turning aluminum and steel in the experiments and its prediction error is measured less than 3%.