This article established a new combining hierarchy genetic algorithm and multivariate linear regression model of WNN (wavelet neural network) for identify the feature of rotary machine. The effection on the question of nonlinear approximation is verified through the simulation and optimization. The test datas of a tandem mill are inputted into the model. After trained, the model has automatic ability of obtained the inspect information and the ability of adapt the changing of worked condition. The self-adaptive study and diagnosis of torsional oscillation state on different work condition are realized. The results verify the combining hierarchy genetic algorithm and multivariate linear regression model has the reliability.