An audio signal has a time series of typical nonlinear characteristics. System dynamics feature of robustness can be extracted from the input of the time series by nonlinear theory. Phase Space Reconstruction (Yang et al. 2011) (PSR) is a common method for nonlinear signal analysis. A new classification has been proposed based on instruments in this paper. The characteristics of the various instruments are analyzed with phase space reconstruction method and the differences between the different instruments are characterized with the probability density function. The probability density function of the parameter values is combined with other timbre characteristic quantities. Flexible neural tree classifier is adopted as the classification. Experimental results showed that the algorithm complexity proposed in this paper is reduced and the accuracy of the classification of different instruments is enhanced to some extent.