Aiming at the non-linearity which exits in the bearing transmission but ignores in the fault diagnosis traditionally, a method of fault diagnosis of bearing based on non-linear time series geometry invariant applying the theory of chaos and fractal is put forward. This method reduced noise by using wavelet transform and re-constructed the phase space of bearing vibrating time serial, after that , the non-linear geometry invariant such as correlation dimension, max Lyapunov exponent, K entropy and relative correlation distance entropy were calculated and were import the nerve network regarded as the character values of fault of bearing. The results of experiments show that the method can implement the fault diagnosis of bearing, otherwise, this method provides a new approach for the fault diagnosis of rotating machinery.