Fault detection of ball bearings by the complex and non-stationary vibration signals with noise is very difficult, especially at the early stages. Also, many failure mechanisms and various adverse operating conditions in ball bearings involve significant nonlinear dynamical properties. The quality of chaotic vibration of ball bearings is studied by the reconstructed phase space. The phase space demonstrates different chaotic vibration of ball bearing for different healthy/faulty conditions. But, to easily use of this procedure in the ball bearing fault detection, the chaotic behavior of vibration signal is quantified by a set of new features. The new set of features based on chaotic behavior, including the largest Lyapunov exponent, approximate entropy and correlation dimension are extracted to acquire more fault characteristic information. The effectiveness of the new features based on chaotic vibrations in the ball bearing fault detection is demonstrated by the experimental data sets. The proposed approach can reliably recognize different fault types and have more accurate results. Also, the performance of the new procedure is robust to the variation of load values and shows good generalization capability for various load values.