For the nonlinear and non-stationary characteristics of rolling bearing fault signals, a feature extraction method based on optimal generalized S transform and pulse coupled neural network(PCNN) was proposed. Generalized S transform was optimized by measuring the time-frequency aggregation, and then utilized to achieve time-frequency representations of bearing fault signals. Time-frequency images were further decomposed into a series of binary images. The capture rate sequences of binary images were then defined and extracted as the bearing fault feature parameters. Rolling bearing signals of four different states were analyzed. The results indicate that the proposed method can extract more effective bearing fault feature parameters which are capable of improving the bearing fault diagnosis accuracy.