The identification of bearing abnormal sound is proposed by using the filter bank characteristic of empirical mode decomposition (EMD) on a white noise and short time fourier transform (STFT). Based on the EMD numerical experiments on uniformly distributed white noise, it is found that the characteristic of the EMD filter bank will be changed with a stopping criteria SD, and a dyadic filter bank, which resembles wavelet decomposition, is only a particular case of the EMD decomposition. Using the adjustable characteristic of the EMD filter bank, numerical implementations of the EMD adaptively decomposition for ball bearing vibration signal are presented, and the numerical performance of the approach to meet the needs of abnormal sound measurement is realized. Through the STFT on first thre eorders of intrinsic mode functions, the impact characteristics and instant frequency of bearing vibration are revealed in time domain and frequency domain, and the stochastic and periodic properties of those are intuitively depicted. The abnormal sound level of the ball bearing can be estimated by setting up a threshold value. Besides, the proposed method has advantage of the pattern expressi on of abnormal sound of ball bearing and can control the manufacture quality of ball bearing by abnormal sound identification.