Pass-by noise signal is non-stationary and time variant. It is difficult to identify sound sources accurately by the traditional FFT method. A new noise source identification technique was developed by using the wavelet packet transform. Firstly, the features of noise sources were extracted by the decomposition and reconstruction of the wavelet packet to time variant signals, and then, contributions were got from all kinds of the noise sources to the standard measuring point with the solving process of the multidimensional linear system of equations. Research showed that the technique could clearly distinguish noise energy change in the time-frequency domain and quantitatively identified noise sources. 