Typical noise signals of aluminum reduction cells were collected to e xtract the indepth features thus to determine the operating conditions exactly . The wavelet method was employed to denoise in preprocessing, and the Burg metho d an d a wavelet packet were used to estimate power spectral density and extract powe r characteristic vectors respectively, and the two methods were compared. The re sults show that the Burg method is simpler and has definite physical meaning, wh ich performs well in extracting indepth features of stationary noise signals, w hile the wavelet packet method does well in extracting indepth features of non  stationary noise signals with much transient information. A combination of the t wo methods can determine operating conditions of aluminum reduction cells accura tely.