Detection of damaged bearings in aircraft engines by visual analysis of vibration spectrograms has proven its efficiency. Enhancement of this visual analysis can be achieved by data processing of the spectrograms. The data processing flow presented in this paper is composed of three successive steps: i) individual statistical normalization of the wideband noise (background noise of the spectrogram), ii) elimination of 'sane' narrowband vibration using a determinist parametrical model, and iii) elimination of residual 'sane' vibration using a statistical model whose parameters have been estimated from a database of 'sane' engine. This 3-step procedure has been tested on a set of vibration data from aircraft engines and the results are discussed.