This thesis intends to build an original process of spectral analysis of stationary signals. Its originality lies in the decision concept that we used. It is not based on the choice of one method for a particular signal, but on a methods comparison. The purpose of this comparison is to estimate the signal spectral structures. We use an iterative interpretation of the spectrum based on the properties of the analysis methods. A complete and detailed theoretical study of each selected Fourier's method is presented. In some special complex cases, techniques of parameters optimisation are developed. The advantages and drawbacks of each method are underlined as well as their comparison. Some interpretation criteria have been developed. On one hand, some pre-analysis criteria are able to forecast the signal nature and to emit alarms : a sampling Shannon test, two measures of the signal period and a non-stationarity detection. On the other hand, some interpretation criteria want to characterise each spectrum frequency : an extraction of the background noise spectrum, a peak detection and a validation criterion of the spectral lines. Finally, a global structure for analysis and interpretation is presented. This process and the criteria have been validated on the signals of the ASPECT data-basis provided by our industrial partners : EDF and the Centre Technique des Systèmes Navals. The main application domains are vibration and acoustic measurements.
