L’analyse vibratoire est l’une des principales méthodes pour le suivi et la maintenance des machines tournantes. Cette thèse s’inscrit dans cette thématique et en particulier dans le diagnostic des roulements en régime variable. L’objectif de la thèse est l’élaboration de capteurs logiciels. Ces derniers pourraient fournir un indicateur pertinent sur l’état de roulements défectueux à partir des signaux de vibrations de type aléatoire qu’ils génèrent dans les cas où ils sont mélangées à d’autres sources vibratoires provenant de défaut d’engrenages générant des signaux déterministes. Afin de diagnostiquer d’une façon efficace les cas d’avarie, il faut au préalable séparer le signal de vibration relatif à chaque source de vibration. En régime stationnaire, l’outil utilisé pour estimer la composante déterministe est la moyenne synchrone. En régime variable, la partie déterministe du signal de vibration évolue avec le régime même si l’enregistrement du signal de vibration est synchronisé avec la vitesse de rotation. Dès lors, cette composante ne devient plus périodique et son estimation par la moyenne synchrone classique devient inappropriée. Nous avons alors proposé une nouvelle approche qui consiste à caractériser les signaux de vibrations en régime variable en introduisant la notion de cyclo-non-stationnarité. Ceci nous a permis de développer un outil d'estimation des composantes déterministes du signal qui sont caractérisées par la cyclo-non-stationnarité d'ordre 1 et de développer un outil de caractérisation de la cyclo-non-stationnarité d'ordre 2 basé sur la synchronisation de la corrélation spectrale. Les performances de cet outil ont été testées avec succès sur des signaux synthétiques ainsi que sur des signaux réels. / Vibration analysis is one of the main methods for rotating machinery monitoring. The topic of this thesis is the diagnosis of rolling element bearings under varying rotating speed. The objective is the development of software sensors providing a relevant indicator of bearing condition monitoring from the vibration signal generated by the bearing. Often, rotating machinery is composed from gears and bearings. Gear faults generate deterministic components and bearing faults generate random components. To diagnose, the key point is to extract the vibration signal caused by bearing from the total vibration signal. In constant rotating speed condition, the used tool to estimate the deterministic component is the time synchronous averaging. In varying rotating speed, the signal becomes "cyclo-non-stationary". Therefore, the traditional synchronous averaging method becomes inappropriate and one needs to extend its definition and estimation process. In this thesis, we develop a new approach based on a decomposition of the vibration signal into a set of cyclo-non-stationary components. This allowed us to develop a tool for estimating the deterministic components of the signal which are characterized by cyclo-non-stationarity order 1 and to develop a tool for characterizing the cyclo-non-stationarity order 2 based on synchronized spectral correlation. The performance of this tool has been tested on synthetic signals and on real signals.

Condition based maintenance is becoming increasingly popular in many industrial contexts, offering substantial savings and minimising accidental damage. When applied to rotating machinery, its most common tool is vibration analysis, which relies on well-established mathematical models rooted in the theory of cyclo-non-stationary processes. However, the extraction of diagnostic information from the real world vibration signals is a delicate task requiring the application of sophisticated signal processing techniques, tailored for specific machines operating under restricted conditions. Such difficulty in the current state of the art of vibration analysis forces the industry to apply methods with reduced diagnostic capabilities but higher adaptability. However in doing so most of the potential of vibration analysis is lost and advanced techniques become of use only for academic endeavours. The aim of this document is to reduce the gap between industrial and academic applications of condition monitoring, offering ductile and automated tools which still show high detection capabilities. Three main lines of research are presented in this document. Firstly, the implementation of stochastic resonance in an electrical circuit to enhance directly the analog signal from an accelerometer, in order to lower the computational requirements in the next digital signal processing step. Secondly, the extension of already well-established digital signal processing techniques, cepstral prewhitening and spectral kurtosis, to a wider range of operating conditions, proving their effectiveness in the case of non-stationary speeds. Thirdly, the main contribution of the thesis: the introduction of two novel techniques capable of separating the vibrations of a defective component from the overall vibrations of the machine, by means of a threshold in the amplitude spectrum. After the separation, the cyclic content of the vibration signal is extracted and the thresholded signals provide an enhanced detection. The two proposed methods, phase editing and amplitude cyclic frequency decomposition, are both intuitive and of low computational complexity, but show the same capabilities as more sophisticated state of the art techniques. Furthermore, all these tools have been successfully tested on numerically simulated signals as well as on real vibration data from different machinery, lasting from laboratory test rigs to wind turbines drive-trains and aircraft engines. So in conclusion, the proposed techniques are a promising step toward the full exploitation of condition based maintenance in industrial contexts.