The quality of the on-line differentiation with a noisy signal plays an important role in various applications involving different areas (1D and 2D). In the present work, we were interested particularly to differentiation method which requires a minimal knowledge on the noise and/or the input signal. Our work is registred in this context and address both aspects theoretical and applications. From the theoretical point of view, a study of two families of differentiation algorithm was performed. Firstly, we studied the algorithm based on the algebraic development [the work of the project INRIA-ALIEN] and secondly the differentiators of higher order sliding modes [work Levant] for detailed study. This study allowed us to describe the advantages and disadvantages of each approach. Following this study, a new version of sliding modes differentiator was proposed, this last one being the main theoretical contribution of this thesis. To test the efficiency of the proposed version, differents applications were carried out. The first application in 1D is shown, which consists in use the proposed differentiator for online parametric identification of a two-axis SCARA robot. The identification method is validated experimentally by comparing the results obtained with those given by other methods (method online and offline). The second application relates to the 2D domain. In this case, we propose the implementation of the differentiator for the edge detection in an image. To validate the results, various tests are carried out for different types of noise. In order to have some comparison basis, we use the filter Canny and Prewitt masks and the gradient. To test the effectiveness of the differentiator in control loops we are interested in electrohydraulic systems. For this, a model of the hydraulic and mechanical system is made. After a consideration of some assumptions on the simulation model, the control models are provided. To separate the dynamic servovalve/ actuator, a local study is performed. The last part of this work is dedicated to the synthesis of control strategy through the backstepping technique for the positioning of the electro-hydraulic axis in order to test the effectiveness of the proposed algorithm in the control loop. As a basis for comparison, the classic version of sliding mode algorithm is also used.
