This paper deals with a new methodology concerning parametric identification designed for non-linear uncertain systems. Non-measurable states of the model are restored through a variable structure observer which converges in a finite time. Thanks to this latter property, it is possible to derive equations of the model in order to obtain a parametric estimation law which converges in finite time to the nominal values of the parameters without the use of the classical persistent excitation for the input signal. The main interest of the approach is its robustness with respect to parameter uncertainties and additive measurement noise. This method will be compared to some classical ones. Finally, simulation results will illustrate the use of the algorithm with an application to a synchronous machine.