Abstract : In this thesis we propose a new method of isolation and identification of singular fault for nonlinear dynamic systems. This method is based on the characteristic of monotonicity of the observer prediction error according to the difference of the parameters. The whole of the admissible values of each parameter is subdivided in a certain number of intervals. One builds an isolation observer for each interval, this observer is initialized in the considered interval. After the fault occurrence, the value of the faulty parameter must be in one of the parameter intervals. The amplitude of the residue calculated by the isolation observer corresponding to this interval (that which contains the faulty parameter value) will be in the area limited by two dynamic thresholds at any time. On the other hand, the residues corresponding to the other intervals will have great amplitudes and their evolutions are not limited by the two corresponding dynamic thresholds. Consequently the interval containing the value of the faulty parameter can be determined and the fault is isolated and identified. Various versions of this method were developed: a first with fixed thresholds, one second with adaptive thresholds and a last without thresholds. One can show that this method has common points with that based on the adaptive observers. However, this last has a major disadvantage which is the isolation time. The proposed method in this work enables us to solve the problem of the isolation time.