In this article are developed a method for the representation of nonlinear systems as well as a scheme of fault diagnosis, both use a multi-model approach. The dynamic behavior of the system is described using different points of operation for every linear time invariant model. For every model a decoupled Kalman filter of faults is used. This allows the creation of a bank of residuals to estimate the robust weighting functions. This bank is directly linked to a scheme for detecting, isolating and estimating faults in order to improve its effectiveness. A simulation example to test the performance of the developed method is presented.