Since accurate models of nonlinear systems are difficult to obtain a-priori, it is necessary to obtain these models from input-output data. As neurofuzzy networks can approximate nonlinear functions with arbitrary accuracy, and can be trained from data, they are used here to model nonlinear systems, It is shown that the residuals generated from the model is Guassian distributed, and that the aymptotic approach can be applied to detect fault, when the residuals computed from the model exceeded a threshold determined by the X²-test for a given false alarm probability. The proposed fault detection procedure is demonstrated by an example.