Abstract: A new method of nonlinear fault diagnosis for sensors is proposed for a class nonlinear system. Transforming sensors faults into systematic ones via extended state variables with the fault derivative function estimated by RBF neural network, the method can adjust on-line the weights of network to implement real-time estimate of fault values. With the threshold processing technique applied to fault diagnosis for the uncertainties in the system, robustness is provided to the algorithm to a certain degree. The Lyapunov stability of the proposed method is proved and its correctness verified through a simulation as instance.