Neural observer-based small fault detection and isolation for uncertain nonlinear systems

Small faults (some weak faults with a tiny magnitude) are difficult to detect and may cause severe problems leading to degrading the system performance. This paper proposes an approach to estimate, detect, and isolate small faults in uncertain nonlinear systems subjected to model uncertainties, disturbances, and measurement noise. A robust observer is developed to alleviate the lack of full state measurement. Using the estimated state, a dynamical radial basis function neural networks observer is designed in form of LMI problem to accurately learn the function of the inseparable mixture between modeling uncertainty and the small fault. By exploiting the knowledge obtained by the learning phase, a bank of observers is constructed for both normal and fault modes. A set of residues is achieved by filtering the differences between the outputs of the bank of observers and the monitored system output. Due to the noise dampening characteristics of the filters and according to the smallest residual principle, the small faults can be detected and isolated successfully. Finally, rigorous analysis is performed to characterize the detection and isolation capabilities of the proposed scheme. Simulation results are used to prove the efficacy and merits of the proposed approach.