Robust neural network RISE observer based fault diagnostics and prediction

Abstract

A novel fault diagnostics and prediction scheme in continuous-time is introduced for a class of nonlinear systems. The proposed method uses a novel neural network (NN) based robust integral sign of the error (RISE) observer, or estimator, allowing for semi-global asymptotic stability in the presence of NN approximation errors, disturbances and unmodeled dynamics. This is in comparison to typical results presented in the literature that show only boundedness in the presence of uncertainties. The output of the observer/estimator is compared with that of the nonlinear system and a residual is used for declaring the presence of a fault when the residual exceeds a user defined threshold. The NN weights are tuned online with no offline tuning phase. The output of the RISE observer is utilized for diagnostics. Additionally, a method for time-to-failure (TTF) prediction, a first step in prognostics, is developed by projecting the developed parameter-update law under the assumption that the nonlinear system satisfies a linear-in-the-parameters (LIP) assumption. The TTF method uses known critical values of a system to predict when an estimated parameter will reach a known failure threshold. The performance of the NN/RISE observer system is evaluated on a nonlinear system and a simply supported beam finite element analysis (FEA) simulation based on laboratory experiments. Results show that the proposed method provides as much as 25% increased accuracy while the TTF scheme renders a more accurate prediction.