Hidden Actuator Degradation Prognosis Considering Unmodeled Dynamic by Neural Network-Assisted Adaptive Kalman Filter

To avoid the deviation caused by the unmodeled dynamic from traditional nonlinear filter on hiddenly degraded actuator in the nonlinear closed-loop control system, a prediction framework by developing a cooperation mode between neural network and adaptive Kalman filter is designed in this paper. Firstly, the structure and gain matrices of the adaptive Kalman filter are redesigned because of the existence of unmodeled dynamic. Then, the weight vector of cooperative neural network is recursively deduced based on developed internal collaboration model, and the convergence of the overall filter is proven. Finally, the real-time distributions of degradation states are re-derived following the iterative adjustment of filter, which is the calculation foundation of the probability density function (PDF) of remaining useful lief (RUL). The accuracy of the proposed algorithm is verified on experimental platform for RUL predicotn of motor under closed-loop control.