Online parameter estimation for fast-time-varying non-smooth dynamical systems with insensitive time window size

For dynamical systems with rapidly and complexly varying parameters, especially when subject to uncertain noise levels, designing a reliable parameter estimation algorithm poses significant challenges. This paper proposes a sliding-window based online parameter estimation for fast-time-varying (FTV) non-smooth dynamical systems, which is less sensitive to time window size. It integrates complex time-varying features into parameter estimation models, steering clear of preset assumptions about parameter trends. Utilizing neural network's (NN) aptitude for fitting complex functions, the NN is employed to describe the mapping from time to estimated parameters, thus reducing the window size's impact on performance. The Levenberg-Marquardt (LM) algorithm is then integrated with NN to facilitate online identification and real-time computation. A longitudinal dynamical model of aircraft with time-varying aerodynamic derivatives is tested as a verified example, and the effect of noise level and time window size is also analyzed.