As a representative nonlinear system, significant progress has been made in the state estimation of the Duffing system in recent years. However, existing methods still suffer from several shortcomings, including heavy reliance on noise statistical parameters, sensitivity to initial conditions, particle degeneracy, and challenges in capturing complex temporal dynamics. To address these issues, this study proposes a novel dynamic noise covariance matrix estimation framework based on Adaptive Variational Bayesian Filtering (AVBF). The framework integrates AVBF with a physically driven system model and employs an exponentially decaying forgetting factor, which adaptively balances historical information with real-time observational data to yield more accurate noise covariance estimates. Extensive simulations demonstrate that the proposed approach significantly reduces estimation errors, providing a theoretically rigorous and practically valuable new paradigm for state estimation in complex nonlinear systems.