26 May 2026 in IEEE Transactions on Instrumentation and Measurement
State estimation in unknown nonlinear dynamical systems has gained significant importance. However, existing methods often suffer from performance degradation when dealing with imperfect and incomplete observations, especially under nonlinear and non-Gaussian measurement conditions. To tackle this challenge, we propose a novel neural stochastic differential equation network (NSDE-Net) for efficient learning of system dynamics under non-ideal observations. The NSDE-Net models unknown system dynamics using a neural stochastic differential equation (SDE), where the hidden state is resolved through an observation-driven neural network. Specifically, the NSDE-Net employs an LSTM-based temporal modeling framework, enabling an observation encoder to create a direct mapping from observations to latent states. The drift and diffusion networks capture deterministic and stochastic dynamics, thereby reducing uncertainty in state estimation under non-ideal observations. This model is optimized through an observation likelihood function augmented with a Lipschitz prior regularization term. Additionally, we provide a theoretical analysis of the mean-square stability of the proposed NSDE-Net under suitable boundedness and Lipschitz conditions. Finally, simulations validate the stability of the proposed framework and the estimation performance of NSDE-Net under various observational conditions, while a real INS/GNSS navigation system is employed to evaluate the overall effectiveness of the algorithm.