Self-Supervised Neural Networks for Model-Free State Estimation

This paper introduces SsnNet-KF, a novel selfsupervised neural network-based Kalman filter designed for highprecision state estimation of aircraft operating in complex, nonlinear environments with significant time-varying noise. Unlike conventional filtering approaches that rely on explicit system models or ground-truth state labels, SsnNet-KF leverages Long Short-Term Memory (LSTM) units to directly learn the optimal Kalman gain. This eliminates the dependency on detailed system dynamics and allows for adaptive noise covariance estimation. Extensive simulations on aerospace vehicle state estimation demonstrate that SsnNet-KF significantly outperforms traditional Kalman filter variants and existing data-driven methods in terms of accuracy, robustness, and adaptability to strong disturbances and unknown time-varying noise profiles. Our approach provides a robust, model-free solution for real-time state estimation in challenging dynamic systems.