Neural Kalman Filter Improves Distributed State Estimation.

Online latent state estimation is a critical challenge in AI, underpinning applications from sequential decision-making to anomaly detection. This research proposes a new distributed sensing framework, the Covariance-Agnostic Neural Kalman Consensus Filter (CA-NKCF), designed for scenarios where multiple agents collaborate to estimate a hidden state. The framework intelligently merges existing domain knowledge with the powerful representation capabilities of deep neural networks. A key innovation of CA-NKCF is its ability to perform decentralized inference without requiring explicit knowledge of noise statistics, a common limitation in traditional methods. It integrates prior estimates, optimized consensus weights, and Kalman-like recursive updates. Extensive testing across linear, chaotic (Lorenz), and practical wireless tracking environments demonstrated that CA-NKCF consistently outperforms conventional distributed Kalman and particle filters, as well as purely model-free deep neural networks, even when underlying motion and observation models are misspecified. Its robustness extends to varying noise levels, communication topologies, and observation clutter.