NDKF: A Neural-Enhanced Distributed Kalman Filter for Nonlinear Multi-Sensor Estimation

TL;DR

The paper introduces NDKF, a neural-enhanced distributed Kalman filter that improves multi-sensor nonlinear state estimation by replacing analytical models with learned neural mappings, reducing communication and centralization.

Contribution

It presents a novel neural-enhanced distributed Kalman filtering approach that replaces explicit models with neural networks, enabling robust, communication-efficient multi-sensor estimation.

Findings

NDKF outperforms distributed EKF under model mismatch.

NDKF achieves improved estimation accuracy with modest communication.

Provides stability conditions and consistency strategies for learned models.

Abstract

We propose a Neural-Enhanced Distributed Kalman Filter (NDKF) for multi-sensor state estimation in nonlinear systems. Unlike traditional Kalman filters that rely on explicit analytical models and assume centralized fusion, NDKF leverages neural networks to replace analytical process and measurement models with learned mappings while each node performs local prediction and update steps and exchanges only compact posterior summaries with its neighbors. This distributed design reduces communication overhead and avoids a central fusion bottleneck. We provide sufficient mean-square stability conditions under bounded Jacobians and well-conditioned innovations, together with practically checkable proxies such as Jacobian norm control and innovation monitoring. We also discuss consistency under learned-model mismatch, including covariance inflation and covariance-intersection fusion when cross-correlations are uncertain. Simulations on a 2D nonlinear system with four partially observing nodes show that NDKF outperforms a distributed EKF baseline under model mismatch and yields improved estimation accuracy with modest communication requirements.