[Paper Review] Decentralized State Estimation In A Dimension-Reduced Linear Regression

Robin Forsling, Fredrik Gustafsson+2 more|arXiv (Cornell University)|Oct 13, 2022

Target Tracking and Data Fusion in Sensor Networks4 citations

TL;DR

This paper proposes a generalized eigenvalue optimization (GEVO) framework for decentralized state estimation in communication-constrained sensor networks, using dimension-reduced (DR) estimates to minimize bandwidth. It derives optimal linear mappings Ψ that minimize mean squared error for fusion under KF, covariance intersection (CI), and largest ellipsoid (LE) methods, with convergence analysis for CI and efficient message encoding. The method significantly improves estimation accuracy over baseline approaches while reducing communication load.

ABSTRACT

Decentralized state estimation in a communication-constrained sensor network is considered. The exchanged estimates are dimension-reduced to reduce the communication load using a linear mapping to a lower-dimensional space. The mean squared error optimal linear mapping depends on the particular estimation method used. Several dimension-reducing algorithms are proposed, where each algorithm corresponds to a commonly applied decentralized estimation method. All except one of the algorithms are shown to be optimal. For the remaining algorithm, we provide a convergence analysis where it is theoretically shown that this algorithm converges to a stationary point and numerically shown that the convergence rate is fast. A message-encoding solution is proposed that allows for efficient communication when using the proposed dimension reduction techniques. We also derive different properties from the proposed framework and show its superiority in relation to baseline methods. The applicability of the different algorithms is demonstrated using a simple fusion example and a more realistic target tracking scenario.

Motivation & Objective

• Address communication constraints in decentralized sensor networks by reducing the dimensionality of exchanged estimates.
• Develop a unified framework to compute optimal linear mappings Ψ for dimension-reduced estimates that minimize mean squared error in fusion.
• Ensure robustness and modularity in decentralized estimation while maintaining performance close to centralized optimal methods.
• Provide convergence guarantees and efficient encoding for practical deployment in real-time sensor networks.
• Demonstrate superiority over baseline methods (e.g., naïve fusion, fixed Ψ) in both synthetic and realistic tracking scenarios.

Proposed method

• Introduces a generalized eigenvalue optimization (GEVO) framework to compute the optimal dimension-reducing matrix Ψ for each agent’s local estimate.
• Derives Ψ for three key fusion methods: Kalman filter (KF), covariance intersection (CI), and largest ellipsoid (LE), using BSC-type formulas generalized to DR estimates.
• For CI, proposes an alternating minimization algorithm to compute Ψ, proven to converge to a stationary point with fast convergence rates.
• Introduces a message-encoding scheme that compresses (Ψy, ΨRΨᵀ, Ψ) into a single, communication-efficient data packet.
• Handles singular cases and provides guidelines for selecting the number of rows in Ψ based on estimation performance and numerical stability.
• Uses Monte Carlo simulations to evaluate performance across multiple fusion methods under realistic noise and network conditions.

Experimental results

Research questions

• RQ1 How can dimension-reducing mappings Ψ be optimally designed to minimize mean squared error in decentralized fusion under communication constraints?
• RQ2 What are the convergence properties of the algorithm used to compute Ψ for the covariance intersection (CI) fusion method?
• RQ3 How does the proposed GEVO framework compare in performance to baseline methods such as naïve fusion and fixed-identity Ψ?
• RQ4 In what scenarios does dimension reduction not degrade estimation performance, particularly for the LE method?
• RQ5 How can the full information in dimension-reduced estimates be efficiently encoded and transmitted to minimize communication load?

Key findings

• The proposed GEVO framework achieves significantly better estimation performance than baseline methods, with RMTR values indicating up to 40% improvement in estimation accuracy when using GEVO over PCO-based methods.
• For the CI method, the alternating minimization algorithm converges to a stationary point, with numerical results showing fast convergence rates.
• The LE method is conservative in terms of ANEES and COIN, and achieves zero performance loss when using DR estimates due to its binary fusion behavior.
• The NKF and DCA-EIG methods are not conservative, with COIN values exceeding 1.5 in some cases, indicating overconfidence in error estimates.
• The message-encoding solution enables efficient transmission of DR estimates, reducing communication overhead without information loss.
• The framework is robust to singular cases and provides clear guidelines for selecting the number of rows in Ψ to balance performance and computational cost.