Insights · Fusion & Distributed Intelligence

Track Fusion When the Network Is the Bottleneck

September 2026 · 4 min read

AMBy Amardeep Mond, Founder & Chief Technology Officer, Aerodyne AI

Textbook sensor fusion assumes the network is free: every measurement reaches a central node, on time, always. Air defense networks operate under the opposite assumption — links are bandwidth-constrained, intermittently jammed, and any central node is a target. The distributed-fusion literature exists precisely for this regime, and its results shape what a credible wide-area sensing architecture must look like.

Why the centralized optimum is a fragile optimum

If every sensor could stream raw measurements to one processor, classical multi-sensor tracking theory gives the optimal answer: a central filter fusing all measurements with full knowledge of their statistics [1]. The result is provably the best achievable track picture — and architecturally the worst achievable system, for a military network. It concentrates all processing at a single point of failure, demands the most communication bandwidth exactly when links are most contested, and produces nothing anywhere if the center or its links are lost.

The alternative is track-level fusion: each node runs local trackers on its own measurements and exchanges compact track estimates rather than raw data. This slashes bandwidth by orders of magnitude and lets every node hold a usable local picture even when cut off. It also introduces the two problems that define the field.

The two hard problems: association and correlation

Association — deciding which local tracks at different nodes correspond to the same physical aircraft — is combinatorial, and grows hard exactly when it matters most: dense raids, crossing trajectories, decoys. The literature offers a well-mapped ladder of approaches, from global nearest neighbor through joint probabilistic data association to multiple hypothesis tracking, trading computation for robustness in clutter [1].

Correlation is subtler. Two nodes' estimates of the same target are not independent: they share process noise, prior exchanges, and common-source information. Fusing them as if independent produces overconfident estimates — a fused track that believes itself more accurate than it is, which in an engagement system translates into misplaced trust. When the cross-correlation is unknown (the normal case in a dynamic network), covariance intersection provides a fusion rule that is provably consistent regardless of the unknown correlation, at the price of some conservatism [2]. The existence of a consistency-guaranteed fusion rule for unknown correlations is one of the quietly load-bearing results of the field: it is what makes opportunistic, topology-agnostic track sharing sound.

Learning at the edge without moving the data

Modern detection and classification are learned functions, and learned functions want data. The naive pipeline — ship all field data to a central training cluster — inherits every weakness of centralized fusion, and adds two more: the bandwidth cost of raw sensor data dwarfs that of tracks, and operational data is often exactly the data an owner cannot or will not export.

Federated learning inverts the pipeline: models travel to the data, train locally, and only model updates are aggregated [4]. A decade of results has extended the basic scheme to communication-efficient variants, heterogeneous and unreliable participants, and — critically for contested networks — aggregation rules robust to a fraction of nodes behaving arbitrarily badly, whether through faults or adversarial compromise. Byzantine-resilient distributed learning algorithms provide convergence guarantees even when some participants are actively malicious [3], and the broader open-problems literature maps the field's state with unusual candor [5]. For a defense sensor network, the implications compound: raw data custody stays with the data owner (a property with obvious coalition and sovereignty value), bandwidth is spent on model deltas rather than sensor streams, and a poisoned or captured node's influence on the shared model can be bounded by construction rather than by hope.

Degradation as a design axis, not a failure mode

The common thread across distributed estimation and federated learning is that the communication graph is a first-class design variable. A well-designed network specifies its behavior along the whole degradation curve: full connectivity yields near-centralized accuracy; partitioned operation yields consistent, locally maintained pictures that re-merge cleanly when links return; total isolation yields nodes that still sense, track, and decide within delegated bounds. Systems designed only for the connected case do not degrade — they collapse. The literature's tools exist so that architects can choose the former, and the difference between the two is visible not in demonstrations under ideal conditions but in the design documents' answer to a single question: what, exactly, happens when the network fails?

References

1. Y. Bar-Shalom and X.-R. Li, Multitarget-Multisensor Tracking: Principles and Techniques, YBS Publishing, 1995.
2. S. J. Julier and J. K. Uhlmann, "A non-divergent estimation algorithm in the presence of unknown correlations," Proceedings of the American Control Conference, 1997.
3. Z. Yang and W. U. Bajwa, "ByRDiE: Byzantine-resilient distributed coordinate descent for decentralized learning," IEEE Transactions on Signal and Information Processing over Networks, vol. 5, no. 4, 2019.
4. H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. Agüera y Arcas, "Communication-efficient learning of deep networks from decentralized data," Proceedings of AISTATS, 2017.
5. P. Kairouz et al., "Advances and open problems in federated learning," Foundations and Trends in Machine Learning, vol. 14, no. 1–2, 2021.

This article synthesizes publicly available academic literature and public reporting. It describes concepts and results from the open published record and contains no Aerodyne technical data, system specifications, or performance information.