Aim

This project will develop new machine learning inspired algorithms to extract, communicate and fuse information from networks of sensors, converting sensing into actionable information.

Objectives

1. Investigate framework for decentralised information fusion capturing both statistical and model uncertainty while working with partial information;
2. Develop new decentralised algorithms for information fusion, enabling robust detection, identification, tracking and self-calibration;
3. Test and evaluate the developed techniques within Dstl’s Stone Soup open-source software environment

Description

Large-scale sensor networks, such as coordinating swarms of drones, provide the opportunity for improved military situational awareness. While centralised fusion architectures for such networks boast near optimal inference based on powerful Bayesian statistics, they introduce a critical point of failure and can impose prohibitive bandwidth constraints. Alternatively, distributed systems offer the potential for modular and scalable solutions. However, current methods do not provide the robustness or assurance needed in military systems.

One of the biggest challenges of distributed fusion that severely limits its use, is that distributed algorithms only have access to local information and are ignorant of the global underlying network topology. This can lead to catastrophic issues when unknown duplicate or correlated information from different nodes is naively combined, resulting in overconfident predictions that corrupt and can potentially destabilise the network inference. This project will develop new machine learning inspired distributed solutions, capable of making decisions with an incomplete network model and resolving conflicting information.