Handling Uncertainty in UAV Sensor Information using Bayesian Belief Network and Large Language Model

This paper describes how UAVs can handle uncertainty in information collected from UAVs with heterogenous sensors. The approach reported here combines Bayesian Belief Network (BBN) with a Large Language Model (LLM). Our primary use case concerns the detection of forest fires but we also report laboratory experiments that are conducted using non-combustible objects. Objects’ colour, shape, are detected and interpreted using on-board sensors. Images from the UAV are also passed for interpretation to an LLM. None of the sources is perfectly applicable in all situations, as such, the UAV requires situation-based confirmation. Each of the sources is mapped to a node in BBN node with relations between nodes pre-defined through a Conditional Probability Distribution (CPD) created with input from Subject Matter Experts. We demonstrate the approach using DJI Ryze Tello programmable UAV and PyBBN scripts. The approach shows flexibility, adaptability, real-time analysis, and data saving (little data is required).