Adaptive Sampling Using an Unsupervised Learning of GMMs Applied to a Fleet of AUVs with CTD Measurements

This paper addresses the problem of real-time adaptive sampling using a coordinated fleet of Autonomous Underwater Vehicles (AUVs). The system setup consists of one leader AUV and one or more follower AUVs, all equipped with conductivity, temperature and depth (CTD) sensor devices and capable of running in real-time an on-line unsupervised learning computer algorithm that uses and updates Gaussian Mixture Models (GMMs) to model the CTD data that is being acquired in real-time. The path to be traced by the leader is predefined. The followers path will depend on the CTD data. More precisely, during each resurfacing of the AUVs (and this has to be done in a coordinated fashion), every follower AUV receives the GMM hypothesis of the leader and computes the variational distance error between its own GMM and the received one. This error, that provides a notion of how different is the CTD data of each follower from the leader, is used to reconfigure the formation by scaling the distance between the AUVs in the formation (making a zoom-in and zoom-out), in order to improve the efficiency of the CTD data acquisition in a given region. The simulation results show the feasibility of the proposed strategy in uniform and more complex environments.