Monitoring the Evolution of Cumulus Clouds with UAVs - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2016

Monitoring the Evolution of Cumulus Clouds with UAVs

Christophe Reymann
Simon Lacroix

Résumé

In this work, developed in the framework of the SkyScanner project, we study the problem of monitoring the evolution of atmospheric variables within low-altitude cumulus clouds with a fleet of Unmanned Aerial Vehicles (UAVs). This challenging objective presents two main issues that need to be overcome: i) building on-line maps of highly dynamic phenomena based on sparse, local and noisy measurements and ii) designing planning algorithms which exploit the obtained map to generate trajectories that optimize the adaptive data sampling process, minimizing the uncertainty in the map, while steering the vehicles within the air flows to perform energy-efficient flights. To deal with the on-line mapping, Gaussian Process Regression (GPR) methods have been adopted. Due to the sparsity of the sampling process in a dynamic 3D environment, the GPR probabilistic framework is indeed particularly suited for this problem. Furthermore, a particular care has been devoted to the on-line optimization of the hyper-parameters defining the Gaussian Processes, as the mapped phenomena may significantly vary in time, and depending on the area within which the UAVs are flying. The proposed regression model is then the basis on which the energy-efficient data gathering strategies are developed. The local map built with the GPR can indeed be used to generate local feasible trajectories and at the same time provide a prediction on two fundamental quantities: the energy needed to perform such trajectories and the information gain their execution will bring to the cloud map. With all these information in hand, the initial goal can be translated in a multi-criteria optimization problem as follows: finding the optimal control inputs to design safe and feasible trajectories which minimize the UAVs energy consumption, taking into account the wind field, while maximizing the total information collected along the paths. The different criteria are combined in a weighted sum, whose weights can be changed during the mission to define the impact of each criterion. For the trajectory generation and evaluation, we consider short-time planning horizons, typically in the order of 20 seconds. This choice is motivated by two main reasons: firstly, the reliability of the local maps significantly decreases (especially in time), making unrealistic any long-term prediction; secondly, the computational constraints would be harder to respect with larger optimization spaces. Each planning horizon is then further divided in sub-intervals of fixed duration in which the controls (optimization variables) are kept constant. To solve this constrained multi-criteria problem, we propose an optimization scheme composed of two successive phases: a good initialization for the set of trajectories is initially obtained thanks to a blind random search algorithm, and then optimized by using a constrained version of the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm. This algorithm is based on a simultaneous stochastic perturbation of the entire set of optimization variables to obtain a fast gradient approximation, thanks to which a convergence to a local optimum of the problem is ensured. The final objective of this work, that is experiment the flight of a fleet of drones within actual cumulus clouds, is yet to achieve. Beforehand, intensive simulations are required to assess the validity of the proposed solutions. With this intent, we present a first integrated simulation architecture, which aims at validating both the mapping and planning algorithms. Some exploration tasks have been depicted, where a fixed area of interest in which the UAVs were tasked to map the 3D wind field has been defined. The obtained results are based on a cumulus cloud model produced by realistic Meso-NH simulations and a realistic motor glider flight dynamics model. Extensive analysis varying the weights on the criteria to optimize and a thorough study of the evolution of the GP hyper-parameters are provided.
Fichier non déposé

Dates et versions

hal-01522326 , version 1 (14-05-2017)

Identifiants

  • HAL Id : hal-01522326 , version 1

Citer

Christophe Reymann, Alessandro Renzaglia, Simon Lacroix. Monitoring the Evolution of Cumulus Clouds with UAVs. 4th Conference of the International Society for Atmospheric Research using Remotely-piloted Aircraft, May 2016, Toulouse, France. ⟨hal-01522326⟩
81 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More