Research On UAV-assisted Data Collection Technology Based On Machine Learning

With the advent of the era of ubiquitous connectivity,a vast array of sensors will be deployed extensively to sample changes in the environment and system status in real time.In order to effectively conserve sensor energy consumption and overcome terrain limitations in sensor deployment areas,unmanned aerial vehicles(UAVs)can be utilized to assist in collecting sensor data,thereby enhancing the real-time and reliable collection of perceptual data.Currently,research on UAV-assisted data collection techniques mainly focuses on optimizing transmission latency or throughput,and cannot accurately reflect the total latency of perceptual data from sensor generation to data center utilization.Therefore,thesis takes information age,a novel performance indicator,as the optimization target and employs various machine learning methods to design a solution that optimizes the freshness of perceptual data collected by UAV-assisted sensing.In addition,thesis proposes single and multi-UAV-assisted data collection algorithms for different-scale sensor network scenarios to effectively solve the trajectory planning problem for UAVassisted data collection.In the design of the single-UAV-assisted data collection algorithm,thesis categorizes the optimization of the freshness of perceptual data into a nonlinear integer programming problem based on UAV energy constraints and other constraints.To solve this complex problem,thesis proposes two optimization algorithms based on Q-learning and deep Qnetwork(DQN),respectively.The former is particularly suitable for applications with a small number of sensors and a small distribution range but faces the "curse of dimensionality" problem when the number of sensors increases.Although the latter has a higher computational complexity,it can effectively solve the trajectory planning problem of a single UAV in complex network scenarios.Algorithm simulations based on PYTHON demonstrate that the two proposed algorithms provide shorter average information age for perceptual data compared with information age priority and distance priority algorithms.Furthermore,to address the problems of long information age and energy constraints that may arise for a single UAV in large-scale sensor network-assisted data collection,thesis proposes a multi-UAV-assisted data collection algorithm.The algorithm first identifies suitable data collection point locations to achieve efficient full coverage of all sensors,and then uses clustering algorithms to appropriately cluster all data collection points.Finally,a flight trajectory is designed for each cluster to optimize the data collection process for sensors within the cluster.Algorithm simulations based on PYTHON demonstrate the advantages of the UAV-assisted data collection algorithm based on data collection points over the algorithm without data collection points in terms of average information age,as well as the advantages of K-means clustering over linear/polynomial kernel clustering methods in terms of average information age.