Simulation-Based Neural Network Models for Improved Bluetooth Ranging and Localization
Author
Klepacs, Mark
Term
4. semester
Education
Publication year
2024
Pages
41
Abstract
This thesis explores how machine learning can improve the accuracy of Bluetooth-based ranging using channel sounding—that is, sending test signals and measuring how they propagate through the environment. We simulated radio channel responses in multiple environments with the Sionna ray tracer, a physics-based ray-tracing simulator, and used these simulations to generate synthetic Bluetooth channel sounding measurements and build datasets. We evaluated different ways to represent these measurements as inputs to neural networks and examined whether models trained on simulated data transfer to real measurements. The results show that the simulated data were realistic enough to train models that generalize well to real-world scenarios. Models trained on simulation achieved ranging accuracy comparable to, and sometimes better than, the well-known MUSIC algorithm, a classical signal processing method. Overall, the findings indicate that realistic simulations can provide an effective basis for developing more accurate Bluetooth ranging methods.
Dette speciale undersøger, hvordan maskinlæring kan forbedre præcisionen af Bluetooth-baseret afstandsmåling ved hjælp af kanalsondering – det vil sige at sende testsignaler og måle, hvordan de udbreder sig i omgivelserne. Vi simulerede radiokanalers respons i flere miljøer med Sionna ray tracer, en fysikbaseret strålesporingssimulator, og brugte disse simuleringer til at generere syntetiske Bluetooth-kanalsonderingsmålinger og opbygge datasæt. Vi afprøvede forskellige måder at repræsentere målingerne som input til neurale netværk og undersøgte, om modeller trænet på simulerede data kan overføres til virkelige målinger. Resultaterne viser, at de simulerede data var realistiske nok til at træne modeller, der generaliserer godt til virkelige scenarier. Modeller trænet på simulation opnåede afstandsnøjagtighed, der var på niveau med eller bedre end den velkendte MUSIC-algoritme, en klassisk metode inden for signalbehandling. Samlet peger arbejdet på, at realistiske simuleringer kan danne et effektivt grundlag for mere præcise Bluetooth-baserede afstandsmålinger.
[This apstract has been rewritten with the help of AI based on the project's original abstract]