Investigating the cost of ML-based fingerprinting localisation for integrated sensing and communication in cell-free massive MIMO systems

Krøyer, Sune ; Thorsager, Mathias Drekjær

4. Semester

Communication Technology

2023

2023-06-02

62

Dette speciale undersøger balancen mellem lokaliseringsnøjagtighed og den ekstra belastning, metoderne påfører det underliggende kommunikationsnet. Vi fokuserer på integreret sensing og kommunikation, hvor positionsestimater udledes af data fra den eksisterende kommunikation. Vi udvikler tre metoder baseret på maskinlæring: to, der bruger maskinlæring som mellemtrin i en ankomstvinkel-baseret fingeraftryksmetode (fingerprinting), og én, der estimerer positionen direkte med maskinlæring. Med fingerprinting menes her at matche målte signalkarakteristika med et katalog over signalmønstre fra kendte positioner. De to fingerprinting-metoder har til formål at forbedre nøjagtigheden i forhold til en referencemetode, mens den direkte metode skal reducere opsætningsomkostningen ved at indsamle mange fingeraftryk. Resultaterne viser, at den direkte metode, trods lavere opsætningsomkostning, ikke fungerede i det undersøgte system. Til gengæld overgik begge fingerprinting-metoder referencemetoden og gjorde det med ringe eller ingen ekstra belastning af kommunikationsnettet.

This thesis examines the trade-off between localization accuracy and the extra load that localization methods place on the underlying communications network. We focus on integrated sensing and communications, where location estimates are derived from data already present in communication signals. We develop three machine learning methods: two that use machine learning as an intermediate step in an angle-of-arrival–based fingerprinting approach, and one that performs location estimation directly with machine learning. Here, fingerprinting means matching measured signal characteristics to a catalog of signal patterns recorded at known locations. The two fingerprinting methods aim to improve accuracy relative to a fingerprinting-based baseline, while the direct method aims to reduce the setup effort of collecting many fingerprints. Results show that, despite its lower setup cost, the direct method did not perform in the studied system. In contrast, both fingerprinting methods outperformed the baseline while adding little to no additional strain on the communications network.

[This summary has been rewritten with the help of AI based on the project's original abstract]

cell-free mMIMO ; machine learning ; DNN ; CNN ; ISAC ; localisation ; channel models ; UMi ; fingerprinting ; AoA estimation

Download
View record in AAU Student Projects