Lärdomar från verkliga nätverk: Ett maskininlärningsramverk för smart SRS-resurstilldelning i 5G-basstationer
Year: 2025
Type: dissertation
Abstract: Modern society increasingly depends on fast and reliable wireless communication, driven bythe rise of emerging technologies such as the Internet of Things (IoT), smart infrastructure,and a growing number of mobile users. These developments are creating a surge in datademand, pushing wireless networks to become more efficient and scalable.To meet this challenge, 5G networks rely on advanced technologies like Massive MIMO,where large numbers of antennas at base stations work together to transmit data moreeffectively. One promising technique within Massive MIMO is reciprocity-based precoding,which can outperform the standard codebook-based approach by transmitting data moreefficiently. However, reciprocity depends on specific channel estimates, which is based ondevices transmitting something called Sounding Reference Signals (SRS). Because SRSresources are limited, especially in dense urban settings, the network must make smartdecisions about which users should be granted access to them.This thesis investigates how machine learning can support more effective SRS resourceallocation. Using real-world data from a commercial 5G network, a system was developedto estimate each user’s spectral efficiency under two transmission strategies: one usingcodebook-based precoding and the other using reciprocity-based precoding. These predictions help determine whether a user would benefit from being allocated SRS resources forreciprocity-based transmission.The results show that the machine learning models were able to predict spectral efficiencywith good accuracy for both transmission methods. However, they were less effective atestimating the performance gain from switching to reciprocity. Despite this limitation,the models still provide a useful tool for guiding SRS allocation decisions. With furtherdevelopment, this approach could help networks operate more efficiently and respond moreintelligently to increasing data demands.
Author: Skäremo, Johan
Institution: Malmö University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Funder: Malmö Högskola
OpenAlex ID: W7057914661
