2026-09-18T19:16:05Zhttps://helda.helsinki.fi/server/oai/request

oai:helda.helsinki.fi:10138/5981192026-07-23T15:15:46Zcom_10138_18086com_10138_17738col_10138_18093

Mann, Shaiza Helsingin yliopisto, Matemaattis-luonnontieteellinen tiedekunta University of Helsinki, Faculty of Science Helsingfors universitet, Matematisk-naturvetenskapliga fakulteten 2025-06-26T06:53:15Z 2025-06-26T06:53:15Z 2025-06-26 2025-05-22 http://hdl.handle.net/10138/598119 URN:NBN:fi:hulib-202506262989 The rapid increase in data traffic, diverse device connections, and strict latency requirements in mobile communication systems have created high demands on the radio access network (RAN), especially at the physical layer. In 5G New Radio (NR) systems, good channel estimation is crucial for demodulating uplink signals like those sent over the Physical Uplink Shared Channel (PUSCH). Demodulation Reference Signals (DMRS) help channel estimation but are often sparse, needing smart interpolation to accurately recreate full channel matrices in challenging conditions. This thesis explores using knowledge distillation (KD) to improve machine learning-based channel estimation in 5G systems. A convolutional neural network (CNN) acts as a teacher model, trained on synthetic data, to create a lightweight student CNN through response-based and feature-based KD techniques. The distillation is formulated as a supervised regression task to reduce losses. The thesis evaluates the effectiveness of different distillation strategies across various operational scenarios relevant to 5G deployments, from ideal conditions to challenging cell-edge environments. It also explores the impact of hyperparameter selection on distillation quality. It investigates how different channel characteristics influence the knowledge transfer process while assessing computational efficiency through training duration measurements to provide a holistic view of each method's practical feasibility. The results demonstrate that KD enables the student model to achieve performance comparable to the teacher, while significantly reducing resource requirements, thereby making real-time deployment on edge hardware in 5G base stations feasible. eng Helsingin yliopisto University of Helsinki Helsingfors universitet In Copyright 1.0 mobile communication networks wireless networks signal processing machine learning radio technology knowledge distillation 5G new radio neural networks channel estimation hardware efficiency Tietojenkäsittelytieteen maisteriohjelma Master's Programme in Computer Science Magisterprogrammet i datavetenskap Knowledge Distillation in DMRS CHEST ML to Optimize Radio Performance and Hardware Efficiency pro gradu -tutkielma master's thesis pro gradu -avhandling open.access