Deep Learning for Direction of Arrival Estimation in Massive MIMO Systems
Author
Jensen, Kasper Steensig
Term
4. term
Education
Publication year
2021
Submitted on
2021-06-03
Pages
53
Abstract
Med udbredelsen af 5G er massive MIMO-systemer (basestationer med mange antenner) blevet mere relevante. Når mange master står tæt, kan de såkaldte pilotsignaler, der bruges til at måle radiokanalen, forstyrre hinanden. Dette kaldes pilotforurening og gør kanalestimering vanskelig. For at mindske afhængigheden af pilotsignaler undersøger dette arbejde blind kildeadskillelse (Blind Source Separation, BSS), som forsøger at udlede kanalparametre direkte fra modtagne signaler. Traditionelle BSS-metoder kan være langsomme, så tre alternativer afprøves: dybe neurale netværk, sparsomhedsbaserede estimationsmetoder (ISTA-baserede) og en hybrid mellem de to (Learned-ISTA). Fokus er på estimering af ankomstretning (Direction of Arrival, DoA). Metoderne opnår sammenlignelig nøjagtighed, målt som normaliseret middelkvadreret fejl (NMSE; lavere er bedre) omkring 10e-1 til 10e-2. Learned-ISTA fremstår lovende, og det er relevant at undersøge træningsstrategier, modelparametre og topdetekteringsalgoritmer til valg af ankomstretninger nærmere.
As 5G networks roll out, massive MIMO systems (base stations with many antennas) are becoming more common. When many sites are placed close together, the pilot signals used to estimate the radio channel can interfere with each other. This “pilot contamination” makes channel estimation unreliable. To reduce reliance on pilots, this work explores blind source separation (BSS), which aims to infer channel parameters directly from received signals. Because traditional BSS methods can be slow, three alternatives are evaluated: deep neural networks, sparse estimation methods based on ISTA, and a hybrid approach (Learned-ISTA). The study focuses on direction-of-arrival (DoA) estimation. The methods achieve comparable accuracy, with normalized mean squared error (NMSE; lower is better) roughly between 10e-1 and 10e-2. Learned-ISTA shows promise, and further work on training procedures, model parameters, and peak-finding algorithms for selecting directions of arrival is warranted.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
MIMO ; 5G ; DoA estimation ; sparse estimation ; ISTA ; deep learning ; DNN
Documents