Gaussian Mixture—Expectation Maximization–Based Learned AMP Network With CNN Deep Residual Network for Millimeter Wave Communication

ABSTRACT In millimeter wave (mmWave) multiple input and multiple output (MIMO) hybrid systems, estimating beamspace channels is complex due to the use of fewer RF chains compared with the number of transceiver antennas. Additionally, beamspace channels exhibit sparsity, making channel estimation a s...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:International journal of communication systems Ročník 38; číslo 4
Hlavní autoři: K, Shoukath Ali, Philip, Sajan P, T, Perarasi
Médium: Journal Article
Jazyk:angličtina
Vydáno: Chichester Wiley Subscription Services, Inc 10.03.2025
Témata:
Accuracy
Algorithms
approximate message passing
Artificial neural networks
Channels
convolutional neural networks
Estimation
Gaussian beams (optics)
Gaussian mixture
Hybrid systems
Iterative algorithms
Matched pursuit
Maximization
Message passing
millimeter wave communication
Millimeter waves
Mixtures
Neural networks
Optimization
residual network
Signal reconstruction
Subtraction
ISSN:1074-5351, 1099-1131
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:ABSTRACT In millimeter wave (mmWave) multiple input and multiple output (MIMO) hybrid systems, estimating beamspace channels is complex due to the use of fewer RF chains compared with the number of transceiver antennas. Additionally, beamspace channels exhibit sparsity, making channel estimation a sparse signal recovery problem. This problem can be addressed using iterative algorithms such as orthogonal matching pursuit (OMP), approximate message passing (AMP), and learned AMP (LAMP) enabled by deep neural networks (DNNs). Two methods—GM‐LAMP and GM‐EM‐LAMP—have been proposed in the literature to improve channel estimation accuracy further. The GM‐LAMP algorithm employs a Gaussian mixture (GM)–based shrinkage function within the LAMP network to enhance estimation accuracy. The GM‐EM‐LAMP algorithm incorporates the expectation maximization (EM) algorithm to recover channel information from the beamspace channel vector. However, both methods result in noisy computed channel matrices. This article addresses this challenge by proposing a GM‐EM‐LAMP network integrated with a convolutional neural network deep residual network (CDRN). The proposed method enhances the contraction function and the DNN, modeling channel estimation as a denoising problem using an element‐wise subtraction structure. The spatial properties of the channel matrix are estimated using GM‐EM‐LAMP, and the outputs of GM‐EM‐LAMP are fed into the CDRN network. Compared with existing algorithms such as LS‐CDRN, OMP‐CDRN, AMP‐CDRN, LAMP‐CDRN, and GM‐LAMP‐CDRN, the proposed approach provides superior results in terms of normalized mean squared error (NMSE) and enhanced spectral gains.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1074-5351
1099-1131
DOI:10.1002/dac.6007