In the high mobility settings emerging in vehicular communications scenarios channel estimation is an extremely challenging task. At mmWaves, multiple-input multiple-output (MIMO) channels show a sparse impulse response in the angular and delay domains, also jointly referred to as the Space-Time (ST) domain. Low-Rank (LR) channel estimation methods propose to improve unconstrained channel estimations by representing sparsity through the algebraic LR structure of the MIMO channel impulse response matrices. In spite of their effectiveness, these models require a large quantity of unconstrained channel estimates to converge to an optimal solution and assume the invariance of ST parameters for several temporal intervals (organized in time slots), which is not always fulfilled in a high mobility environment owing to its intrinsic variational features. Therefore, considering a line-of-sight (LOS) condition for an urban vehicular scenario, we propose a deep learning model for the prediction of Low-Rank orthonormal bases representing ST subspaces, starting from single Least-Squares (LS) channel estimates. In order to train and evaluate our model, we implement a data generation framework for pre-processing and interpolating vehicular traffic and mmWave urban radio-propagation data generated through specialized simulation software. Experimental results over new simulated vehicular trajectories for the selected urban scenario show that our deep learning model, by using a single LS channel estimate, achieves on average a Mean Squared Error performance comparable with the one attained by the utilized optimal LR estimation algorithm.
Nelle situazioni a elevata mobilità emergenti in scenari veicolari la stima di canale è una procedura estremamente impegnativa. A onde millimetriche, i canali multiple-input multiple-output (MIMO) presentano una risposta impulsiva sparsa nel dominio angolare e dei ritardi temporali, anche congiuntamente denotati come dominio Spazio-Tempo (ST). I metodi di stima a rango ridotto, o Low-Rank (LR), propongono di migliorare le tradizionali stime di canale rappresentando la sparsità mediante la struttura algebrica LR delle matrici di stima di canale MIMO. Nonostante la loro efficacia, tali modelli richiedono una grande quantità di stime di canale tradizionali per convergere a una soluzione ottimale e assumono l’invarianza dei parametri ST per un elevato numero di intervalli temporali. Tali ipotesi non risultano sempre soddisfatte in scenari veicolari a causa delle loro caratteristiche intrinseche di mobilità. Perciò, considerando uno scenario veicolare urbano in condizione di line-of-sight (LOS), proponiamo un modello di deep learning per la predizione di basi ortonormali LR (rappresentanti sottospazi ST) a partire da singole stime Least-Squares (LS). Per l’addestramento e la valutazione del modello abbiamo implementato un framework per la pre-processazione e l’interpolazione di dati relativi a traiettorie veicolari e alla radio-propagazione a onde millimetriche in ambiente urbano, generati con software di simulazione specializzati. I risultati sperimentali ottenuti su nuove realizzazioni delle traiettorie simulate mostrano che il modello proposto, a partire da singole stime LS, raggiunge in media performance Mean Squared Error comparabili con quelle dell’algoritmo di stima LR ottimale utilizzato.
Deep learning-based low-rank space-time subspace prediction for mmWave MIMO channels
CAZZELLA, LORENZO
2018/2019
Abstract
In the high mobility settings emerging in vehicular communications scenarios channel estimation is an extremely challenging task. At mmWaves, multiple-input multiple-output (MIMO) channels show a sparse impulse response in the angular and delay domains, also jointly referred to as the Space-Time (ST) domain. Low-Rank (LR) channel estimation methods propose to improve unconstrained channel estimations by representing sparsity through the algebraic LR structure of the MIMO channel impulse response matrices. In spite of their effectiveness, these models require a large quantity of unconstrained channel estimates to converge to an optimal solution and assume the invariance of ST parameters for several temporal intervals (organized in time slots), which is not always fulfilled in a high mobility environment owing to its intrinsic variational features. Therefore, considering a line-of-sight (LOS) condition for an urban vehicular scenario, we propose a deep learning model for the prediction of Low-Rank orthonormal bases representing ST subspaces, starting from single Least-Squares (LS) channel estimates. In order to train and evaluate our model, we implement a data generation framework for pre-processing and interpolating vehicular traffic and mmWave urban radio-propagation data generated through specialized simulation software. Experimental results over new simulated vehicular trajectories for the selected urban scenario show that our deep learning model, by using a single LS channel estimate, achieves on average a Mean Squared Error performance comparable with the one attained by the utilized optimal LR estimation algorithm.File | Dimensione | Formato | |
---|---|---|---|
Cazzella_thesis.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
8.66 MB
Formato
Adobe PDF
|
8.66 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/154509