CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Optimización de la Golden Angle Modulation mediante Redes Neuronales para IoT en Redes NTN


Estudiants que han llegit aquest projecte:


Director/a: MONTORO LÓPEZ, GABRIEL

Departament: TSC

Títol: Optimización de la Golden Angle Modulation mediante Redes Neuronales para IoT en Redes NTN

Data inici oferta: 12-01-2026     Data finalització oferta: 12-09-2026



Estudis d'assignació del projecte:
    GR ENG SIS TELECOMUN
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
Golden Angle Modulation, CP-OFDM, Bit Mapping, NR-NTN, Neural Networks, GMI, Rapp, LUT
 
Descripció del contingut i pla d'activitats:
Diseño y evaluación de un simulador 3GPP de Capa Física en Python para IoT en redes no terrestres (NTN) y análisis comparativo entre modulaciones QAM tradicionales y Golden Angle optimizadas por Redes Neuronales (PyTorch).
Validación de resultados a través de métricas de rendimiento como BER, BLER, GMI, PAPR, SNR mínima, etc.
 
Overview (resum en anglès):
This Bachelor's Thesis investigates the optimization of Golden Angle Modulation by means of neural networks for Internet of Things downlink transmissions in non-terrestrial networks. The evolution towards 6G and massive machine-type communications demand simultaneous improvements in spectral and energy efficiency for low-cost devices. Conventional QAM constellations impose a rectangular geometry that leaves part of the channel capacity unexploited, a phenomenon known as the shaping gap, and concentrate power at the corners, which increases the peak-to-average power ratio. This limitation is exacerbated in the satellite link, where the power amplifier operates near saturation and introduces nonlinear distortion, clipping and spectral regrowth, while the channel exhibits Rician fading and residual Doppler.

To address this problem, a fully differentiable 3GPP 5G-NR transceiver chain is developed, comprising CP-OFDM, Zadoff-Chu synchronization, channel estimation, MMSE equalization, CRC, LDPC, rate matching and interleaving. An end-to-end autoencoder is trained on this chain to jointly optimize constellation geometry and bit mapping. The constellation starts from the Fermat spiral and the golden angle and is deformed through regularized radial and angular perturbations. The receiver uses a log-likelihood demapper based on the Mahalanobis distance, and the labeling is periodically reassigned with the Hungarian algorithm towards a quasi-Gray mapping. The loss, based on binary cross-entropy, is equivalent to maximizing generalized mutual information, and training is split into two phases: one over a Gaussian channel and another with the Rapp model, clipping and Rician fading.

Validation uses Monte Carlo simulations of the physical layer in the S and sub-GHz bands, with a Rician factor of 9 dB, residual Doppler of 350 Hz and clipping at 8 dB. Results show a stable PAPR and, from modulation orders of 32 symbols onwards, an advantage over QAM: the required signal-to-noise ratio is reduced by 2 to 3 dB, with gains of up to 4 dB for 32-GAM and 3 dB for 128-GAM, together with higher generalized mutual information. At lower orders performance is similar or slightly worse, although 8-GAM provides intermediate granularity. Finally, lookup tables are generated and their deployment in MATLAB, C and software-defined radio is described, confirming the feasibility of the proposal for IoT devices.


© CBLTIC Campus del Baix Llobregat - UPC