Clasificación automática de señales para comunicaciones inalámbricas basada en técnicas de inteligencia artificial

Abstract

This work addresses the identification of digital modulation schemes under realistic chan- nel conditions—multipath fading, additive white Gaussian noise (AWGN), and carrier phase offset—across six formats: BPSK, QPSK, 16-QAM, cross-32-QAM, 64-QAM, and cross-128- QAM; a complementary analysis with 256-QAM is also included to assess performance under higher constellation orders. To capture channel variability, scenarios with different signal-to- noise ratio (SNR) levels, numbers of propagation paths, and phase-offset values were gene- rated using the TorchSig toolkit; the resulting baseband in-phase/quadrature (I–Q) signals fed three neural-network architectures trained and evaluated in supervised classification tasks. Finally, a voting-based ensemble that combines the top k models from each network is pro- posed, yielding consistent gains in correct classification probability (PCC) across all evaluated scenarios.