A Hybrid Framework for Blind Channel Estimation and Equalization in Multipath OFDMA Systems

This paper addresses the problem of blind channel estimation and equalization in orthogonal frequency-division multiple access (OFDMA) systems over multipath Rayleigh fading channels, modeled according to BRAN-A, BRAN-B, and Proakis B profiles. We investigate and compare three algorithms: the classical Fourth-Order Cumulant method based on higher-order statistics (HOS), a deep learning approach using an autoencoder, and a novel hybrid technique combining cumulant-derived statistical features with autoencoder learning. The Fourth-Order Cumulant method leverages higher-order statistical moments to extract channel information without requiring pilot signals, making it computationally lightweight and analytically interpretable. It performs well in stationary or low-noise conditions with slow channel variations. However, its performance degrades in the presence of high noise, nonlinearities, or rapid fading due to its reliance on rigid statistical assumptions. The autoencoder-based approach employs a neural network to implicitly learn channel effects directly from the magnitude or complex samples of the received signal, without explicit channel modeling. This method offers high adaptability and robustness to nonlinear distortions and time-varying fading, significantly outperforming cumulant-based estimation in complex scenarios. However, it incurs higher computational costs and reduced transparency, as its learned parameters are not directly interpretable. The proposed hybrid method integrates cumulant-based features into the autoencoder’s input space, enabling the network to benefit from both the explicit statistical structure provided by HOS and the data-driven learning capacity of deep neural models. Simulations conducted over a wide range of signal-to-noise ratio (SNR) values and under realistic multipath fading conditions demonstrate that: The cumulant method is optimal for low-noise, slowly varying channels, The autoencoder provides robust performance in nonlinear or rapidly changing environments, and, The hybrid approach consistently achieves the lowest mean square error (MSE) and bit error rate (BER), combining the interpretability and efficiency of cumulants with the adaptability of deep learning. In conclusion, this work highlights the complementarity between statistical and deep learning methods for blind channel estimation and equalization, demonstrating that hybridization can yield a balanced and powerful solution for next-generation wireless systems, such as OFDMA.