Deep Hybrid Learning Model for Fading Mitigation in MC-DCSK Receivers

In Multicarrier Differential Chaos Shift Keying (MC-DCSK) systems, a reference chaotic sequence is transmitted over a dedicated sub-carrier, while information-bearing chaotic sequences are transmitted over the remaining sub-carriers. At the receiver, the reference sequence is first recovered and then correlated with each of the data-carrying chaotic sequences to extract the transmitted information. However, due to the intrinsic sensitivity of chaotic signals to additive noise and channel fading, the accurate reconstruction of the reference sequence becomes highly challenging, leading to significant degradation in overall system performance. To address this issue, we propose a Hybrid UNet-Transformer architecture designed to enhance the robustness of MC-DCSK receivers under fading and noise conditions. The model combines 1D-convolutional blocks inspired by UNet for effective local feature extraction, with a patch-based Transformer encoder that captures global dependencies through positional encoding. This architecture processes long and complex chaotic sequences via overlapped chunking, supported by gradient clipping and adaptive learning rate scheduling to ensure stable and scalable training. The proposed model is assessed using diverse datasets that represent various channel conditions and modulation parameter settings. Experimental results, assessed using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Signal-to-Noise Ratio (SNR), Structural Similarity Index (SSIM), correlation coefficient and Bit Error Rate (BER), demonstrate that the proposed model significantly improves reference sequence reconstruction and enhances the BER performance of the MC-DCSK system in realistic wireless environments.