Abstract
Communication systems require robust techniques for channel estimation, coding, and reconstruction to ensure reliable data transmission in noisy environments. Traditional methods like turbo codes and Low-Density Parity-Check (LDPC) codes, while historically efficient, struggle with capturing intricate channel features and adapting to dynamic conditions. Consequently, there is a need for novel approaches to improve channel estimation, coding, and reconstruction in communication systems. The thesis emphasizes the importance of designing optimal channel codes with efficient error correction techniques, highlighting the potential of machine learning applications in mitigating the current challenges. To address these limitations, novel approaches have been proposed to integrate attention mechanisms into turbo code autoencoder referred to as ATT-TurboAE. These mechanisms selectively focus on informative features, enhancing system accuracy and robustness. To assess the performance of the proposed system under different noise conditions, Bit Error Rate (BER) and Block Error Rate (BLER) were compared across AWGN, non-IID AWGN, and T-distribution channels within an SNR range of -1.5 dB to 4 dB.
Comparative evaluations of the proposed model simulations over AWGN channel produced 65% improvement against standard TurboAE and 95% improvement against TurboAE-without-interleaver. The proposed model also generated 52% improvement over TurboAE and 91% better than TurboAE-without-interleaver under the non-IID AWGN channel condition. Furthermore, the proposed attention based model produced an impressive 44% better than TurboAE and 90% improvement against TurboAE-without-interleaver over the T-distribution channel. The adaptability of ATT-TurboAE-CNN over standard TurboAE trained on AWGN shows an average 63% and 43% performance improvement in non-IID-AWGN and in T-distribution noise types respectively, highlighting the proposed attention-based model superior ability to handle more complex, non-Gaussian noise types. The channel estimation result analysis over the CDL-A channel, shows that the proposed Attention mechanism based model offers a 25-30% improvement over LDAMP and Maximum Likelihood and, around a 50-60% improvement over WGAN at high SNR values. While in the Rician fading channel, the Attention
iii
Mechanism provides a 15-20% improvement over LDAMP and Maximum Likelihood and a 30-40% improvement over WGAN at higher SNRs. These results show that the proposed attention mechanism-based model performs even better under more complex channel conditions in CDL-A channels.
In conclusion, ATT-TurboAE demonstrates significantly better performance across all noise models, with the improvements most pronounced in harsher conditions such as non-IID-AWGN and T-distribution. This proves the attention mechanism's effectiveness in channel coding and noise mitigation. In summary, the thesis signifies the transition towards machine learning-based communication applications as the future of efficient communication systems.