Joint Identification of Encoders and Interleavers Using Deep Learning with Hardware Validation

Abstract

In the fifth-generation (5G) communication era, efficient transmitter operations are essential, with channel encoding and interleaving mitigating random and burst errors to ensure reliable communication. In military communications and spectrum surveillance systems, which operate in non-cooperative environments, and cognitive radio systems, encoder and interleaver parameters are often unknown, necessitating blind identification. While existing research focuses on separately identifying channel encoders and interleavers with high accuracy across various channels, their joint identification remains largely unexplored. We propose a joint identification framework for channel encoders and interleavers over Rayleigh and additive white Gaussian noise (AWGN) channels. Our approach leverages advanced deep learning models, including hybrid convolutional neural network (CNN)-support vector machine (SVM), deep residual networks (DRN), and capsule networks (CapsNet), and compares their performance against standard CNN, dense neural network (DenseNet), recurrent neural network (RNN), CNN+attention, and autoencoder network models. Experimental results demonstrate that our proposed models outperform the baselines, with CapsNet achieving the highest accuracy. Notably, we observe that increasing the input size from 1024 to 16,384 (in powers of two) significantly improves classification accuracy. We analyze classification time and model complexity, showing our joint model reduces latency by 26.6% compared to separate encoder and interleaver identification. We also perform an ablation study on the CapsNet model to determine its optimal architecture, achieving a balance between accuracy and classification time. Finally, we validate our models on a hardware setup using a universal software radio peripheral (USRP) B210-based GNU Radio companion interface. The hardware results closely align with the simulated outcomes, underscoring the robustness and practical applicability of our approach. © 2015 IEEE.