Adaptive Iteration Control for Turbo Decoding: A Lightweight and Quantized Deep Learning Framework

Turbo codes have been widely adopted in 3G\/4G systems and deep space communications due to their excellent error correction capabilities. However, conventional Turbo decoders employ a fixed number of iterations, leading to significant resource wastage and processing latency, particularly under high Signal-to-Noise Ratio (SNR) conditions. Existing Early Termination (ET) strategies, such as Cyclic Redundancy Check (CRC) and Sign Change Ratio (SCR), are widely used in engineering but suffer from defects such as false convergence judgments in low SNR environments or the inability to identify non-converging \dead loops.\ Addressing these issues, this paper proposes a Quantization-Aware Training Neural Network (QAT-NN) early stopping framework driven by feature engineering. By extracting statistical features of Log-Likelihood Ratios (LLR)---such as mean, variance, and extreme values---we construct a lightweight Multi-Layer Perceptron (MLP) to predict the decoding convergence status. To mitigate the accuracy loss associated with deploying deep learning models on edge devices, a Quantization-Aware Training (QAT) mechanism is introduced, enabling high-precision 8-bit fixed-point inference. Simulation results demonstrate that, while maintaining lossless Bit Error Rate (BER) performance, the proposed scheme reduces the average number of iterations by 10\\% to 66\\% within the SNR range of -2 dB to 3 dB, exhibiting robust performance against quantization noise.