TDMRNet: Exploiting multiscale residual learning for blind estimation of polar code parameters

Abstract

Polar codes are widely utilized in modern communication systems, where accurate identification of their parameters from received signals is crucial for reliable information decoding. However, existing methods often perform inadequately in non-cooperative scenarios due to the lack of prior information. In this study, we propose a two-dimensional multiscale residual network (TDMRNet) for the blind identification of polar code parameters. Considering the structural characteristics of polar codes, the received signal sequence is first reshaped into a two-dimensional matrix to enable more effective parameter recognition. To further process the reshaped data, TDMRNet integrates an inter-code feature extraction module (ICFEM) and a data-coupling module (DCM). The ICFEM employs a unit convolution layer and three parallel standard convolution layers to construct a multiscale feature extractor, thereby enhancing the model's generalization capability and recognition accuracy. The DCM combines standard and dimension-reducing residual blocks to strengthen feature extraction while mitigating overfitting. Simulation results demonstrate that TDMRNet achieves a recognition accuracy of 93.80 %, surpassing existing methods, and reducing computational complexity, confirming its effectiveness in scenarios without prior knowledge.