基于Foster的调制识别增量学习方法

Incremental learning method for modulation recognition based on Foster

摘要:

针对新的调制识别信号在通信场景中动态出现的问题，提出了一种基于Foster的增量学习调制识别算法，该算法采用基于通道共享阈值机制的深度残差收缩网络，通过自适应软阈值化机制实现噪声抑制并提取有效特征；结合一种两阶段学习范式动态扩展新模块来适应新类别，通过残差拟合模块动态扩展网络容量以适应新调制类型，为解决训练过程中参数不断增多导致参数爆炸的问题，通过一种知识蒸馏策略减少特征维度和冗余参数，保持模型的主干部分；实验结果表明该方法性能明显优于其他增量学习算法，能够有效解决动态环境下的自动调制识别，展现出良好的应用价值。

Abstract:

A modulation recognition algorithm based on Foster incremental learning is proposed to address the challenge of dynamically emerging novel modulation signals in communication sc-enarios.The algorithm employs a Deep Residual Shrinkage Network incorporating a channelshared threshold mechanism.This mechanism leverages adaptive soft-thresholding to simultaneously suppress noise and extract discriminative features.Furthermore，the framework integrates a dualphase learni-ng paradigm to dynamically incorporate new modules，thereby accommodating novel modulation categories.A dedicated residual fitting module scales the network capacity dynamically to adapt to em-erging modulation types.To mitigate the issue of exponential parameter explosion during training，a knowledge distillation strategy is adopted.This strategy reduces feature dimensionality and prunes redundant parameters while preserving the core architecture of the model backbone. The experime-ntal results demonstrate that this method achieves superior performance compared to other increme-ntal learning algorithms.It can effectively address automatic modulation recognition in dynamic env-ironments and shows significant application potential.