Scholarlink

<p>The classification of radar and communication signals plays a crucial role in advancing cognitive radio and radar systems. While radar and communication systems often employ distinct modulation techniques, there are instances where the two domains closely overlap. To address the challenge of deep learning models failing to classify unseen signal types in signal classification tasks, this study proposes a novel model for open-set signal classification, leveraging zero-shot learning and autoencoders. The autoencoder is utilized to extract discriminative features from modulated signals, with cross-entropy loss, center loss, and reconstruction loss incorporated to ensure effective feature separation across different signal types. Open-set recognition is achieved by analyzing the feature space distribution, enabling the model to distinguish unseen signal types effectively. Moreover, a decoder is integrated into the training process to reconstruct signals, further enhancing the model’s recognition accuracy. Experimental results validate the effectiveness of the proposed approach, demonstrating its ability to identify unseen classes while maintaining a high recognition rate for known classes. The model achieves an overall accuracy (OA) exceeding 87.11%, outperforming existing signal recognition methods and highlighting its potential in cognitive signal processing applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Zero-Shot short time fourier transform radar and communication signal classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月26日

查看全文
  • Asif Shakil Ahamed,
  • Tanvir Ahmed,
  • Shekh Iftesham Iftee,
  • Asfika Jaman Tanha ,
  • Douaa Dellal,
  • Taharimul Islam

摘要

The classification of radar and communication signals plays a crucial role in advancing cognitive radio and radar systems. While radar and communication systems often employ distinct modulation techniques, there are instances where the two domains closely overlap. To address the challenge of deep learning models failing to classify unseen signal types in signal classification tasks, this study proposes a novel model for open-set signal classification, leveraging zero-shot learning and autoencoders. The autoencoder is utilized to extract discriminative features from modulated signals, with cross-entropy loss, center loss, and reconstruction loss incorporated to ensure effective feature separation across different signal types. Open-set recognition is achieved by analyzing the feature space distribution, enabling the model to distinguish unseen signal types effectively. Moreover, a decoder is integrated into the training process to reconstruct signals, further enhancing the model’s recognition accuracy. Experimental results validate the effectiveness of the proposed approach, demonstrating its ability to identify unseen classes while maintaining a high recognition rate for known classes. The model achieves an overall accuracy (OA) exceeding 87.11%, outperforming existing signal recognition methods and highlighting its potential in cognitive signal processing applications.