.png){kind=icon}
.png){kind=icon}
Return
CILAD: Adaptive Feature Distillation and Prototype Replay-Based Incremental Automatic Modulation Recognition
.png){kind=icon}
PRE
AI
X
C
Z
X
Y
DOI：10.1109/tgcn.2026.3711969.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To improve transmission efficiency, modulation classes adopted in wireless communications are becoming increasingly complex, while new modulation classes continue to emerge. These trends necessitate incremental automatic modulation recognition (IAMR) to understand and adapt to dynamically changing communication environments. However, most automatic modulation recognition (AMR) methods are only suitable for identifying fixed modulation classes. Moreover, direct application of computer vision (CV)-oriented incremental learning methods to AMR leads to severe catastrophic forgetting, resulting in degraded AMR accuracy. To address the aforementioned issues, a class-incremental learning (CIL) method based on adaptive feature distillation and prototype replay (CILAD) is proposed. First, a new feature extraction network for IAMR is designed. Second, in the incremental learning stage, a small number of representative old-class exemplars are preserved for subsequent training to enable the incremental model to retain representative features of old modulation classes. Third, an adaptive feature distillation method is utilized to constrain the output representations of each hidden layer and output layer of the model, thereby overcoming feature drift and improving recognition accuracy. Experimental results on both public and private datasets collected from a software-defined radio platform demonstrate that CILAD achieves higher average recognition accuracy and lower forgetting rate than four incremental learning methods.
Keywords:
Automatic modulation recognition
class-incremental learning
adaptive feature distillation
prototype replay
catastrophic forgetting
Journal
I
IF:
6.7
Papers:
1.3K
Citations:
4.3K
Organization
Cited Papers