Discover Novel Unknown Signal Classes in Open Scenarios With Multi-Representation Discriminant Network

This letter aims to achieve the automatic classification and annotation of novel unknown signal classes in open scenarios, enabling the continuous discovery and accumulation of novel modulation information across various tasks. It proposes a multi-representation transformation swapped prediction mechanism to learn invariant signal features and utilize fusion information, promoting consistent training and improving the clustering performance. Furthermore, a discriminative classification loss is introduced, enhancing the ability to discover previously unseen classes. We perform a five-fold cross-validation on two datasets. Experimental results demonstrate that our approach outperforms other comparative approaches in several aspects.