A Deep Learning-Based Novel Class Discovery Approach for Automatic Modulation Classification

The existing Automatic Modulation Classification (AMC) methods require the training and testing datasets share a common set of modulation categories. However, the AMC model may encounter the need to discriminate novel classes in non-cooperative environments. To the best of our knowledge, the research reporting AMC in the class-disjoint environment has not been addressed yet. In this letter, a novel class discovery method is proposed for AMC leveraging the information contained in the labeled training dataset. Specifically, a 3-stage deep learning method is introduced to recognize samples of the known classes and cluster samples of novel classes. The extracted features and the pairwise similarity relationship are considered as the common knowledge between the two class-disjoint datasets and are utilized to help the construction and training of the classifier for novel classes. The simulation results validate the effectiveness and performance of the proposed method.