Unsupervised modulation classification based on multi-level deep subspace clustering

Automatic modulation classification (AMC) recognizes modulation patterns based on features extracted from received signals. Benefiting from its powerful feature extraction capabilities, deep learning is widely and successfully applied for AMC. However, existing AMC mainly based on supervised deep learning, which is highly dependent on labelled data and has difficulty constructing large-scale and well-labelled datasets, still faces challenges. To address this challenge, we propose an unsupervised learning architecture for AMC (USLAMC) by uniting feature extraction and deep subspace clustering (DSC). Feature extraction can utilize multi-level information to obtain the original dataset features. DSC can generate self-representation relationships for features, and use spectral clustering realizes unsupervised subspace feature clustering.