Fine-Grained Open Set Signal Modulation Classification via Self-Supervised Pre-Training

Open set automatic modulation classification (OAMC) can identify unexpected unknown modulation types in real-world scenarios, which is receiving increasing interests in electronic countermeasures and signal recognition. However, existing OAMC methods rely on a large number of accurate labeled data, which makes the generalization ability of models poor and undoubtedly brings tremendous difficulties and challenges, especially in the case of limited labeled information. In addition, they only implement simple unknown detection without fine-grained unknown classification, which is necessary especially in areas that require high adaptability and robustness to unknown classes. Therefore, this paper proposes a CNN-based learning framework via self-supervised pre-training with out-of-distribution (OOD) generation and quadratic discrimination for fine-grained OAMC, named FOSSP, to address related problems. First, we use self-supervised pre-training learning paradigm in which a reformed de-biasing weighted contrastive loss is introduced to handle sample imbalance. To effectively capture the semantic features of the signals, we employ a tailored blend of loss functions including reconstruction loss, center loss, and cross-entropy. These are complemented by a suitable distance metric to ensure that the minimal distance between semantic features exceeds the maximum intra-class distance. Despite in the absence of training data for specific signal classes, the proposed FOSSP can distinguish between known and unseen signals. Moreover, FOSSP can determine what kind of unknown class it belongs to when a new signal instance appears. Extensive experiments are conducted on four benchmarks and the results demonstrate the effectiveness of FOSSP.