Class Feature Space Reconstruction for Automatic Modulation Open Set Recognition

Automatic modulation classification (AMC) is a critical technology in the field of wireless communications. However, as the wireless communication environment becomes increasingly complex, the received signal may include unknown modulation types. Traditional closed-set recognition algorithms assume that all test samples belong to known categories. Consequently, these algorithms often misclassify unknown modulation types, failing to meet practical needs. To address this issue, in this paper, we propose a Deep Learning-Based Class Feature Space Reconstruction (CFSR) method for open-set recognition of radio signals, aiming to achieve accurate classification of known categories and effective detection of unknown categories. First, CFSR sets a specific class feature space for each known class, and the corresponding class samples are fitted using the AE manifold in the class feature space, thereby overcoming the class feature overlap problem caused by compressing the intra-class features into a limited space. Then, The class space open-set loss is introduced based on the class feature space. By treating the remaining class samples except the current class as unknown samples, unknown information is effectively introduced during the training process, thereby enhancing the open set recognition ability of the model. Finally, the experimental results on multiple public datasets show that the proposed CFSR has great open-set recognition performance and outperforms the existing open-set recognition algorithms.