Abstract: Radio frequency fingerprinting(RFF) identification, a crucial technology in physical layer security, enables device authentication through the analysis of hardware impairment characteristics embedded within signals. To address the semantic shift challenge posed by unknown radiation sources in open environments, this paper presents CutOE, a novel framework for open-set RFF identification that implements open-set recognition through both data-centric and loss function approaches. At the data level, At the data level, considering the specific characteristics of electromagnetic data, unlabeled anomaly samples are introduced during the training phase, and a random cut-smooth technique is employed to generate virtual anomaly samples, effectively expanding the prior knowledge of the open space. At the loss function level, the framework introduces soft labels and Jensen-Shannon divergence loss to constrain the model's confidence decay smoothly, while implementing efficient open-set recognition based on confidence thresholds. Extensive experiments conducted on Wi-Fi datasets demonstrate that under high signal-to-noise ratio conditions, the CutOE framework achieves F1 scores of 93.57% and 72.2% at openness ratios of 0.105 57 and 0.396 98, respectively. Furthermore, ablation studies validate the effectiveness and scalability of each framework component, with random cutting strategies and Jensen-Shannon divergence significantly enhancing open-set recognition performance.