Radio frequency fingerprint (RFF) identification (RFFI) is a promising technique for device authentication at the physical layer of the communication stacks. However, practical challenges, particularly in low signal-to-noise ratio (SNR) scenar-ios, and the lack of comprehensive studies on open-set recognition hinder the widespread application of RFFI. This paper presents an unsupervised open-set RFF identification algorithm designed to address the robustness challenges associated with low SNR. Our approach integrates the Noise2Noise method for de noising, drawing inspiration from its successful applications in image and speech processing. The proposed framework utilizes an image-based autoencoder (AE) to extract features from the differential constellation trace figure (DCTF) of the signals after Noise2Noise denoising. The open-set recognition task is performed by cosine distance measurement. We carried out extensive experimental evaluation involving 18 ZigBee devices and a USRP software-defined radio platform. Our proposed method can achieve a gain up to 25% under low SNRs.