Viscoelastic characterization is of great importance for the investigation of biomolecular condensates. Single-particle-tracking-based rotational diffusion analysis of single nanorods is an effective approach for quantitative viscosity measurement. However, in the case of high background and noise with high-speed image acquisition, accurate extraction of diffusivity from the data is a challenging task. Here, we develop a novel frequency-domain-based deep learning (DL) method for single nanorod rotational tracking analysis. We synthesized Brownian rotational time-series data for training, designed a data preprocessing module to reduce the effect of noise, and extracted rotational diffusion coefficient using recurrent neural networks in the frequency domain. Compared with the traditional curve-fitting-based methods, our method shows higher accuracy and a wider detection range for viscosity measurement. We verified our method using experimental data from plasmonic imaging of single gold nanorods (AuNRs) in glycerol solution and PGL droplets. Our method can be potentially applied to the viscosity measurement of different biomolecular condensates in vitro and in vivo.