To address the low classification and recognition rate of radio frequency fingerprints in internet of things (IoT) terminal devices, this paper proposes an improved convolutional neural network-gated recurrent unit (CNN-GRU) method based on comb filtering for radio frequency signal "gene" classification and recognition. First, the time-frequency characteristics of RF signals are enhanced using a comb filter to construct a unique "gene map" for each device. Second, the traditional one-dimensional CNN is expanded into a three-layer two-dimensional structure, combined with a dual-layer GRU to achieve joint time-frequency feature extraction and sequence modeling. Finally, hybrid pooling and exponential linear unit (ELU) activation functions are introduced to optimize feature representation. Experimental results show that the proposed method achieves a identification accuracy of 100% in simulated data and 95.52% in real-world data, outperforming traditional algorithms by 5%-22%, significantly enhancing the security and manage-ment efficiency of IoT devices.