Abstract: The identification of routers is significant for maintaining network security. Traditional router feature recognition fingerprints have problems such as difficulty in data acquisition, hardware limitations, and complex fingerprint feature analysis. For this reason, a router identification method based on deep learning is proposed. The high-precision delay consumed by the data packets being forwarded and processed in the router, which was an inevitable side channel information, was used as the basis for device fingerprint identification. ResFGRU model that combined the advantages of the fully convolutional network, gated recurrent neural network and residual network was proposed to improve the identification accuracy. The experiments show that using this model to classify on the constructed router high-precision delay dataset can achieve 99.8% recognition accuracy, proving the effectiveness of high-precision delay as the characteristic fingerprint of router devices for classification and recognition.