A physical frame segmentation method based on convolutional neural network is innovatively proposed in the present study to address the difficulty of threshold selection in the frame segmentation method and the problem of poor universality of the method. Firstly, the digital sequence is transformed into an image by following three steps including matrix construction, data compression and matrix expansion. Then, the convolutional neural network is trained with the existing samples and the trained convolutional neural network is employed to identify the frame length of the unknown protocol. Finally, on the basis of frame length recognition, the initial position of the frame is identified using correlation filtering method. Therefore, each frame can be extracted from the bit stream. The proposed method, which has higher accurate recognition than existing algorithms suggested by simulation results, has significant potential in engineering application.