Currently, the safety of subway tunnels mainly relies on the manual track inspection of subway track inspectors when there are no trains on the track. This method is slow and inefficient, and the inspection results are completely dependent on the experience and status of the track inspector. Aiming at this problem, a video anomaly alarm system of subway tunnel inspection based on deep separable convolution is proposed, this system uses the proposed SubwayNet convolutional neural network to complete the classification of inspection video images. The built-in convolutional neural network is trained and saved by using the produced subway tunnel inspection dataset. The graphical user interface is created and the alarm function is added. Finally, the program files are packaged into an executable file. The experimental results show that the classification accuracy of the system can reach 96%, and the speed can reach 52 frames/second, which meets the requirements of real-time and accurate analysis of video.