Abstract: In order to make the network intrusion detection model have the ability of self-adaptive learning for new network traffic data without destroying the learned performance, this paper proposes an intrusion detection method based on incremental Multi-Scale Convolution Neural Network (iMSCNN). Firstly, by training the parameters in the Multi-Scale Convolution Neural Network (MSCNN) model, the local features of the original network traffic data are extracted and spliced. Then the model is controlled, and the update of parameters in the controlled element is frozen. Next, the Linear Converter (LC) is set for each convolution layer as the trainable parameter in the incremental learning process, and the new network traffic is learned by the fewer parameters in the LC, so that the intrusion detection model is updated. Finally, the abnormal network traffic is identified and detected by the iMSCNN intrusion detection model. The experimental results verify the effectiveness of this method, and prove that this incremental intrusion detection method can retain the detection performance and improve the efficiency of learning new knowledge.