In today’s complex connected network environment, any abnormal network attack may cause serious problems, ranging from loss of data to great economic and social losses. How to distinguish normal traffic and abnormal attack traffic by analyzing the characteristics of network traffic has become the focus of network security research, among which the anomaly detection system is a very important branch. At present, neural network has been widely used in anomaly detection. Most studies combine CNN and LSTM structure to extract spatiotemporal characteristics of network traffic to classify abnormal traffic. However, when LSTM is used for high-dimensional original traffic data, the gradient will disappear, and the training process is complex and time-consuming. Therefore, based on the convolutional neural network CNN, this paper proposes a model CONBSCNN with global feature correlation ability. This paper proposes the CON structure and combines it with the BSCNN structure to form the CONBSCNN model. This network model can fully learn the interaction protocol characteristics and timing information characteristics of traffic. Compared with the traditional BSCNN, the classification accuracy of UNSW-NB15, KDDCUP99 and as well as the encrypted traffic data collected by ourselves has been improved, and the accuracy of KDD data sets is 99.9%. In addition, CON structure is added to other CNN models, and the recognition accuracy is also improved effectively. This is a 2% improvement on both datasets.