Aiming at the problem of low accuracy of network traffic categorization due to the temporal, long sequence characteristics and category imbalance that network traffic data usually have, this paper proposes a hybrid network architecture that combines bidirectional long and short-term memory network and Informer model. The method preprocesses the raw traffic data by employing the edit nearest neighbor algorithm undersampling and Kmeans-SMOTE algorithm oversampling operations on the network traffic data, and then extracts the temporal and spatial features using the fusion of the Informer and the bi-directional long and short-term memory neural network models. The method is validated on the NSL-KDD dataset and the UNSW-NB15 dataset. The experimental results show that the method has an accuracy of 84.73% and 83.5% in recognizing encrypted traffic services and has better classification performance compared to the latest deep learning models.