Facing the complex and changing network environment in the context of modern society, Network Intrusion Detection System (NIDS) is an important network safety defense tool. In this paper, an improved RFS-XGBoost deep learning model is developed to predict and identify potential attacks and intrusions in the network. The model uses a random forest classifier to select features, performs hyperparameter processing and then uses a grid search to optimize the XGBoost method for intrusion type identification. Our experiments were conducted using the UNSW-NB15 dataset, where we trained and tested our models. We then proceeded to compare our results with those obtained using long and short-term memory recurrent neural networks, convolutional neural networks, and decision tree methods. Comparison of the prediction results with actual attack scenarios and comparison of the prediction results of multiple methods show that the provided prediction method is able to select a subset of features with the highest classification accuracy, is more adaptive, and has higher accuracy and faster training speed. The accuracy of the final obtained training model reaches 96.7%, which effectively improves the accuracy and adaptive ability of intrusion detection.