With the rapid development of the Internet, network security is becoming increasingly important. However, the high dimensionality of network traffic characteristics and the complexity of data make traditional intrusion detection models have long response times and low detection accuracy. This article constructs an FDA-PCC feature selection algorithm for multi-stage feature selection. Meanwhile, we use Convolutional Neural Network for local perception, combined with multi-layer BiLSTM to capture contextual information, extract long-range dependent features, and finally use the softmax function for classification. Using PSO algorithm to optimize the model network structure, allowing the model to autonomously select the optimal network structure and initial parameters under non empirical guidance. The experimental results on the NSL-KDD dataset show that after preprocessing and feature extraction, the multi classification detection performance of the dataset on this model is significantly improved compared to existing traditional detection models.