Abnormal traffic detection utilizes various techniques to analyze and identify potential attacks within network traffic, playing a critical role in securing and monitoring network environments. To address the problem of low detection accuracy caused by the imbalanced distribution of normal traffic and unknown attack traffic in abnormal traffic detection, we propose a method called Synthetic Minority Over Sampling Technique and Wasserstein GAN-Based Convolutional Neural Network-Bidirectional Long Short-Term Memory (SWGCNN-BiLSTM). For imbalanced samples, SWG, which combines SMOTE and WGAN, is employed to balance the minority class samples. Subsequently, CNN is utilized to extract the spatial features of the samples, followed by BiLSTM to capture the temporal features. Our method is evaluated on the NSL-KDD dataset, achieving an accuracy of 98.67%. Experimental results demonstrate that our method enhances the accuracy of unknown attack traffic detection and outperforms detection methods based on traditional Machine Learning and Deep Learning.