This article describes how to use Transformer models to detect irregular network traffic in order to enhance model detection performance. By comparing traditional machine learning and deep learning methods, the Transformer model is validated for its superiority in processing high-dimensional data and capturing timing features. In the experiment, compared with traditional methods, the Transformer model achieved an accuracy of up to 98.5%. In the second experiment, the effectiveness of the multi-head self-attention mechanism in improving the accuracy of network traffic anomaly detection was also verified. When the data scale is 10,000, the accuracy of Transformer model reaches 94.2%. In the real-time detection experiment, Transformer model also has good real-time detection ability. Experimental data results show that Transformer model has significant advantages and practicability in network traffic anomaly detection.