Research on Time Series Anomaly Detection Algorithm Based on Transformer Coupled with GAN
Mengfei Ye, Zhanquan Wang, Li Fei · 2024
Timing series anomaly detection plays an important role in several fields. Currently existing methods lack generators with strong generalization ability, and do not sufficiently consider contextual features, time series anomaly detection algorithm based on Transformer coupled with GAN is proposed to solve the problems. The method uses frequency domain preprocessing to extract contextual features, then Transformer-like encoder and decoder architecture is designed by using multi-head self-attention and positional encoding mechanisms to enhance the temporal pattern learning ability, finally uses temporal con-volutional neural network-based discriminator for adversarial training to obtain temporal data anomaly detection results. Experiments on four public datasets show that the algorithm exhibits good performance in time series anomaly detection, especially the F1-score in SMD and MSL reaches 0.974 and 0.968 respectively, which surpass the current state-of-the-art time series anomaly detection algorithm by 1.5%, and the overall experimental results show that the algorithm is reasonable and effective.