Time series anomaly detection based on TimeGAN and LSTM neural network

Anomaly detection is one of the most mature applications in time series data analysis. However, due to the nonlinearity of time step and sequence, it often leads to abnormal state of a specific step in time series. On the one hand, previous GAN studies focused on the method of generating sequence data, but did not pay attention to the autocorrelation of sequences. After all, a good model should learn not only the distribution of features across each timestamp, but also the underlying relationships between different points in time. On the other hand, a good generation model should ensure the generation of high-quality synthetic data. Most of the researches on this aspect are based on RCGAN, so the TimeGAN applied in this paper is also conducted on this basis, and the visualization analysis of data is combined with T-SNE to evaluate the quality of the model qualitatively. Then LSTM neural network is used to evaluate the influence of periodicity, noise, step size and other factors on the model. The results show that both methods can detect anomalies effectively, but TimeGAN is better than traditional RCGAN, which verifies the feasibility of this method.