TSI-GAN: Unsupervised Time Series Anomaly Detection Using Convolutional Cycle-Consistent Generative Adversarial Networks

Anomaly detection is widely used in network intrusion detection, autonomous driving, medical diagnosis, credit card frauds, etc. However, several key challenges remain open, such as lack of ground truth labels, presence complex temporal patterns, and generalizing over different datasets. This paper proposes TSI-GAN, an unsupervised anomaly model for time-series that can learn patterns automatically generalize well, i.e., no need choosing dataset-specific parameters, making statistical assumptions about underlying data, or changing architectures. To achieve these goals, we convert each input into a sequence 2D images using two encoding techniques with the intent capturing various types deviance. Moreover, design reconstructive GAN uses convolutional layers encoder-decoder employs cycle-consistency loss during training to ensure inverse mappings are accurate well. In addition, also instrument Hodrick-Prescott filter post-processing mitigate false positives. We evaluate TSI-GAN 250 well-curated harder-than-usual datasets compare 8 state-of-the-art baseline methods. The results demonstrate superiority all baselines, offering overall performance improvement 13% 31% second-best performer MERLIN third-best LSTM-AE, respectively.