Anomaly Detection in Time Series via Adversarial Autoencoder with Reverse Sampling

This paper proposes RevAAE, an unsupervised time series anomaly detection model based on Adversarial Autoencoder (AAE). To address the overgeneralization issue in traditional AE-based models, RevAAE introduces a novel reverse sampling strategy and an output discriminator to transform latent representations of anomalies into normal reconstructions. Additionally, a “Same Sample for Adjacent Time Series” strategy is adopted to stabilize adversarial training. Experiments on three public datasets show that RevAAE outperforms state-of-the-art statistical and AE-based methods.