Dual-Pass Autoencoder: Tackling Temporal Drift in Time Series Anomaly Detection

Detecting anomalies in multivariate time series remains difficult when data distributions shift over time. Existing approaches often struggle to remain effective under such temporal drift. We present a Dual-Pass Autoencoder, a new architecture that directly addresses this issue by learning two complementary perspectives of the data: one from the raw sequence and another from its temporal context. These dual representations are fused through a drift-adaptation mechanism, enabling the model to distinguish true anomalies from natural distributional changes. Unlike conventional autoencoder-based methods, our design explicitly incorporates temporal self-attention and adaptive latent fusion, resulting in drift-aware embeddings that are both expressive and robust. The reconstructed feature vectors, supported by an auxiliary classifier, provide strong discriminative capacity for anomaly detection. Comprehensive experiments on diverse real-world datasets demonstrate that our approach consistently outperforms leading baselines in both accuracy and robustness, establishing Dual-Pass Autoencoder as an effective solution for anomaly detection in evolving temporal environments.