Seeing Patterns Differently: Topological Geometry for Anomaly Detection in Multivariate Time Series

Anomaly detection in time-series data is essential for safeguarding critical systems against malfunction, cyber intrusion, and costly downtime. Recent deep-learning approaches achieve significant results, yet segment-level detection in multivariate streams remains challenging since high-dimensional spatio-temporal dependencies demand large, computationally intensive models. To address this challenge, we use a geometric view of time series to achieve comparable or better accuracy with lower overhead. In this paper, we show that mapping multivariate time series into a stable topological representation, followed by classification with a lightweight neural network, achieves competitive segment-level detection while drastically reducing model complexity. Compared to purely deep learning baselines, our topology-driven framework captures inherent shape information that distinguishes normal from anomalous behavior without expensive feature extraction. Across two public benchmark datasets, the Server Machine Dataset (SMD) and the Controlled Anomalies Time Series (CATS), our method achieves 96.1% and 97.8% accuracy respectively, performing on par with or better than state-of-the-art models while requiring significantly less inference time and offering interpretable geometric insights.