Cross-Domain Adaptive Multi-Scale Representation Learning for Unified Time Series Anomaly Detection

ICLR 2026 Conference Submission17085 Authors

Keywords: anomaly detection, cross-domain, multi-granularity

TL;DR: We propose UniAnomaly, a cross-domain anomaly detection framework equipped with a multi-scale encoder that effectively captures temporal dependencies at different granularity.

Abstract: Time series anomaly detection has received growing attention due to its importance in a wide range of real-world applications. However, two critical challenges remain under-explored. First, most existing methods train separate models for individual domains, which severely limits their generalization ability and neglects the potential of cross-domain anomaly detection. Second, when extending to the cross-domain setting, the inconsistency of temporal granularityacross datasets makes it difficult to learn unified representations. To address these issues, we propose UniAnomaly, a cross-domain anomaly detection framework equipped with a multi-scale encoder that effectively captures temporal dependencies at different granularity. Our approach enables robust and transferable representation learning across heterogeneous datasets. Extensive experiments on multiple real-world benchmarks demonstrate that UniAnomaly consistently achieves state-of-the-art performance, highlighting the effectiveness of cross-domain multi-scale modeling for time series anomaly detection. Our code is available at https://anonymous.4open.science/r/UniAnomaly-B923.

Primary Area: learning on time series and dynamical systems

Submission Number: 17085