KANomaly: Fourier-KAN-based Multi-Scale Patch Mixer for Multivariate Time Series Anomaly Detection

Jiwon Seo, Taehyeong Kim, Sung Won Han

Keywords: Multivariate Time Series Anomaly Detection, Kolmogorov–Arnold Networks, Fourier Basis Functions

TL;DR: KANomaly combines Fourier-KAN with multi-scale patching to detect point and pattern anomalies, capturing frequency-aware structures across scales.

Abstract: Multivariate Time Series Anomaly Detection (MTSAD) is crucial for system stability in domains such as industrial monitoring, yet rare events, nonlinear dependencies, and limited labels necessitate unsupervised methods. However, existing approaches struggle to model subtle anomalies and detect diverse patterns, as they rarely integrate explicit frequency-domain representations and rely on fixed-scale analysis. To address these limitations, we propose KANomaly, a novel model inspired by Fourier-based Kolmogorov–Arnold Networks (KANs) for MTSAD. The model incorporates three key innovations: (i) Fourier basis functions embedded within the KAN architecture to capture subtle periodic and spectral anomalies; (ii) a coarse-to-fine multi-scale patching strategy that enhances detection of both point and pattern anomalies; and (iii) a Fourier-KAN Mixer that aggregates information across channel, patch, and temporal dimensions to model complex local and global interdependencies. Extensive experiments demonstrate that KANomaly consistently outperforms state-of-the-art models on multiple real-world datasets, validating the effectiveness of each component.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17736