MoE-KAN-AD: Mixture-of-Experts Kolmogorov-Arnold Autoencoders for Univariate Time Series Anomaly Detection

Abstract

Univariate time series anomaly detection remains difficult because real-world signals are nonlinear, noisy, and often contaminated with outliers. To overcome the parameter explosion and noise sensitivity of standard Kolmogorov-Arnold Networks (KANs), we propose MoE-KAN-AD. Our architecture introduces a shared dictionary of learnable B-spline experts within a Mixture of Experts (MoE) framework. By dynamically routing inputs to these shared experts, the model captures complex local temporal patterns while maintaining a highly compact parameter footprint. Furthermore, because highly expressive B-splines are prone to memorizing outliers, we integrate the Geman-McClure robust loss as an essential regularizer to prevent overfitting on contaminated training data. Evaluated on five benchmark datasets, MoE-KAN-AD achieves competitive detection performance against existing baselines with only 904 parameters. These results highlight that our synergistic design provides a highly efficient and CPU-friendly model.