ADKANet: Multivariate Time Series Anomaly Detection with Kolmogorov-Arnold Network and Transformer

Abstract

The evolving technology landscape and explosive data proliferation make time series anomaly detection critical. Temporal anomalies often indicate operational failures, security threats, or atypical patterns that require urgent analysis. Multivariate time series analysis introduces greater complexity through interdependent variables with temporal evolution, which traditional methods are unable to cope with. To address these issues, we propose ADKANet - an anomaly detection framework that combines Kolmogorov-Arnold Network (KAN) and Transformer architecture. Our approach improves detection accuracy by distinguishing the correlation differences between normal and anomalous data through the correlation difference attention mechanism. At the same time, the computational efficiency of multivariate temporal patterns is optimized by introducing KAN convolution to capture the complex nonlinear characteristics of the data. Experimental validation using benchmark datasets shows excellent performance: the server machine dataset (SMD) achieves an F1 score of 92.17% and the application server dataset (ASD) reaches 96.08%, outperforming existing methods in terms of precision-recall balance. These results confirm that ADKANet is able to decode complex variable interdependencies while maintaining operational efficiency in real detection scenarios.