Anomnet: a Hybrid Framework for Multivariate Time Series Anomaly Detection Using Attention-Augmented State Space Models

Anomaly detection in multivariate time series plays a crucial role in diverse applications, from monitoring system performance to detecting financial fraud. Conventional deep learning approaches often face challenges in capturing intricate distributional changes and effectively combining both local and global anomaly insights. In this study, we present AnomNet, a hybrid framework for anomaly detection that merges state space models with attention-based mechanisms to detect anomalies at both granular and broad levels. By integrating dynamic sequence modeling and a Kolmogorov-Arnold Network (KAN) to ensure data consistency, AnomNet facilitates superior feature differentiation and anomaly identification. The attention mechanism significantly enhances the framework's ability to capture temporal relationships, while a novel reconstruction strategy boosts the detection of anomalies across diverse patterns. Comprehensive experiments on benchmark datasets show that AnomNet surpasses current methods in accuracy, resilience, and the detection of both detailed and large-scale anomalies, setting a new benchmark for time series anomaly detection in complex domains.