A method for detecting spatio-temporal correlation anomalies of WSN nodes based on topological information enhancement and time-frequency feature extraction

TL;DR

This paper introduces TE-MSTAD, a novel anomaly detection method for WSNs that enhances spatio-temporal feature extraction using topology and time-frequency analysis, achieving high accuracy.

Contribution

The paper proposes a topology-enhanced spatio-temporal feature fusion method combining graph neural networks and dual-branch networks for improved WSN anomaly detection.

Findings

Achieved F1 scores of 92.52% and 93.28% on public and real datasets.

Outperforms existing methods in detection accuracy and generalization.

Effectively captures spatial and temporal correlations with reduced computational overhead.

Abstract

Existing anomaly detection methods for Wireless Sensor Networks (WSNs) generally suffer from insufficient extraction of spatio-temporal correlation features, reliance on either timedomain or frequencydomain information alone, and high computational overhead. To address these limitations, this paper proposes a topology-enhanced spatio-temporal feature fusion anomaly detection method, TE-MSTAD. First, building upon the RWKV model with linear attention mechanisms, a Cross modal Feature Extraction (CFE) module is introduced to fully extract spatial correlation features among multiple nodes while reducing computational resource consumption. Second, a strategy is designed to construct an adjacency matrix by jointly learning spatial correlation from time-frequency domain features. Different graph neural networks are integrated to enhance spatial correlation feature extraction, thereby fully capturing spatial relationships among multiple nodes. Finally, a dualbranch network TE-MSTAD is designed for time-frequency domain feature fusion, overcoming the limitations of relying solely on the time or frequency domain to improve WSN anomaly detection performance. Testing on both public and realworld datasets demonstrates that the TE-MSTAD model achieves F1 scores of 92.52% and 93.28%, respectively, exhibiting superior detection performance and generalization capabilities compared to existing methods.