Log Semantic Graph Construction and System Event Anomaly Detection via Convolutional Neural Networks

Abstract

This paper proposes a system event anomaly detection method that integrates convolutional semantic encoding with structure-aware graph fusion. The method first introduces a convolution-based log semantic modeling module, which uses a multi-scale convolutional perception mechanism to extract contextual features and build embedding representations from raw log sequences. This significantly enhances the ability to capture local semantic dependencies and behavioral cues. Then, a structure-aware graph fusion module is designed to construct dynamic graphs based on event associations, modeling and reinforcing the potential interactions among logs. This captures multiple dependencies of log events across both temporal and structural dimensions. To improve the discriminability of the embedding space and the consistency of structural representations, a joint optimization objective is constructed. It combines contrastive loss and structural alignment constraints for collaborative training, enabling accurate semantic differentiation and anomaly detection at the event level. The proposed method demonstrates strong semantic parsing and structural modeling capabilities, achieving precise anomaly detection and pattern extraction in high-dimensional and sparse system logs. Modeling experiments on real log data verify the effectiveness and applicability of the approach in complex system scenarios.