Enhancing Lightweight Log Anomaly Detection with Deep Semantic Encoding and Nonlinear Feature Compression

Abstract

Anomaly detection within system logs remains a critical task aimed at maintaining the stability as well as reliability of modern software systems. As large-scale cloud and distributed computing environments continue to expand, system logs have become increasingly voluminous, heterogeneous, and complex, making traditional rule-based and manual analysis approaches less effective. Although learning-based methods have shown promising potential, most existing approaches still struggle to maintain an effective trade-off across detection performance and computational efficiency, particularly in resource-constrained environments. In this work, a lightweight and context-aware log anomaly detection approach, referred to as Lightweight Semantic TCN (LS-TCN) is introduced based on an enhanced Temporal Convolutional Network (TCN). The proposed method relies on compact semantic log representations and incorporates a nonlinear dimensionality reduction strategy based on PCA–tSNE to preserve discriminative information while reducing model complexity. In addition, the TCN architecture is designed to efficiently capture temporal dependencies in log sequences. Experimental results across the BGL and HDFS benchmark datasets indicate that our method achieves competitive or improved detection performance, while maintaining lower computational cost.