Work

Structure-Aware Cross-Layer Gated Attention Transformer for Unsupervised Multivariate Server Time-Series Anomaly Detection

Year: 2026
Type: preprint
Abstract: Unsupervised anomaly detection in multivariate server time series is challenging because temporal dependencies vary across variables and time scales. Reconstruction-based Transformers can model them, but they often lose fine-grained temporal detail, reuse cross-layer context indirectly, and rely mainly on reconstruction error. This paper proposes the Cross-Layer Gated Attention Transformer (CLGAT), an unsupervised reconstruction framework for server monitoring data. CLGAT combines four components. Overlapping Patch Embedding preserves local context while restoring the original temporal resolution through linear interpolation. Cross-Layer Dynamic Context Residual retrieves representations from preceding encoder layers and injects them into the current layer through time-step gating. Dual-stage Gated-Attention modulates temporal attention weights and aggregated feature responses. The Window Attention Structure aggregates selected long-range attention responses, normalizes them through temperature-controlled chunk-wise Softmax, and fuses the resulting structural score with a Mean--Top-$K$ reconstruction score. The threshold is set at the 99.5th percentile of validation anomaly energy. On the SMD benchmark, CLGAT achieves 94.59\% Precision, 96.04\% Recall, and 95.31\% F1-score under the point-adjusted protocol, outperforming 15 baselines. Ablation confirms all four components contribute, with the largest drop after removing the Window Attention Structure.
Language: English
Open Access status: green
OpenAlex ID: W7235939431
Cites:
Cited by:
Related to: 10