CGSTA: Cross-Scale Graph Contrast with Stability-Aware Alignment for Multivariate Time-Series Anomaly Detection

TL;DR

The CGSTA framework is proposed, which fuse the multi-scale and temporal features and use a conditional density estimator to produce per-time-step anomaly scores and is comparable to the baseline methods on SWaT and SMAP.

Abstract

Multivariate time-series anomaly detection is essential for reliable industrial control, telemetry, and service monitoring. However, the evolving inter-variable dependencies and inevitable noise render it challenging. Existing methods often use single-scale graphs or instance-level contrast. Moreover, learned dynamic graphs can overfit noise without a stable anchor, causing false alarms or misses. To address these challenges, we propose the CGSTA framework with two key innovations. First, Dynamic Layered Graph Construction (DLGC) forms local, regional, and global views of variable relations for each sliding window; rather than contrasting whole windows, Contrastive Discrimination across Scales (CDS) contrasts graph representations within each view and aligns the same window across views to make learning structure-aware. Second, Stability-Aware Alignment (SAA) maintains a per-scale stable reference learned from normal data and guides the current window's fast-changing graphs toward it to suppress noise. We fuse the multi-scale and temporal features and use a conditional density estimator to produce per-time-step anomaly scores. Across four benchmarks, CGSTA delivers optimal performance on PSM and WADI, and is comparable to the baseline methods on SWaT and SMAP.

Figures (4)

• Figure 1: Overview of the proposed CGSTA framework for multivariate time series anomaly detection. The framework consists of (i) DLGC, which builds local, regional, and global graphs with nodes as variables and edges as temporal dependencies; (ii) CDS, which learns discriminative and consistent representations across graph layers; and (iii) SAA maintains stable graphs and enforces consistency with dynamic ones.
• Figure 2: Sensitivity of AUROC and AUPRC to Hyperparameters: CDS weight $\alpha$ ($[0.70,0.95]$) and EMA momentum $\gamma$ ($[0.80,0.95]$). Results are shown as mean$\pm$std across PSM, WADI, SWaT, and SMAP (left: AUROC; right: AUPRC).
• Figure 3: Sensitivity to SAA’s weight $\beta$ — AUROC and AUPRC across datasets
• Figure 4: Case on SWaT (index 7894, label Attack). Top: per-sensor anomaly scores for Top-$K$ sensors (brighter is more anomalous). Bottom (from left to right): dynamic local dependency (row-wise softmax), regional adjacency (group-wise), global adjacency (cross-group), $\Delta_{\text{local}}=|A_{\text{local}}^{\text{dyn}}-A_{\text{local}}^{\text{stable}}|$, and the EMA-anchored stable local graph. Bright vertical bands in the dynamic graph indicate hub targets (e.g., P601/FIT401); $\Delta_{\text{local}}$ concentrates on edges incident to these hubs, evidencing structural deviations rather than transient noise.