STGAD: Saliency–Local Overlap@5 vs. Max|Δ| for Multivariate Time Series Anomaly Detection

STGAD is a dual-score generative-adversarial framework for unsupervised anomaly detection in multivariate time series. The 5.5 KB XLS file, uploaded by Xiao Liao on May 21, 2026, contains results comparing the model's saliency-local overlap metric against a baseline. Experiments were conducted on five benchmark datasets covering server monitoring, aerospace telemetry, industrial control, and ECG signals.

Use Cases

• Benchmarking anomaly detection models based on the saliency-local overlap@5 metric described in the paper.
• Analyzing the stability of adversarial learning frameworks for time-series modeling as discussed in the description.
• Evaluating the effectiveness of dual-score fusion for anomaly localization in multivariate sensor data.
• Studying temporal dependency modeling using Transformer encoders within a WGAN-GP critic architecture.

Strengths

• Results are derived from experiments on five established benchmark datasets (SMD, SMAP, MSL, SWaT, MIT-BIH).
• The dataset is openly available under a CC-BY-4.0 license.
• The associated research paper provides a detailed methodological framework for the results.

Limitations

• The dataset is very small at 5.5 KB, suggesting it contains summary results rather than raw time-series data.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for certain analyses.

Provenance

Source

Xiao Liao via figshare.

Collection Method

Likely contains experimental results (e.g., metric comparisons) from the STGAD model evaluation.

Freshness

Last updated 2026-05-21 17:42:23; freshness should be verified.