Research map / Time series / Anomaly detection
SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
SARAD trains a Transformer autoencoder on transposed time windows to learn spatial feature associations, then autoencodes the reduction of those associations (progression) to detect and diagnose multivariate time series anomalies.
Problem
Time series anomaly detection and diagnosis in industrial control and service monitoring systems mainly relies on temporal modeling, which overlooks or dilutes spatial (inter-feature) associations that anomalies disrupt.
Approach
Apply an encoder-only Transformer to transposed windows to learn pairwise inter-feature associations; split windows into halves to compute non-negative association decreases (progression) and autoencode these in an MLP, combining data-space and association-space reconstruction errors into a joint anomaly score.
Key results
- SARAD reaches 79.97 VUS-ROC on SMD, outperforming the best baseline MAD-GAN (64.35 AROC) by about 15.5 points on VUS-ROC as claimed in text.
- On HAI, SARAD achieves 96.87 AUC-ROC and 96.17 VUS-ROC, the best among compared methods.
- For diagnosis, SARAD gets 56.73 HR@100% and 71.93 HR@150% on SMD, versus 33.33 and 45.26 for TranAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ReLU to association progression to retain reduction-only changes (ablation) | SARAD · SMD, PSM, SWaT, and HAI | VUS-ROC: 75.16 -> 79.67; 60.21 -> 61.77; 86.75 -> 87.52; 94.93 -> 96.17 | Helped |
| use column-sum instead of row-sum aggregation (ablation) | SARAD · SMD, PSM, SWaT, and HAI | VUS-ROC: 79.17 -> 79.67; 63.20 -> 61.77; 87.40 -> 87.52; 95.80 -> 96.17 | Mixed |
| sum progression values during aggregation rather than concatenate them (ablation) | SARAD · SMD, PSM, SWaT, and HAI | VUS-ROC: 56.98 -> 79.67; 47.34 -> 61.77; 85.26 -> 87.52; 90.70 -> 96.17 | Helped |
| reconstruct progression instead of using progression directly as the anomaly criterion (ablation) | SARAD · PSM, SWaT, and HAI | VUS-ROC: 60.49 -> 61.77; 86.70 -> 87.52; 94.98 -> 96.17 | Helped |
| combine data reconstruction and progression reconstruction errors instead of using data error only (ablation) | SARAD · SMD, PSM, SWaT, and HAI | VUS-ROC: 72.72 -> 79.67; 61.45 -> 61.77; 86.15 -> 87.52; 95.56 -> 96.17 | Mixed |
| combine data and progression reconstruction errors instead of using progression error only (ablation) | SARAD · SMD, PSM, SWaT, and HAI | VUS-ROC: 79.39 -> 79.67; 62.20 -> 61.77; 65.31 -> 87.52; 94.77 -> 96.17 | Mixed |
Surprising findings
- Removing the ReLU from progression (capturing both increases and decreases) hurts detection, and using progression alone underperforms on most datasets except SMD.
Limitations
- Model time complexity is quadratic in number of features, limiting scalability to very large systems.
- Scarcity of forensically labeled datasets for diagnosis; SWaT and HAI diagnosis labels only mark attack origins.
Experimental setup
Datasets: SMD PSM SWaT HAI
Benchmarks: SMD PSM SWaT HAI
Metrics: AUC-ROC AUC-PR VUS-ROC VUS-PR HR@P% NDCG@P% IPS@P% Precision Recall F1
Baselines: IF DIF TranAD ATF-UAD AT DCdetector USAD GDN MAD-GAN DiffAD
Builds on: Transformer iTransformer USAD Isolation Forest
Closest papers by meaning
- Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.90)
- Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.86)
- GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.86)
- CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling (NeurIPS 2025 · similarity 0.86)
- ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.86)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.