Research map / Time series / Anomaly detection
When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection
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.
Proposes a test-time adaptation strategy for unsupervised time-series anomaly detection that detrends inputs with an EMA trend estimate and updates an autoencoder online using only self-predicted normal instances, addressing the "new normal" distribution-shift problem.
Problem
Unsupervised time-series anomaly detectors trained on normal data degrade under distribution shift between training and test data, producing false alarms when normal behavior changes (the "new normal problem").
Approach
Combine an exponential-moving-average trend estimate used to detrend inputs with fully test-time online updates of a reconstruction-based detector, where only instances the model itself predicts as normal are used for gradient updates.
Key results
- MLP+Ours reaches AUROC 0.624 on WADI versus 0.485 for off-the-shelf MLP, and AUPRC 0.575 on MSL (P-15) versus 0.061.
- On the AnoShift benchmark, MLP+Ours improves FAR-split AUROC by 0.216 over the base MLP (0.208 to 0.424).
- Ablation on MSL (P-15) shows detrending alone and TTA alone each underperform the combination (F1 0.440 with both vs 0.221 DT only and 0.019 TTA only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add detrending and test-time adaptation to MLP | MLP · WADI | F1: 0.131 -> 0.148 | Hurt |
| Add detrending and test-time adaptation to MLP | MLP · Yahoo (A1-R20) | F1: 0.067 -> 0.678 | Helped |
| Add detrending and test-time adaptation to MLP | MLP · CreditCard | F1: 0.127 -> 0.135 | Mixed |
| Add detrending only; omit test-time adaptation (ablation) | MLP · SWaT | F1: 0.765 -> 0.762 | No clear effect |
| Add test-time adaptation only; omit detrending (ablation) | MLP · SWaT | F1: 0.765 -> 0.784 | Helped |
| Add detrending only; omit test-time adaptation (ablation) | MLP · SMD (M-2-1) | AUROC: 0.674 -> 0.749 | Helped |
| Add test-time adaptation only; omit detrending (ablation) | MLP · MSL (P-15) | F1: 0.263 -> 0.019 | Hurt |
| Combine detrending and test-time adaptation (ablation) | MLP · MSL (P-15) | AUPRC: 0.061 -> 0.575 | Helped |
| Add MLP test-time adaptation to the MLP baseline | MLP · AnoShift FAR | AUROC: 0.208 -> 0.424 (+0.216) | Helped |
| Add MLP test-time adaptation to the MLP baseline | MLP · AnoShift NEAR | AUROC: 0.441 -> 0.639 (+0.194) | Helped |
Surprising findings
- Test-time model updates without detrending can substantially harm performance (e.g., MSL F1 drops to 0.019 with TTA alone), as the model overfits to pre-shift sequences.
- Anomaly Transformer shows comparable F1-PA but much lower F1, AUROC and AUPRC than simple MLP baselines on several benchmarks.
Limitations
- Assumes test streams contain more normal than anomalous instances; ambiguous labeling of distribution shift vs. anomaly.
- Threshold selection relies on training statistics; performance depends on threshold choice, e.g., WADI F1 lower than baseline.
- Experiments focus on an MLP autoencoder; applicability to other backbones is claimed but not shown.
Experimental setup
Datasets: SWaT WADI SMD SMAP MSL CreditCard Yahoo A1 AnoShift
Benchmarks: SWaT WADI SMD SMAP MSL CreditCard Yahoo A1 AnoShift
Metrics: F1 F1-PA AUROC AUPRC AUPRC-in AUPRC-out
Baselines: MLP LSTMEncDec USAD THOC Anomaly Transformer SO-GAAL deepSVDD LUNAR ICL BERT
Builds on: fully test-time adaptation online gradient descent Online RNN-AD anomaly transformer
Closest papers by meaning
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.89)
- Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.87)
- Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.87)
- Position: Quo Vadis, Unsupervised Time Series Anomaly Detection? (ICML 2024 · similarity 0.87)
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.87)
- Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.86)
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.