Time-Series Anomaly Detection Benchmark Datasets
Match the dataset to your domain: NAB for general streaming anomaly research, the Server Machine Dataset for IT operations, NASA SMAP/MSL for aerospace telemetry, SWaT/WADI for industrial control systems, and the UCR anomaly archive when you need many precisely labeled single-anomaly series to stress-test an algorithm. Below is what each dataset contains, its access terms, and what to watch for in benchmark comparisons.
The short answer
There is no single default time-series anomaly dataset the way COCO is for vision — the right choice depends on your domain and on how the anomalies are labeled. NAB is the broadest general-purpose starting point; SMD and NASA SMAP/MSL are the standard references for IT-operations and spacecraft-telemetry research respectively; SWaT and WADI cover industrial control systems but require a signed data-use agreement; the UCR anomaly archive is best when you want cleanly isolated, unambiguous anomalies to validate an algorithm rather than a noisy production-like signal.
General-purpose streaming anomaly benchmarks
Numenta Anomaly Benchmark (NAB)
NAB provides 58 labeled time-series files combining real-world data (AWS server metrics, Twitter volume, road traffic) with synthetic series designed to test specific anomaly patterns, along with a scoring system built for streaming, real-time detection rather than batch evaluation. It is a good first stop when you want a small, varied, well-documented benchmark to sanity-check a new detection method before moving to a domain-specific dataset. It is MIT licensed. Get it from github.com/numenta/NAB.
UCR Time Series Anomaly Archive
The UCR anomaly archive, maintained by Eamonn Keogh's group with Hexagon ML, was built specifically to fix problems in earlier benchmarks — many series in older datasets contain multiple ambiguous or mislabeled anomalies that let simple heuristics score artificially well. Each UCR series instead contains exactly one clearly defined anomaly, which makes the archive well suited to isolating whether an algorithm actually detects the anomaly pattern rather than exploiting a labeling shortcut. Review the license terms on the download page before use. Get it from cs.ucr.edu/~eamonn/time_series_data_2018.
IT operations and infrastructure monitoring
Server Machine Dataset (SMD)
SMD is a five-week-long multivariate dataset collected from a large internet company's production servers, covering 28 machines across three groups with 38 monitored metrics per machine, split into per-machine training and testing sets with roughly a 4% anomaly rate in the test data. It is one of the standard reference datasets for multivariate IT-operations anomaly detection research, alongside methods like OmniAnomaly that were built and evaluated on it. It is released under the MIT license. Get it from github.com/NetManAIOps/OmniAnomaly (the SMD data ships inside this repository).
Aerospace telemetry
NASA SMAP and MSL
This dataset provides real, anonymized spacecraft telemetry and expert-labeled anomalies from two NASA missions: the Soil Moisture Active Passive satellite (SMAP) and the Curiosity rover on Mars (MSL), totaling 105 labeled anomaly sequences across 82 telemetry channels. Because the anomalies were identified and labeled by domain experts rather than synthetically injected, it is a useful benchmark for evaluating detection methods against real operational judgment calls, though it has documented limitations including a small number of anomalies per channel. The Telemanom framework built around it is Apache 2.0 licensed; the underlying telemetry data is distributed via Kaggle. Get it from github.com/khundman/telemanom.
Industrial control systems
SWaT and WADI
SWaT (Secure Water Treatment) and WADI (Water Distribution) are testbed datasets from Singapore University of Technology and Design's iTrust center, capturing sensor and actuator readings from physical water-treatment and water-distribution testbeds under both normal operation and staged cyber-physical attacks. They are the standard reference datasets for anomaly and intrusion detection research on industrial control systems and critical infrastructure. Access is free but requires submitting a request and agreeing to iTrust's data-use terms, typically restricted to non-commercial research with an academic or institutional affiliation; approval is manual. Request access via sutd.edu.sg/itrust.
Datasets side by side
| Dataset | Domain | Scale | Access |
|---|---|---|---|
| NAB | General / streaming | 58 labeled series | Free, MIT license |
| UCR anomaly archive | General / algorithm testing | Many single-anomaly series | Free, check page for terms |
| Server Machine Dataset (SMD) | IT operations | 28 machines, 38 metrics each, 5 weeks | Free, MIT license |
| NASA SMAP / MSL | Aerospace telemetry | 105 anomaly sequences, 82 channels | Free via Kaggle; code Apache 2.0 |
| SWaT / WADI | Industrial control systems | Multi-day sensor/actuator logs | Free, request + data-use agreement |
How to choose
Start from your deployment domain rather than picking the biggest or most popular dataset. If you are building general infrastructure monitoring, SMD is closer to a real production signal than NAB's mixed synthetic series. If you work in aerospace, defense or scientific instrumentation, NASA SMAP/MSL gives you expert-labeled anomalies from an operational system. If your target is industrial control or critical infrastructure security, budget time for the SWaT/WADI approval process early, since it is not instant. Whatever you choose, read the dataset's own documentation on labeling limitations — several widely cited benchmarks, including early NAB and SMAP/MSL evaluation protocols, have been criticized for evaluation metrics that make scores look better than real-world performance, so validate any published number against the raw, non-adjusted evaluation before trusting it.
FAQ
Which time-series anomaly detection dataset should I start with?
It depends on your domain. For general streaming-anomaly research, start with NAB. For IT operations and server monitoring, use SMD. For aerospace or spacecraft telemetry, use NASA SMAP/MSL. For industrial control systems, use SWaT or WADI, which require a data-use agreement. The UCR anomaly archive is useful when you want many small, precisely labeled single-anomaly series to stress-test an algorithm rather than a single large production-like dataset.
Why do many anomaly detection benchmarks report near-perfect scores that don't hold up in practice?
Several widely used benchmarks have documented issues such as very few anomalies per series, mislabeled or ambiguous anomaly windows, or evaluation metrics like point-adjustment that inflate F1 scores. The UCR anomaly archive was created partly to address this by providing single, well-isolated, precisely labeled anomalies per series. When comparing benchmark numbers, check whether the paper uses point-adjusted or raw evaluation and read the dataset's own documented critiques before trusting a leaderboard score.
Are SWaT and WADI free to download?
They are free but not open — access requires submitting a request through iTrust at the Singapore University of Technology and Design and agreeing to their data-use terms, typically restricted to non-commercial research with an academic or institutional affiliation. Approval is manual and can take time, so request access early if your project depends on these datasets.
Still cannot find the right anomaly data? Leave it to us
If your domain is not covered by these benchmarks — finance, manufacturing sensors, network traffic in a specific industry — tell us your task goal and the conditions it must meet. DeepSData can run a real availability search and verify sources for you, checking reputable data platforms and judging matches and gaps against each requirement, with every source link kept traceable.
New accounts get 3 free searches; every result comes with verifiable source links. See also curated datasets.