Skip to content

NASA SMAP / MSL

What is this data? Real spacecraft telemetry from two NASA missions — the SMAP satellite (soil moisture) and the MSL Curiosity rover. Each channel pairs one telemetry value with 24 one-hot command flags; anomalies are incidents from NASA's ISA reports. The canonical aerospace telemetry benchmark.

  • Source: NASA telemanom (Hundman et al., KDD 2018) — https://github.com/khundman/telemanom
  • Contents: 55 SMAP channels and 27 MSL channels; per-channel train/test npy arrays (25 dims: 1 telemetry value + 24 command features)
  • License: NASA open data. The telemanom code itself is not used (only the label CSV).
  • Download: tsad-forge download smap — S3 (telemanom/data.zip) + label CSV. If S3 is blocked in your environment, download the HuggingFace mirror appleparan/telemanom manually and place files under data/smap_msl/train/<chan>.npy and data/smap_msl/test/<chan>.npy.
  • Loader: load_smap(channel="P-1"), load_msl(channel="M-6")

Known flaws (stated openly)

  • Quasi-binary channels: 24 of the 25 dimensions are one-hot-like command features. Almost all information lives in the single telemetry dimension — the benefit of multivariate models can be overestimated here.
  • Sparse labels: 1–3 anomaly events per channel, so event-level metrics have high variance.
  • History of PA inflation: most prior state-of-the-art claims on this dataset were PA-F1 based and are not comparable with the VUS-PR results in this repository (Kim et al., AAAI 2022).