Chemical Process Anomaly Detection Dataset: Where the Labeled Plant Data Lives

The archive measures ~295 MB compressed - 309,354,827 bytes per the public Kaggle API - and unpacks into multiple directories covering two families of case studies: batch distillation runs and continuous reactive processes. Each run arrives as a high-frequency multivariate time series with thousands of signal columns from plant sensors and actuators, and each carries labels marking when faults occurred. Industrial provenance, per-run labels and wide sensor coverage together are what separate it from generic sensor corpora.

Treat that fourth step as exploratory analysis rather than setup overhead, and pin the version you downloaded beside your experiment logs.

The labels are the asset. Because each run ships with intervals marking when faults occurred, you score a detector against ground truth instead of eyeballing reconstructed residuals - the difference between reporting a precision figure and reporting a feeling. Three task families map straight onto the structure: point-wise anomaly classification over the multivariate series, segment-level detection of the labeled intervals, and early-warning models scored on lead time bought before a labeled onset.

The two case-study families reward separate treatment. Batch distillation progresses through phases, so a deviation that is anomalous in one phase is routine in another; continuous reactive processes call instead for steady-state residual monitoring across thousands of parallel signals. One model per family, then compared, mirrors how the benchmark itself is organized.

Keep provenance in mind when generalizing. This is telemetry from BASF SE plants - one operator's sensor layout and control philosophy. It transfers well to benchmark comparisons and method papers, more cautiously to other companies' plants. It also explains why four personas in Datadory's tagging carry this record: data scientists and journalists, academics and students at relevance 3, competitive-intel product teams and developers at relevance 2.

Yes - which makes it the exception in this slice's machine-learning cluster. Kaggle lists NoBOOM under Attribution 4.0 International (commercial delivery terms 4.0): sharing and adaptation are permitted, commercial use included, provided you attribute. One caveat belongs in your README: the underlying data originates from BASF SE, so check any additional restrictions stated on the dataset card before shipping a product.

NoBOOM is the only record in the pool operating at unit-operation scale. Its three neighbors in the ML sub-cluster work at molecular or computational scale, so treat them as complementary layers rather than substitutes - process telemetry for fault detection, compound corpora for cheminformatics, DFT calculations for catalyst potentials.

The table puts all four side by side so each can be matched to the layer of the problem being modeled: NoBOOM for the plant floor, the BASF-AI collection and PubChem CSV for molecular representation work, AQCat25 for catalyst potentials where spin polarization matters.

A fault-detection model earns its keep against everything else known about the sites it watches, and the same catalog supplies that context free of charge.

• Environmental ground truth. The EPA Toxics Release Inventory (TRI) Basic Data Files publish facility-level toxic release, transfer and waste-management CSVs with 100 fields per row and roughly 90k+ records per recent national reporting year, running from reporting year 1987 through 2024 in the commercial delivery terms. Aligning detected excursions against a facility's reported history gives alerts an auditable backdrop.
• Site and volume framing. EPA's Chemical Data Reporting rule covers roughly 8,650 chemicals in its 2024 cycle, spanning reporting years 2020-2023 per substance and per manufacturing or import site - useful for weighting which plants and chemistries justify monitoring spend.
• Feedstock context. The US Census Bureau International Trade Data API returns commercial delivery terms monthly import and export values, quantities and shipping weights by HS2-HS10 code across 18 datasets, needing only a free key - a cheap way to line upstream supply shocks up against the periods your detector flags.

For contrast, the synthetic option lives one industry over: commodity-chemicals hosts a simulated reactor set at minute resolution, useful for volume but without plant-grade provenance. Data scientists assembling the full stack will find these sources ranked together on our diversified-chemicals page for data scientists.