PdMData: Datasets for Predictive Maintenance

A Python toolkit to download, load, and visualize predictive-maintenance datasets for research. Use a shared API with dataset-specific loaders; download only what you need.

Prerequisites

• uv on your PATH
• Python 3.14 (installed automatically by uv from .python-version)
• Optional: a system libarchive library if you will download IMS (nested ZIP/7z/RAR). macOS usually includes one; Debian/Ubuntu: sudo apt-get install libarchive-dev; see Usage for Windows notes.

Windows users can run the same uv commands in PowerShell, or use Git Bash / WSL for ./scripts/bootstrap.sh.

Getting started

From the repository root:

./scripts/bootstrap.sh

Or, without the shell script:

uv sync --locked
uv run --locked python -m ipykernel install --user \
  --name pdmdata --display-name "Python (pdmdata)"

Confirm the install:

uv run --locked python -c "import pdmdata; pdmdata.summary()"

Setup creates the locked Python 3.14 environment (runtime, Jupyter, and dev tooling) and registers the Python (pdmdata) notebook kernel. It does not download datasets. Data lands under data/raw/ by default (see pdmdata.toml).

Start JupyterLab, select the Python (pdmdata) kernel, and run:

uv run --locked jupyter lab

import pdmdata

pdmdata.summary()
pdmdata.download("cmapss")  # direct URL; no Kaggle login
frame = pdmdata.load("cmapss", subset="FD001", split="train")
frame.head()

Importing or loading never downloads data. Download each dataset explicitly first. Loaders return Polars DataFrames or LazyFrames, depending on the dataset.

figure = pdmdata.visualize("cmapss", "sensor_trajectory", frame, entity=1)
figure.show()

Next steps

Goal	Where to go
Choose a dataset and compare support	Dataset catalog
Browse the task list and experiment views	Task overview
Configure paths, downloads, and Kaggle auth	Usage guide · Documentation index
Explore data or task examples	Notebooks
Run CLI and maintenance scripts	Scripts guide
Run tests	uv run --locked python -m unittest discover -s tests -q

Open a dataset README from the catalog for selectors, schema details, and terms of use. Open a task README for experiment bundles and scoring protocols.

Acknowledgments

Thanks to the researchers, institutions, and data providers who create and share the predictive-maintenance datasets that this project downloads and loads. PdMData exists to make those sources easier to use; credit and citation belong with the original authors. Provider names, licenses, and suggested citations are listed in each dataset README and in the dataset catalog.

License and data use

Repository code and documentation are available under the MIT License.

This repository does not redistribute dataset files. Downloads fetch data from the original providers (or their documented mirrors) into your local data directory. Each dataset remains under its own license and terms of use. Before downloading or using any dataset, review the provider’s rights, citation requirements, and restrictions, and proceed at your own responsibility. See the dataset catalog and the per-dataset READMEs for source links and terms.