PATH: A Discrete-sequence Dataset for Evaluating Online Unsupervised Anomaly Detection Approaches for Multivariate Time Series

This record comprises the 0_simulation and 1_postsim folders. The 0_simulation folder consists of the raw .mat files, each representing a multivariate time series, whereas the 1_postsim folder consists of the following pickle files:

• normal.pkl, which contains all nominal sequences
• anomalous.pkl, which contains all anomalous sequences
• control.pkl, which contains all control-counterparts to anomalous.pkl
• training.pkl, which contains all pre-determined folds for training
• training_clean.pkl, a version of training.pkl without anomalous sequences
• testing.pkl, which contains all pre-determined folds for testing
• testing_clean.pkl, a version of testing.pkl without anomalous sequences

Each pickle file is a list of 2D NumPy arrays, each representing a multivariate time series. The name of the corresponding .mat file (and, by extension, the label) is present in the metadata. For NumPy object array, it can be read by calling array.dtype.metadata['file_name'].

We decided to omit the 2_preprocessed folder as the contents are specific to the TensorFlow data pipeline and the same data host limitations would apply.

For more information, refer to the publication.

For access to the source code, refer to the repository on GitHub.

If you use this dataset for your research, please consider citing it through the menu on the right, or using the following bibtex entry:
@misc{correiaPATHDiscretesequenceDataset2025,
  title = {{{PATH}}: {{A Discrete-sequence Dataset}} for {{Evaluating Online Unsupervised Anomaly Detection Approaches}} for {{Multivariate Time Series}}},
  author = {Correia, Lucas and Goos, Jan-Christoph and B{\"a}ck, Thomas and Kononova, Anna V.},
  year = {2025},
  publisher = {Zenodo},
  doi = {10.5281/ZENODO.13255120},
  copyright = {MIT License}
}