Datasets

HAI-CPPS v2.2 provides ten modular process-plant topologies with increasing complexity. A complete generated campaign has 350 recordings: 160 healthy and 190 single-fault recordings. Each topology has training, validation, calibration, and healthy test recordings. Every enabled fault proxy has persistent and temporary test recordings paired with a seed-matched healthy control.

Each benchmark recording has three Parquet measurement views: continuous process values, discrete sensor states, and hybrid combined measurements. A separate commands.parquet table contains directly recorded nominal controller commands. The permitted_inputs.json policy lists which measurement and command columns can be used for diagnosis.

Simulator-derived oracle_states.parquet, fault labels, structural knowledge, timing, and audit data are separate offline artifacts. Do not join them into online model inputs. The three identifier columns are for alignment and are excluded as diagnostic features.

The release is organized in topology shards, such as output/v2.2/ds1/ds1/ds1_016_healthy/. Every recording spans times 0–25,000 with 25,001 rows. See Campaign Output and Data Directory for the complete tree, file-by-file roles, naming scheme, fault windows, and an input-loading example.

Download

HAI-CPPS datasets are hosted on IEEE DataPort. You can find the published benchmark there. Check the publication’s version before comparing it with a locally generated v2.2 campaign.