Benchmark

The goal of HAI-CPS is to provide a comprehensive benchmark that enables you to evaluate your machine learning algorithms along two axes:

  • CPS complexity

  • Anomaly complexity

HAI-CPS includes ten different datasets, each recorded from a unique process plant setup.
The datasets are available in discrete, continuous, and hybrid formats to suit a wide range of algorithmic needs.

You can find a full overview of all datasets here.

CPS Complexity

HAI-CPS features ten different process plant configurations, each exhibiting increasing structural and behavioral complexity.

HAI-CPS setups

Anomaly Complexity

For each individual anomaly — and all meaningful combinations of anomalies — a dedicated dataset has been recorded.
This allows for a systematic evaluation of your model’s performance under varying anomaly scenarios and interactions.

Data Formats

HAI-CPS datasets are provided in four formats, depending on what your model requires:

  • discrete: Includes only binary/discrete control and sensor values.

  • continuous: Includes only real-valued sensor data and process measurements.

  • hybrid: A combination of discrete and continuous features.

  • including-states: Hybrid data plus automaton states from each module (i.e., internal process states during execution).