Benchmark

HAI-CPPS evaluates machine-learning methods along two axes: plant complexity and fault complexity. The ten topologies use different combinations of mixer, filter, distill, and bottling modules with source and sink modules.

Campaign design

A complete v2.2 campaign has 350 recordings: 160 healthy and 190 single-fault. Each topology supplies 10 healthy training, 3 validation, 2 calibration, and 1 test recording. Every selected fault proxy produces a persistent and a temporary recording, both paired with the seed-42 healthy test control. Fault activation begins at time 2,500; the temporary variant ends at 7,500. The campaign output guide gives per-topology counts and scenario names.

Measurement and diagnosis inputs

Each recording offers continuous, discrete, and hybrid measurement views. Nominal controller commands are saved separately and can be added if they are observable in the intended application. permitted_inputs.json defines the allowlists. Simulator oracle states, injected-fault labels, technical timing, and verification traces remain offline artifacts.

The campaign validates raw results, exported tables, channel alignment, fault pairs, and release structure. Each recording has a validation.json and provenance.json; each topology shard has a campaign.json manifest and a frozen copy of generation sources.