sankalpsthakur/forge-industrial-control-scenarios

Forge Industrial Control and Telemetry Traces

Deterministic synthetic traces spanning device ingress, signature/quality/range failures, offline store-and-forward, local inference, sequential agent review, L0–L4 policy outcomes, electrolyser ramp sequences and multivariate telemetry anomalies.

No operational plant data, customer data, secrets or real equipment identifiers are included. machine.press-03 and every measurement are fictitious.

Files

• —edge_action_traces.jsonl: edge-to-agent outcomes.
• —safe_action_traces.jsonl: focused tool-capability cases.
• —control_scenario_traces.jsonl: approved/unapproved ramps and deterministic trips.
• —train.csv: 80 normal telemetry episodes from training-only seeds.
• —validation.csv: 40 separate normal episodes used for threshold calibration.
• —test.csv: 40 normal and 20 anomalous episodes from disjoint evaluation seeds.
• —export_electrolyser_dataset.py: deterministic CSV exporter.
• —release_manifest.json: row/run counts and SHA-256 digests.
• —SHA256SUMS: checksums for the three published CSV splits.

Intended use

Use the JSONL traces to test schemas, routers and safety-policy regressions. Use the telemetry splits to reproduce the public robust-PCA baseline and evaluate event recall, false-alert episodes and detection delay. Complete episodes—not overlapping windows—are assigned to splits, preventing temporal leakage.

The telemetry is not large or representative enough to train or validate a production anomaly detector. It contains normalized signals rather than engineering units, capacities, recipes or site protection settings.

Safety

Model detections remain L0 evidence. Agents may create L3 proposals but cannot approve them. L4 motion and energy commands are excluded from the model and agent surfaces.

Reproduce

# Standalone Hub dataset repository:
python export_electrolyser_dataset.py --output-dir exported-data

# From the full Forge lab source instead:
# python data/export_electrolyser_dataset.py --output-dir exported-data

The exporter writes 28,800 training rows, 14,400 calibration rows and 21,600 test rows at a fictitious one-second cadence. SHA-256 checksums are printed at the end of the run and written into the release manifest. Exact bytes are tied to the published generator and recorded numerical environment; split isolation and schema/row counts are the portable validation contract.