VaidhyaMegha/assetops-kg

AssetOps Knowledge Graph

12,647 nodes. 12,662 edges. An industrial asset-operations property graph built from IBM AssetOpsBench — chiller equipment, their sensors, the maintenance work orders raised against them, and the anomaly and alert events they emit.

Part of the Samyama ecosystem. The loader that rebuilds this graph from the upstream source lives at samyama-ai/assetops-kg.

Licence and attribution

The upstream data is IBM AssetOpsBench, released under Apache-2.0 — stated on the repository's LICENSE file and confirmed by IBM's own public dataset ibm-research/AssetOpsBench, which carries license: apache-2.0 and is ungated. Apache-2.0 permits redistribution and derivative works, and requires that the licence and attribution be preserved — which this card does.

Attribution: derived from IBM AssetOpsBench, © IBM, Apache-2.0. IBM does not endorse this dataset or any analysis derived from it.

The loader code in the GitHub repository is separately Apache-2.0. That is the code licence and does not govern the data; both happen to be Apache-2.0 here.

Structure

9 node labels

Label	Rows	What it is
Event	6,256	Operational events logged against equipment
WorkOrder	4,249	Maintenance work orders, with labour hours and problem codes
AlertEvent	1,466	Rule-triggered alerts, with start/end times
AnomalyEvent	541	Detected anomalies, with KPI, score and bounds
Sensor	110	Sensors attached to equipment
FailureMode	12	ISO 14224-style failure modes, with severity
Equipment	11	The chillers themselves
Site	1	MAIN
Location	1	Main Building

5 edge types — FOR_EQUIPMENT (12,512), HAS_SENSOR (110), MONITORS (28), CONTAINS_EQUIPMENT (11), CONTAINS_LOCATION (1).

Files

• nodes/*.csv — one file per label, previewable in the dataset viewer
• edges/*.csv — one file per edge type; columns are id, src, tgt referencing node id
• assetops.sgsnap — the native snapshot. The viewer cannot render it (it shows as an opaque binary); it is a gzipped NDJSON snapshot that loads into the Samyama graph engine in one step, without re-running the ETL.

Example query

Once loaded into the engine — the work orders raised against the busiest chiller:

MATCH (e:Equipment {name: 'Chiller 6'})<-[:FOR_EQUIPMENT]-(w:WorkOrder)
RETURN w.primary_code_description, w.actual_labor_hours
ORDER BY w.actual_labor_hours DESC
LIMIT 10

Verification

Counts were measured from the snapshot, not copied from a README, and match both the snapshot header and the source repository's stated totals exactly: 12,647 / 12,662.

• Orphans: 5 of 12,647 (0.04%) — all FailureMode. The graph contains 11 pieces of equipment, all chillers; 5 of the 12 failure modes describe AHU (air-handling unit) assets, which have no equipment node here to attach to. The other 7 chiller failure modes are connected via MONITORS.
• No personal data. No names, emails or addresses; every identifier is an equipment, work-order or event code.

Known limitations

• This is the base graph, loaded directly from IBM's sources — 9 labels and 5 edge types. The DEPENDS_ON failure-propagation topology and criticality_score shown in the GitHub repository's demo are an analytical layer added on top of IBM's data and are not included here. Neither are the LLM-derived (source: "LLM-derived") nodes from the Generation-Augmented Knowledge demo. Nothing in this dataset is model-generated.
• Edge distribution is heavily skewed — 12,512 of 12,662 edges are FOR_EQUIPMENT, because every event, alert and work order points at one of only 11 machines. This is inherent to the source, not an export defect.
• Only chillers are represented, despite AssetOpsBench covering more asset classes.

Links

Loader / ETL	github.com/samyama-ai/assetops-kg
Samyama Graph	github.com/samyama-ai/samyama-graph
Upstream source	github.com/IBM/AssetOpsBench
Contact	samyama.dev/contact