anonymous-reviewer-207/CAFA-GNN-Federated-Benchmarks

CAFA-GNN Federated Benchmarks

Three federated heterogeneous-graph datasets built to reproduce and benchmark the CAFA-GNN framework (Concept-Aware Federated Agentic GNN for Industrial Predictive Maintenance). Each federation ships as a single .zip containing one torch_geometric.data.HeteroData per client.

Federation	Source	Clients	Node types	Task	Per-node feature
cmapss	NASA C-MAPSS turbofan archive (PHM Society mirror)	3,000	sensor (14), engine (1), operator (6)	regression (RUL per engine, capped at 125)	(W=30, D=14)
batadal	BATADAL water-distribution training set	2,000	pump (11), tank (7), sensor (12)	binary classification (anomaly per pump)	(W=30, D=14)
d3_synthetic	procedurally generated, 3-regime physics model	2,000	sensor (60), machine (15), line (3)	regression (RUL per machine, capped at 125)	(W=30, D=14)

7,000 client files total. Five relation types per heterograph in each federation (see node-type pairs printed by print(data.edge_types) after loading).

Why a federated benchmark?

Predictive-maintenance research on CMAPSS and BATADAL has historically been done centrally — one model trained on the union of all engines / all timestamps. CAFA-GNN argues that real-world deployments are intrinsically federated (one factory per client) and that clusters of clients share similar degradation regimes. This benchmark exposes that structure explicitly: each .pt file is a self-contained per-factory graph snapshot with its own targets, ready to drop into Flower or any other FL simulation framework.

Files

cmapss_federated.zip       # 3000 × client_*.pt  (≈16 MB)
batadal_federated.zip      # 2000 × client_*.pt  (≈14 MB)
synthetic_d3_K100.zip      # 2000 × client_*.pt  (≈27 MB)

Loading

import zipfile, pathlib, torch
from torch_geometric.data import HeteroData

zipfile.ZipFile("cmapss_federated.zip").extractall("cmapss/")
data: HeteroData = torch.load("cmapss/client_0.pt")

print(data.node_types)      # ['sensor', 'engine', 'operator']
print(data.edge_types)      # 5 relation triples
print(data['engine'].x.shape)   # (1, 30, 14)
print(data.y.shape)             # (1,)  — RUL target
print(data.cmapss_subset, data.cmapss_engine_id)  # provenance

Each HeteroData carries:

• x on every node type, shape (N_nodes, W=30, D=14). Channel 0 is the normalised per-entity signal; channels 1–3 are the operating context (CMAPSS op-settings / BATADAL pressure context / D3 cluster progress); remaining channels are zero-padded so the schema is shape-uniform.
• edge_index for each of 5 relation types per federation.
• y — the regression target (RUL for cmapss / d3, ≤125) or binary tensor (per-pump anomaly for batadal).
• Provenance attributes: client_id, task, plus federation-specific metadata (cmapss_subset, cmapss_engine_id, cmapss_end_cycle, cmapss_dominant_regime, batadal_window_start, batadal_window_attacked, or cluster_id for D3).

Reproducibility

Every byte of these archives is produced by a single deterministic script: scripts/generate_hf_datasets.py in the parent CAFA-GNN repository, driven by a fixed master seed (default 42). CMAPSS and BATADAL features are derived from the real public datasets (NASA PHM Society S3 + batadal.net); D3 features come from a three-regime physics model (vibration-dominant, monotonic drift with end-of-life acceleration, stepwise faults). Re-running the script with the same seed yields bit-identical HeteroData payloads.

Upstream licensing

• CMAPSS — released by NASA Prognostics Center of Excellence; freely usable for research with attribution.
• BATADAL — released alongside the BATADAL challenge; freely usable for research with attribution to the BATADAL organisers.
• D3 Synthetic — generated by this project; CC-BY-4.0.

When citing this bundle, please also cite the upstream CMAPSS (Saxena et al. PHM '08) and BATADAL (Taormina et al. 2018) datasets where applicable.

Citation

@inproceedings{cafa-gnn,
  title  = {CAFA-GNN: A Concept-Aware Federated Agentic Framework for Industrial Predictive Maintenance},
  author = {...},
  year   = {2026}
}