FactoryNet

Overview

Manufacturing accounts for roughly 15% of global GDP, yet industrial AI remains confined to single-machine, bespoke deployments. Foundation models have transformed vision and language by pretraining on large, structurally coherent corpora, but no analogous substrate exists for industrial time-series. The gap is not merely volume: existing anomaly detection and forecasting datasets record sensor outcomes without separating commanded intent from measured response. For actuated systems, learning transferable dynamics requires observing the full control loop, from target trajectory through actuation effort to the resulting physical state.

FactoryNet introduces the first universal pretraining corpus for industrial time-series. It unifies novel laboratory recordings on UR3 and KUKA KR10 industrial arms with standardized adaptations of voraus-AD, AURSAD, and UMich CNC, alongside a parallel synthetic track from NVIDIA Isaac Sim. Every signal is mapped into the Setpoint–Effort–Feedback–Context (S-E-F-C) schema, a control-theoretic decomposition grounded in IEC 81346 that enables cross-embodiment learning across arbitrary actuated systems.

The S-E-F-C taxonomy

Most existing benchmarks log raw sensor streams, tangling the controller's target variables with the machine's actual physical execution. FactoryNet maps over 300 heterogeneous signals into four standardized roles. The taxonomy lets a single dataloader work across a 6-DOF rotational arm and a 4-axis CNC gantry: they expose the same four roles, only with different axis counts and units.

Corpus composition

FactoryNet v1.0 spans three pillars: real-world laboratory recordings, standardized open-source adaptations, and a synthetic generation pipeline. Of the 9,114 lab episodes, approximately 40% are healthy and 60% contain injected faults across 27 anomaly types spanning Pick & Place, Screwdriving, and Peg-in-Hole.

Synthetic track & sim-to-real

The synthetic pipeline procedurally generates UR5 Pick & Place episodes in NVIDIA Isaac Sim with domain randomization across payload mass (0.10–0.30 kg), surface friction (0.30–0.50), controller gains, sensor noise, and task geometry. Each episode comes with aligned S-E-F-C metadata and matched healthy twins for controlled fault-deviation analysis. Batch sim-to-real validation across 1,155 paired episodes reports a median joint RMSE of 2.83° and a TCP position RMSE of 13.26 mm, with residual TCP rotation spread attributed to gripper-geometry mismatch between the Isaac (Robotiq 2FG85) and lab (OnRobot 2FG14) setups.

Headline results

Anomaly detection on voraus-AD

Trained on healthy episodes only, an S-E-F-C MLP regresses motor torque from 18 setpoint signals and uses per-episode prediction error as the anomaly score. On just 24 schema-aligned signals, it matches or beats every full-channel (130-signal) baseline except the strongest unstructured methods, with notable wins on mechanically distinctive faults (miscommutation 99.2, additional axis weight 95.8).

Zero-shot cross-embodiment transfer

A 105k-parameter TCN-Transformer trained solely on voraus-AD (Yu-Cobot) Pick & Place is evaluated on 1,433 AURSAD UR3e Screwdriving episodes: a different robot doing a different task. Under the mean-centered MAE metric, which isolates dynamic forces from static payload biases, the structured S-E-F-C model is the only learned baseline that beats a strong non-learned kinematic prior.

Getting started

Citation

@inproceedings{factorynet2026,
  title  = {FactoryNet: A Large-Scale Dataset toward Industrial
            Time-Series Foundation Models},
  author = {Othman, Karim and Petersen, Jonas and
            Ignuta-Ciuncanu, Matei and Maggioni, Riccardo and
            Mazzoleni, Camilla and Martelli, Federico and
            Petersen, Philipp},
  year   = {2026}
}