Title: FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

Abstract:

We present FactoryNet, the inaugural universal pretraining corpus designed specifically for industrial time-series data. This comprehensive dataset comprises 51 million data points derived from 23,000 end-to-end task executions, split between 13,300 real-world scenarios and 9,800 synthetic simulations. These executions span six distinct embodiments, all unified under a shared schema that facilitates robust zero-shot cross-embodiment transfer and enables highly parameter-efficient anomaly detection.

Central to our pipeline is a novel S-E-F-C schema—comprising Setpoint, Effort, Feedback, and Context—which maps any actuated system into a standardized representational framework. The corpus includes 27 annotated anomaly types, supported by healthy baselines and counterfactual pairs, covering both robotic manipulation and machining domains.

Our experiments on cross-embodiment transfer demonstrate promising outcomes: under bias-aware metrics, the model exhibits fair transfer capabilities across the evaluated source-target pairs. Furthermore, utilizing 24 schema-aligned signals yields competitive anomaly detection performance relative to high-dimensional baselines. We are releasing FactoryNet as an expanding, multi-embodiment dataset to accelerate advancements in industrial foundation models.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC