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

Researchers introduce FactoryNet, the first universal pretraining dataset for industrial time-series data containing 51M datapoints across 23k task executions in robotic and machining domains. The dataset employs a novel S-E-F-C schema enabling cross-embodiment transfer and efficient anomaly detection, advancing toward industrial foundation models.

FactoryNet represents a significant step toward democratizing industrial AI by creating the first standardized dataset for pretraining foundation models on manufacturing systems. The 51M datapoint corpus—combining 13.3k real and 9.8k synthetic executions across six different embodiments—addresses a critical gap in AI infrastructure where industrial automation has lagged behind vision and language model development. The Setpoint-Effort-Feedback-Context (S-E-F-C) schema innovatively abstracts any actuated system into a unified representational framework, enabling models trained on one robot or machine to transfer knowledge to fundamentally different hardware configurations.

Industrially, this development matters because factories operate heterogeneous equipment ecosystems where rigid, task-specific models prove economically inefficient. The dataset's inclusion of 27 annotated anomaly types alongside counterfactual pairs directly addresses manufacturing's most critical pain point: unplanned downtime detection. Cross-embodiment transfer capabilities unlock scenarios where a model trained primarily on robotic systems can meaningfully contribute to machinery diagnostics without extensive retraining.

For the broader AI sector, FactoryNet validates that foundation model scaling extends beyond consumer applications into infrastructure domains worth trillions annually. The parameter-efficient anomaly detection aligns with growing enterprise demand for edge-deployable models, reducing computational overhead in bandwidth-constrained factory environments. The release of this growing dataset signals increasing collaboration between academic researchers and industrial stakeholders, establishing shared benchmarks that accelerate development cycles.

Market trajectory depends on whether this catalyzes similar industrial datasets across domains like power grids, water systems, and chemical processes—each representing multi-billion-dollar optimization opportunities.

• →FactoryNet contains 51M datapoints across diverse manufacturing systems with a unified schema enabling cross-hardware transfer learning.
• →The novel S-E-F-C framework abstracts actuated systems into common representations, addressing heterogeneous factory equipment challenges.
• →Cross-embodiment transfer experiments demonstrate competitive anomaly detection using only 24 schema-aligned signals versus high-dimensional baselines.
• →The dataset combines real and synthetic data with 27 annotated anomaly types, directly targeting manufacturing's downtime prediction problem.
• →Release as a growing, multi-embodiment corpus establishes shared infrastructure for industrial foundation model development.