Multimodal industrial pretraining with STL-regularized objectives and causal priors for machine-checkable rationales and controller-aware outputs.
Plans and controllers synthesized under temporal-logic contracts. Reference governors guarantee safety; residual RL optimizes within verified tubes.
Physics-guided features + causal graphs + STL abstractions to reconstruct propagation chains and recommend cost-aware interventions across fleets.
Click a section to expand. Each direction includes a lab-website–friendly introduction, theory-forward core idea, key tasks, representative pilots, metrics, and alignment with national strategy.
We build the next generation of manufacturing AI by unifying data-driven learning with logic, physics, and causality. By embedding Signal Temporal Logic (STL) constraints and causal structure into the representation space, our models produce machine-checkable rationales and calibrated uncertainty, enabling cross-factory transfer and controller-aware recommendations.
Pretrain on multimodal industrial data (time series, events, logs) with STL-regularized objectives, causal priors from knowledge graphs, and physics-consistent invariants. Dual formulations turn logic satisfaction into differentiable losses; distilled symbolic rules align human audit with closed-loop control.
Flexible job-shops Robotic assembly Rotating machinery Rationale fidelity Zero/low-shot transfer Causal precision/recall
National fit: 工业软件自主可控 / 新质生产力。
We engineer robot autonomy that is safe by construction and fast by design. Safety is specified up-front as temporal logic contracts and embedded throughout planning, learning, and control. The result is dual-arm and mobile manipulation that adapts to changing workcells while certifying task timing, force limits, and human–robot interaction rules.
A hierarchical stack where high-level plans and low-level controllers are synthesized and trained under STL specifications. Reference governors project actions onto contract-satisfying manifolds; residual RL optimizes performance inside verified tubes. Runtime monitors provide certificate-based intervention and graceful recovery.
Force-controlled assembly AGV–arm handoff Rapid changeover STL satisfaction Verified accuracy Near-miss reduction
National fit: 先进制造装备与智能机器人。
We move beyond alarms to actionable causality in Prognostics & Health Management. By coupling physics-guided signal processing with causal graphs and temporal logic abstraction, we reconstruct fault propagation chains across components, systems, and fleets—turning heterogeneous data into interventions that save uptime, energy, and cost.
Fuse order/transfer-path–aware features with causal discovery and do-calculus–informed reasoning, then elevate events to STL templates that encode mechanism-consistent signatures. Perform inference on a multi-layer PHM knowledge graph to generate interventions (derating, maintenance windows) with transparent justifications.
Wind & gas turbines Rail traction CNC & spindles Lead time↑ Forced outage↓ RUL precision↑
National fit: 能源安全 / 交通强国 / 高端装备可靠性工程。