PRAETOR

PREATOR a physics-grounded digital twin for factory predictive maintenance PREATOR is a live digital twin of a factory floor that predicts machine failures without adding any new sensors using only the sensors already present, combined with physics. The system has three layers, modeled on a real factory's chain of command. At the base, each machine is represented by a Multi-Head Physics-Informed Neural Network (MH-PINN) that embeds governing equations from bearing kinematics (BPFO/BPFI/BSF), cure chemistry (Arrhenius, Kamal-Sourour), and fatigue mechanics (Hertzian contact, Basquin, Miner) directly into the model. Because the physics constrains the solution space, the model can reconstruct physical quantities at points that were never instrumented — we demonstrate this by withholding a real sensor entirely and reconstructing its signal with 13% accuracy on real bearing data, consistent with published PINN research. The middle layer is a line-level LLM that reads each machine's physical state, reasons over it, and explains it to a non-technical operator in plain language, always citing the actual data it consulted. The top layer is a central LLM that aggregates every machine's state through a shared store to hold the plant-wide picture. The operator stays in control: they can question any recommendation and override it in the moment. Every override feeds a Champion/Challenger learning loop — a live model serves predictions while a challenger retrains in the background on operator corrections, then takes over when it outperforms — so both the reasoning layer and the physical model sharpen with every human decision. The use case is tire manufacturing, focused on the curing press, using public datasets (NASA C-MAPSS, CWRU and NASA IMS bearings, AI4I 2020) as physically-plausible stand-ins no confidential manufacturer data is used. The LLM layers run on Nvidia's Nemotron Ultra, served through Crusoe Managed Inference for real-time latency across all three layers.