Fleet Twins and Transfer Learning
When many similar machines run, their twins share learned patterns and priors while each keeps its own state and history.
Many machines, shared learning
A single machine's twin learns only from that machine. When a design is built as a fleet, the twins can learn from each other: a behavior, fault precursor, or calibration insight discovered on one unit can inform the others. Fleet twins balance two needs, each unit keeps its own individual twin, tracking its own wear and history, while the fleet accumulates shared knowledge that improves every unit.
What is shared and what is not
- Shared: model structure, learned failure precursors, prior distributions for parameters
- Shared: surrogates trained on pooled data across units
- Individual: the current state estimate of each unit
- Individual: each unit's calibration, wear ledger, and operating history
Transfer learning
Transfer learning uses knowledge gained on one system to accelerate learning on another. A new unit need not start from ignorance; it inherits informed priors from the fleet and refines them with its own data. This is especially valuable early in a unit's life, when it has little history of its own, and for rare events that no single unit sees often but the fleet does.
The pitfalls
Units are similar but not identical, differing in construction, siting, and duty. Naively pooling their data can wash out real individual differences or transfer a pattern that does not apply. Sound fleet learning models the units as related but distinct, so shared knowledge informs each unit's twin without overriding what that unit's own data says.
In the Kronos program
The burner is planned as fleets, deployed in Aegis housings for fixed defense installations and MetroVolt housings for data centers, making fleet learning directly applicable: the same machine in many units, ideal for shared precursors and transferred priors. Successive breeder units likewise benefit as they come online. Fleet learning is designed into the twin architecture and data standards now, and becomes active once multiple units operate, after first tritium near 2030. See data standards and predictive maintenance.