RailMind RailMind

Technology

Tracking the development of RailMind's online adaptive inference architecture from early research to deployable implementation.

Development Timeline

MilestoneFocusStatus
FoundationCore architecture and dynamicsCompleted
RepresentationInternal structure formationCompleted
ValidationSubsystem stress testingCompleted
CompetitionMulti-unit interaction dynamicsCompleted
EdgeHardware-optimized deploymentIn progress

Technical Highlights

  • Local online learning without end-to-end backpropagation — no loss functions, no optimizer
  • Hundreds of computational units competing in real-time under finite resource constraints
  • No pre-training required — no labeled datasets, no offline preparation
  • 35+ public datasets across 19 domains — each domain modelled and validated independently, no transfer assumed
  • Over 1.8 million controlled experiment runs across multiple research lines

Multi-Channel Architecture

RailMind outputs through three parallel channels — each serving a distinct integration use case:

ChannelOutputUse Case
Ch.1 Health VectorAdaptive internal representation projected to compact diagnostic spaceFault detection and classification
Ch.2 Drift GateZero-compute distribution shift detectorAlerts when the input distribution shifts
Ch.3 Raw StateFull internal representationCustom downstream tasks and research

Validation Breadth

Validated on 35+ public datasets across 19 domains. Each domain is modelled and validated independently — we do not assume transfer between domains.

DomainNamed public datasets
Rotating machinery — bearings, gearsCWRU · Paderborn/KAT · XJTU-SY · FEMTO/PHM2012 · SEU · IMS · MFPT · HUST · KAIST · UORED · MCC5 · CITEF
Turbofan remaining-lifeC-MAPSS (FD001–004)
Machine audio anomalyDCASE (ToyCar / Fan)

Under validation

DomainNamed public datasets
Bridge structural healthZ24 · KW51
Satellite telemetryESA-ADB SMAP · MSL
VideoCLIP-video · AI City · seg19
Human activity recognitionUCI HAR

Edge Performance

Production-verified performance on commodity hardware:

Runtime memory

Full configuration runtime memory: approximately 200 KB, including all PdM modules, with zero dynamic allocation.

MCU latency

L-tier configuration (not the full configuration): mean engine-step latency of 0.90 ms on an STM32F767 (Cortex-M7, 216 MHz).

Note: the memory and latency figures above correspond to different configurations and do not constitute a latency–memory pairing for a single configuration.

Zero heap allocation in the hot path — all static buffers. No garbage collection pauses, no fragmentation over 24/7 industrial deployment.

Drift Detection

Built-in distribution shift detection with zero additional computation:

  • Detects distribution drift and sensor degradation as it happens
  • Operating-condition changes are handled via explicit condition input (ConditionInput) — not inferred from the signal