RailMind RailMind

Technology

Tracking the development of RailMind's dissipative neural architecture from inception to deployment.

Development Timeline

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

Technical Highlights

  • Gradient-free online learning — no 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
  • 0.1ms per-step deterministic inference on Raspberry Pi 5
  • AUC 0.996 on CWRU bearing fault detection
  • 10 benchmark datasets spanning 6 signal domains — each validated independently, same architecture throughout
  • ~40KB model footprint — deployable on MCU-class hardware (varies by application)
  • 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

Industrial Validation

Validated independently across six distinct signal domains — same architecture, per-domain deployment:

DomainDatasetKey MetricDomain Type
Vibration (PdM)CWRU BearingAUC 0.996Industrial fault
Vibration (PdM)PHM 2022 Rock DrillF1 0.91Industrial fault
Vibration (PdM)CASPER UR3e RobotMulti-axis regime detection (6-axis, 1.76M rows)Robotics PdM
Vibration (SHM)Z24 BridgeAUC 0.99 (within-stream temporal protocol)Structural health
ElectrochemistryNASA BatteryAUC 0.958 (+14.7 pp)Degradation
Human MotionUCI HAR / WISDM v2Multi-regime detectionActivity recognition
AudioESC-50 / DCASEPositive lift (+0.9 pp)Acoustic scene
Satellite TelemetryESA-ADB SMAPCEF0.5 = 0.911Aerospace
Satellite Telemetry (temporal)ESA-ADB SMAPAUC 0.979Aerospace

The engine’s internal representation consistently outperforms scalar readouts by 13–49 percentage points across tested domains. Catastrophic forgetting after sequential multi-fault exposure: 2.3% (CWRU 3-fault continual protocol).

Aerospace Differentiator

Satellite telemetry anomaly detection on the ESA-ADB SMAP benchmark:

  • CEF0.5 = 0.911 [0.897, 0.923] — exceeds published reference of 0.888
  • Zero dynamic memory allocation — satellite-grade reliability
  • Static computational budget: O(1) per sample, deterministic execution path

Note: Non-apple-to-apple comparison with different train/test splits; confidence interval provided.

Edge Performance

Production-verified performance on commodity hardware:

MetricValue
RPi5 median latency0.1 ms (103 μs)
RPi5 P95 latency217 μs
Mac M4 median latency53 μs
Throughput (RPi5)8,102 steps/s
Total memory112.3 KB

Memory breakdown:

ComponentSize
Engine state (static)61.6 KB
PCA projection matrix10.3 KB
RF classifier model40.4 KB

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:

  • True Detection Rate: 84%
  • False Alarm Rate: 1.4%
  • Detects distribution drift and sensor degradation as it happens
  • Operating-condition changes are handled via explicit condition input (ConditionInput) — not inferred from the signal