sankalpsthakur/forge-tiny-drift-multiruntime

Forge Tiny Drift

An intentionally small, interpretable model for demonstrating cross-runtime industrial edge deployment. It reads 50 cycles with two channels—peak force in kN and part deviation in mm—and returns an L0 drift probability.

This model is decision support only. It has no action or equipment-control output.

Architecture

Six deterministic features are extracted from the window: force slope, early-to-late force shift, force standard deviation, maximum deviation, last deviation and force range. A trained logistic head operates on fixed normalized features.

Artifact	Intended runtime
tiny_drift_pytorch.pt	PyTorch training/reference
tiny_drift.onnx	ONNX Runtime on x86/ARM IPCs
tensorflow_saved_model/	TensorFlow parity/reference
tiny_drift.tflite	LiteRT mobile/embedded runtime
weights.json	NumPy fallback and inspectable weights

Data and metrics

Training uses 2,048 deterministic synthetic windows generated by models/export_models.py with seed 17. No customer or plant telemetry is included. Metrics and exact versions are recorded in model_manifest.json; runtime errors and classification agreement are recorded in conformance.json.

The publication gate requires:

• maximum probability error ≤ 1e-4 against the NumPy reference;
• 100% threshold-decision agreement across PyTorch, ONNX Runtime, TensorFlow and LiteRT;
• deterministic training inputs and seeds, with the numerical stack recorded in model_manifest.json.

Floating-point artifacts are not claimed to be byte-identical across different NumPy, BLAS, operating-system or framework builds. The published artifacts and SHA256SUMS are the release reference; reproduce parity and decision agreement, not an unqualified cross-platform binary hash.

Reproduce

pip install -e '.[model-export]'
# Standalone Hub model repository:
python export_models.py
python run_conformance.py

# From the full Forge lab source instead:
# python models/export_models.py
# python models/run_conformance.py

Limitations

The synthetic generator is not representative of a particular machine, material, sensor, failure distribution or operating envelope. Thresholds and calibration must be validated for each real deployment. Field quality, replay protection and engineering ranges are enforced by the edge gateway, while plant safety remains the responsibility of trusted PLC/RTU logic, hardware interlocks and authenticated operator workflows.

For the separate multivariate electrolyser/auxiliary demonstration, see the telemetry anomaly baseline card. That baseline adds whole-run splitting, event recall, false-alert rate and detection-delay metrics; it does not replace this model's cross-runtime export evidence.