| --- |
| language: |
| - en |
| license: apache-2.0 |
| tags: |
| - sofia |
| - numpy |
| - knowledge-distillation |
| - quantum-inspired |
| - verifiable-kernels |
| - synthetic-data |
| - vibration-analysis |
| - experimental |
| --- |
| |
| # Sofia Edge Distilled v0.1 |
|
|
| A **trained 325-parameter NumPy classifier** for five synthetic vibration scenarios: healthy, imbalance, misalignment, bearing impulses and rubbing. The teacher is centred kernel ridge regression using a product-rotation kernel with an exact efficient classical evaluation. |
|
|
| Developed by Rootcastle Engineering & Innovation. [Source and reproducible training](https://github.com/rootcastleco/sofia-distilled). |
|
|
| ## Training |
|
|
| 14 statistical/spectral features feed a 14 → 16 → 5 ReLU MLP. The teacher uses 600 fit examples and 200 validation examples; its 25-fit scale/penalty search is matched by an RBF control. After selection it is refitted on the 800 training/validation anchors. The student is trained for 120 epochs on 6,000 separately generated waveforms, with 1,000 student-validation examples for checkpoint selection. |
|
|
| The student objective combines hard-label cross-entropy and temperature-2 teacher-target KL. A supervised-only student with identical initialization, data, architecture and optimization budget is retained as a control. Train-only standardization precedes fixed arctangent angle encoding for the teacher. |
|
|
| ## Held-out results |
|
|
| | Model | Synthetic test accuracy / balanced accuracy | |
| | --- | ---: | |
| | Product-kernel teacher | 99.933% | |
| | Matched-budget RBF teacher | 99.933% | |
| | Distilled student | 99.933% | |
| | Supervised-only student | 100.000% | |
|
|
| The balanced test set has 1,500 examples. The distilled student reaches 100.000% on a separately generated shaft-frequency shift of 43–65 Hz, versus 18–42 Hz during training. These results concern easy analytical formulas, not machine diagnosis. The control outperforms distillation on the in-distribution accuracy measure; no distillation quality advantage is claimed. |
|
|
| ## Use |
|
|
| Download this repository, then run its standalone loader: |
|
|
| ```python |
| import numpy as np |
| from inference import EdgeModel |
| |
| model = EdgeModel(".") |
| signal = np.load("acceleration.npy", allow_pickle=False) |
| result = model.predict_signal(signal, sample_rate=2048, shaft_hz=25) |
| print(result) |
| ``` |
|
|
| Acceleration values must be in g. Training windows have 1,024 samples at 2,048 Hz. The loader validates feature order, dimensions, finiteness and the weight SHA-256 before inference. |
|
|
| ## Artifacts and provenance |
|
|
| - `model.npz`: final float32 student parameters and training-only scaler. |
| - `config.json`: exact ordered feature schema and weight digest. |
| - `kernel_teacher.npz`: anchors, ridge coefficients, centring statistics and selected hyperparameters. |
| - `dataset.npz`: all fit/validation/test feature matrices, labels and held-out scores. |
| - `metrics.json`, `protocol.json`, `training_history.json`: measured results and run configuration. |
| - `verification_report.json`: independent kernel and noise control experiments. |
|
|
| The method follows [Quantum Artificial Intelligence with Verifiable Kernels](https://www.academia.edu/175377730/Quantum_Artificial_Intelligence_with_Verifiable_Kernels). Calculations use classical CPU hardware. Read `VERIFICATION.md` for exact scope. |
|
|
| ## Limitations |
|
|
| No measured machine data, field validation or calibrated confidence. Softmax values are scores, not diagnostic certainty. The range check is a feature-distance heuristic. The model cannot validate arbitrary sensor placement, machines or units. It is intended for research and demonstration; it has no machinery actuation API. Only one Edge training seed was run. No quantum processor or quantum computational advantage is claimed. |
|
|
