Edge Sentinel Neural
Edge Sentinel Neural is a compact temporal convolutional autoencoder for industrial telemetry. It trains only on normal windows from devices 0-7, selects its anomaly threshold on devices 8-9, and reports final performance on entirely unseen devices 10-11.
It consumes the reproducible telemetry generated by the sibling edge-sentinel-ml
project and complements that project's supervised and Isolation Forest baselines.
Reproduce
uv run python projects/edge-sentinel-ml/generate_data.py
uv run python projects/edge-sentinel-neural/train.py
Verified results
The 2,860-parameter model trained on 3,946 normal windows. Its threshold was selected on devices 8-9 and then frozen:
| Split | ROC-AUC | Average precision | F1 | Recall | False-positive rate |
|---|---|---|---|---|---|
| Validation devices 8-9 | 0.9665 | 0.9539 | 0.9365 | 0.8906 | 0.0031 |
| Test devices 10-11 | 0.9784 | 0.9695 | 0.9532 | 0.9176 | 0.0020 |
The test confusion matrix was [[987, 2], [21, 234]] over 1,244 overlapping temporal
windows. Because a window is labeled anomalous when any constituent timestamp is
anomalous, these figures should not be compared directly with the row-level classical
benchmark.