← Back to the chapter · Next detail →
Technical detail · IoT Machine Learning Pipeline
A pump-classifier release record
This example keeps the decision, inputs, feature recipe and operational response together. It makes the release concrete without replacing the evidence behind each field. The recorded gateway inference target is one deployment choice; the release review must still establish that it meets the latency and resource requirements.
iot_ml_pipeline_record:
decision: classify pump vibration windows for maintenance review
data_sources: accelerometer stream, pump state, maintenance ticket outcomes
labels: reviewed tickets mapped to time windows, label delay recorded
features: fixed-length vibration summaries, spectral peaks, operating-state context
split: chronological holdout plus pump-held-out check for generalization
metric: recall for likely faults, false-alarm rate for review queue, latency for action
deployment: gateway inference with versioned model, scaler, threshold, and rollback
fallback: unsupported input or low confidence routes to manual review
monitoring: drift, missing windows, alert rate, feedback labels, retraining gate
retest_trigger: sensor, firmware, pump class, feature code, threshold, or workflow change