Appliance fingerprints: teaching software to read a power meter

Fourteen months ago I started logging plug-level power draw across the apartment, one sample every two seconds. The original question was cost. The interesting question turned out to be identity: which appliance is drawing right now?

The approach is unsophisticated but effective. Steady-state draw separates the big consumers immediately. For the ambiguous ones, the turn-on transient -- the first ten seconds of the envelope -- carries a shape that is stable per device.

# feature sketch per switch-on event
# mean, peak, time-to-peak, slope(10s) -> nearest centroid

The classifier is nearest-centroid over five features, retrained monthly. Accuracy is not research-grade; the kettle impersonating the air conditioner was the first documented misclassification, and the fix was one more week of samples.

The payoff is a baseline: the system now flags anomalies -- the freezer compressor drawing 15% above its usual envelope -- before anything fails. That early warning alone justified the whole project.