Yup, I think Sense “predicts” on and off events somewhat separately from steady state usage. This blog gives a hint of how it is done, at least for devices that have short on-signatures.
That’s why we see situations where Sense spots an on-event, but misses an off-event, leaving a device bubble on the display that has clearly turned off in the household (the bubble stays on until a watchdog timer turns it off).
I’m guessing that the two features that play the biggest role in identification and classification of the examples given are quick changes in current and phase angle. But that’s all in the context of AC voltage and current waveforms which are continuously varying at 60Hz (or 50Hz elsewhere). Sense has smart digital signal processing guys that have figured out the best way to extract these changes from AC waveforms (preconditioning), then feed them to the AI “magic box”.
I think you are thinking about AI the right way - a magic black box. You just have to think of everything Sense identifies as merely a high-percentage “prediction”. If you have the data to confirm (or deny) those predictions (like via a smartplug), Sense will improve it’s predictions for that device, the next time it goes through training. To be clear, Sense is still in data collection and experimenting mode with smartplug data, as @RyanAtSense states above, so don’t expect any improved magic yet.
Prediction kind of looks like this, except imagine inputs multi-dimensional in current, phase, others, and time.