A few thoughts.
- Yes, I have seen lost off-detections in a number of detected devices. In most cases I don’t see them as harbingers of some thing bad.
- I think about Sense in these terms: The Sense monitor is tuned to “listen” for and report on on and off transitions within a sub-second time window. You see these transitions of interest that are tagged in Power Meter in the iOS/Android app.
- So there are some transitions the Sense monitor never “hears” at all (slow ramps like EVs, mini-splits) and Sense sorts out later at the mothership, and others in the Sense monitor sweet spot that get lost to noise.
- If a transition is reported by the monitor, the monitor takes a shot at classifying based on a set of models associated with that monitor. I’m assuming that classification of negative going transitions are also conditioned on the device, as Sense views it, being on. I also assume, based on what I have seen, that Sense doesn’t associate a model until it has “found” both an on and off signature for a prospective device.
- So your missing off signature could be a transition that was lost to noise and never reported by the monitor, or one that couldn’t be classified by the model as the off-signature for an on-device.
Here’s a case where blinking Christmas tree lights were enough to obscure off-signatures of a washing machine, in a house where the blinking lights weren’t on during the learning for detection.
As for Always On, it is purposefully slow to responding to changes in power usage. More here: