After some recent discussion on if Sense has given up or not, I wanted to share a recent story about how my Sense has NOT given up on native device detection.
My wife and I had a Keurig that Sense natively detected. However, she didn’t like that it was on the opposite side of the kitchen from the sink, so refilling it every couple days was a pain, so she moved it. Of course that meant not just a different circuit, but a different leg.
In the next couple months we had it set up, Sense didn’t refund it. Then my wife wanted to get out the old coffee pot to make more than one cup at a time. Sense never found that coffee maker, but it also wasn’t used every day.
Fast forward to a couple weeks ago, and we had gotten a variety pack of K-cups for Christmas. One morning, I swapped out the drip coffee machine for the Keurig. I ran two empty loads through to clean it, made one cup of coffee, and sat down at the table for breakfast.
BAM! “Sense has found a new device”. It was the Keurig, and the Sense community ranked it with 100% certainty that it was a coffee maker.
I’m guessing that Sense’s algorithms had found the Keurig, but was waiting for it to turn back on and confirm. Once I gave it that on signature, within FOUR MINUTES I got the new device notification. I’m excited to have it back and merged it with the previous detection.
Great to hear that it refound your device! I’ve had great success with that as well when Support team has told me to delete a device because of an issue (merged signatures, overlap, incorrect) and it usually comes back in a week or two!
I just added a section to my guide to native detection, that touches on how and when Sense finds new devices or retunes old devices. Based on new info and hints from Sense.
How do Sense “Native Detection” Models Improve Over Time ?
This is a really useful article. I learned more about the inner workings (and what to expect) than any (perhaps all) the other information to date. HUGE thanks for making this available, great work.
Thanks for the feedback @andy ! I was trying to create a “theory of operation” guide that would help technically minded folks gain a little bit knowledge and agency when trying to figure out how to get the best results out of Sense (or at least understand less good results than they expect).
Sense recently gave me a few hints on the new section.
They recently that Sense models run iteratively on the past 60 days of data for the home, allowing each model to update according to the most recent behavior of the device. For example, if you started running a space heater twice as long as you normally do, the expected duration (duration model) would adjust according to the most recent behavior.
A Sense support guy described a little bit more of the process they use to help users with detection on problem devices.
The cluster / dimension space model also helps explain a little more technically why manual training is hard without using the “many people talking analogy”. Even if people were able to effectively mark the start and end transitions of only the devices that fit the native detection model (that’s a big if), they would end up capturing 2 new connected points in the dimension space. Do it 29 more times and you have 30 pairs of points in the dimension space. Those points mark roughly where that device lives but don’t help a whole lot in informing the clustering algorithms, because so much depends on what the neighborhood around those points looks like.
Great to hear! This is a correct guess - when we detect a new device now, we wait to issue the new device notification until the next time to we register the device as ‘ON’ so you have a better idea of what it could be. Thanks for sharing
