Why Sense needs time to detect device
Sense detects fluctuations in electrical energy use and our software utilizes machine learning to distinguish one appliance from another. The nature of machine learning requires a lot of data before accurate models can be formed which is why it takes time for device detection to improve. Sense needs to see your devices in its typical operating context in order to accurately identify them. Our data science team is constantly working to confirm and push out our device models to ensure they will take into account the variation between home to home, and even the same device running at different times.
I can see the deviceâs graph when I turn it on and off, why canât I tag it?
Device detection is very much like trying to single out one personâs voice while 30 people are talking at the same time. It wouldnât help to know that personâs voice in isolation from the crowd. In order to detect them, you would need listen hard through the crowd. Likewise, our homes are busy, electrically loud places. Sense needs to learn the signatures of your devices in the context of this cacophony. An isolated signature of a television turning on would offer little to help Sense identify your TV in the unique contexts in which it is often used. Hereâs a video of one of our data scientists demonstrating how electrical signals are identified. Feel free to check out our relevant blog article: How Sense Learns About Your Devices?
How often do new devices get pushed?
Device detection updates are pushed automatically, and will vary from home to home based on what it has been learning. Sense needs to see many repetitions of a device working in your home before we can model it, and weâd only want to show you what weâre sure of.
I get that we shouldnât âtrainâ sense, but realistically I think this option would go a long way in helping it learn devices more quickly. If at the very least Sense had an initial questionnaire that included some basic questions (electric/gas heater, electric/gas water heater, number of garage door openers, electric/gas range) this could greatly reduce the number of unknowns that Sense has to figure out on itâs own.
Along those same lines, what about having the Sense app immediately prompt questions when it detects something different. There are only a few inputs sense can use to narrow these things down so whatâs wrong with a little human input in order to speed things up a bit? A good example is that when I turn on the lights in my living room (four incandescent bulbs) I can literally see the watts jump on the app. "Why canât the app bring up a notification like âwe saw something change did you turn something on/off?â You can then have a drop down of common things - lights, dishwasher, washer, etc. - and then an input or dropdown list of rooms (perhaps thatâs part of the initial questionnaire⌠name your rooms). While it doesnât have to be the final say, you can certainly use the inputs to help the algorithm narrow it down.
Thanks for the suggestion, David! Weâre definitely looking at more ways that users can be involved in the device detection process. We recently implemented a feature that gives users the ability to mark a device as ânot onâ when it incorrectly appears on, and we are planning on adding additional features in this vein.
Oooh, ânot onâ sounds interesting. Right now Iâm âDelete this Deviceâ when something never comes on or is obviously wrong. Right now I have an âUnnamed motorâ that came on for 14 secs at 1.5 amps randomly (about 6 times), starting at 2 am thru 8 am+, and only happened for one day. I couldnât figure it out even though I was looking for it when it came on, I think I will delete this âfindâ.
I think in rev 2 of the hardware, you could provide a pass through 110v power outlet for learning and detection, you could single out individual devices for learning purposes instead of trying to pick everything out of a crowd. Then we could individually help train sense about problem devices by providing a clean stream of data about itâs power signature.
Weâre looking into ways to incorporate smart outlets that could be plugged in anywhere. A pass thru would require access to the electrical panel. While many of our users on this forum might be comfortable with that (but we of course do not advise it), we wouldnât expect all of users to be.
This type of machine learning is dependent on samples. You take the sample, along with the user description, and stick it in repository. The AI still works as-is, but the end user could more readily tag and monitor their device and sense benefits from a clean âsampleâ to analyze. Of course you wouldnât trust user input implicitly or automatically push this upstream to the rest of the sense users.
I may simply be misunderstanding. Are you saying have a pass thru outlet on the Sense monitor itself? Or include an outlet with Sense that connects directly to Sense?
That would be a VERY big change in strategy for Sense, whoâs claim to fame is âmachine learningâ, sensing devices from their power signature, etc.
Smart outlets and reporting applications are widely available (as cheap as $7 eachâŚjust browse Amazon!), work quite well, and donât need the expensive/complex Sense system to do their job. They also control the connected devices, which Sense doesnât do. If we look at the cost of Sense, that would pay for about 4 smart outlets, or many more devices than most Sense customers can reliably monitor AND with full remote control.
Will be really interesting to see how this turns out.
I may be in the minority but this seems to me a very logical step. What better to improve machine learning then to feed âpoint of useâ data where possible?
Cancelling out the ânoiseyâ electronics like plasma tvs could overall improve the detection of other devices.
I certainly havenât seen a $7 smart outlet that provides graphed historical usage data. Iâd own multiple if I found one.
I agree with a Sense strategy to include âground truthâ sensors to improve and speed detection, even if it seems somewhat redundant. If you already own a couple of smart outlets that monitor power (Kill-A-Watts, TP-Link, Eve Energy), you know that there is huge value in an integrated UI that brings all the power usage together, rather than gluing one-time-readings together in a spreadsheet. And I would assert that the value/pricetag for a âground truth meterâ varies depending on what it does:
simple binary device on / off information to speed learning - 10$ / probe
rough time history (30 -> 1/2 sec samples) of device power/energy consumption for better identification of complex signatures, assessing accurate power usage between on and off events, and canceling out simple noise generators, plus could be displayed power meter. Could also be used to fill in / replace Always On devices - 40$-60$ / probe
detailed time history at the same scale as Sense - suspect that this type of device could be used to do accurate analysis of virtually all devices and home wiring effects. 150$ / probe
