I can’t help but wonder why you’ve structured the device discovery process the way it presently works. Wouldn’t it be a lot easier for users to create a list of what they have and let your software do the work of “detecting” them? Then you could simply have them disconnect or toggle specific devices if your software needed assistance in determining which specific unit was responsible for a particular bit of usage data?
Unlike TRAINING, I have a hunch this might be possible in the future. Sense first needs to build a pretty significant database of appliances which it has learned… But then? It may be possible to simply enter you model numbers and then wait for Sense to find them.
My feeling is that it will continue to get easier into the future, but it ain’t gonna be all roses and puppies.
My problem is that this house is huge and has a lot of appliances. Sense says it has discovered a fridge. Well, there are three fridge/freezer units and one freezer. I have no idea how to identify which one might be “device 1”. Ditto with all of the other devices it has discovered. There’s simply no clear way to even begin identifying them. But if it was working with a specific list of devices it seems like it would be relatively easy to give me instructions on which one to unplug so that it could correctly match usage data either the responsible devices.
Hey @ananias, if you haven’t already, I recommend taking a look at these two blog articles we have on the topic of device detection:
We do have plans to allow users to enter information about undetected devices into Sense. Ultimately, detection comes down to Sense seeing devices operate a number of times in their usual context, so even with that information it will still take a little while to detect certain devices.
With regards to your current situation, fridges can be a little tough since they’re more passive in consuming energy than other some devices. You could wait till the fridge in the app is detected as on, and try unplugging a fridge to see if it matches up. Once you figure out which fridge it is, there is a section in device details to mark the ‘Location’ and the make/model of the device. We use the make/model information to help inform our detection algorithms (eventually down to the make/model).
Here’s an article that may be helpful on some tips for identifying devices in your home: https://blog.sense.com/articles/follow-the-breadcrumbs-with-device-usage-history/.
Hey Ben, just bumping this thread. Specifically I’m interested in the possible enhancement to input unknown devices. I think I’ve seen it mentioned a few times, and agree it could be very helpful in identifying what the big hitters would be for development of new device models. Could say how many users have x device. Or if the data contains estimates of wattage and runtimes, a function of the estimated kwh.
Additionally I wonder if it would be possible to merge patterns across users to accelerate detection. For example, user1 and user 2 have a known device you should be able to join the two for patterns. Possibly using estimated wattage to narrow the scope. So if it takes 50 cycles to develop a model for a device it might be possible to do it in half the time.
Of course I’m not a data scientist so all of this may seem like a good idea in theory, but difficult in practice. I’m sure not every device behaves the same in the real word either.
The ability to input unknown devices is definitely still on the docket but I don’t have a specific timeline for it quite yet.
With regards to detection model development, the data science team does look at and use a vast amount of data (across homes in the Sense community). In terms of using these models to detect devices in homes, given that there is so much going on in a home at one time, it can take a little while to be able to say with relative confidence that an electrical signal matches a known model.
@BenAtSense I understand that Sense is designed on the principles of Unsupervised Machine Learning, but wouldn’t it be beneficial to give Sense a list of appliances to target? This could easily be incorporated in the setup process, where the user has the ability to check off of a list the appliances they currently have. At a minimum, it would eliminate the foolishness of incorrect guesses, and provide a “wishlist” of sorts for appliances that the user is looking for Sense to recognize.
@Justin_time, this is similar to what we’re thinking with regards to inputting information about unknown devices. Still no specific timeline for that, but definitely something we’ll be tackling this year!
Even if Sense can’t identify the device, can’t it still do pattern recognition to simply determine when a device turns on and recognize it as a specific device by it’s wave pattern? It would be very easy to walk around the house and turn on/off devices and then label them in Sense once it detects it. Heck, I’d even go so far as turning everything off and doing items one at a time just so Sense could identify the wave pattern it produces. It would also help break down the big cluster of “Always On” devices.
Given that Sense is using machine learning, both pattern recognition and identification are are one in the same. Google the terms “neural network classification” to learn more. Machine learning typically requires thousands of repetitions under a variety of background condition to “learn” a device, so you probably don’t want to sign up to manually providing a training set…
I just installed my Sense last night… Thing is soooo cool… What I think would be nice is to just be able to add the devices manually… Turn on the air fryer, watts goes up 800, I should then be able to label / classify that as a device…
Can’t wait until mine starts discovering things! Very amazing tech!
Hi @kmccb - and welcome to the community!
We definitely hear you on manual entry of devices. It’s an interesting challenge, and something we’ve heard about a lot. If you haven’t seen these already, we have a couple of articles on our blog that go into a bit more of the mechanics as to how the device detection algorithms work, and your input can help refine them:
How Sense Learns About Your Devices
Machine Learning vs. Human Learning
Hopefully, you should start seeing some devices detected in the coming days, weeks, and months, and depending on how often you use that air fryer, it might just pop up as a heat source, and you’ll be able to give it the name, and help the whole community.
Thanks for your support!
I too wonder about why Sense wants to crunch on the unknown list for so long. If it would just release the devices homeowners could identify and label them in a fraction of the time. This would still allow the Sense database to be built but much quicker.
You guys are kind of missing that your specific device in your house looks different than the same exact device in a different house, with a different power distribution network and noise contributors.
Kevin, you are correct, I was not aware that the exact same appliance would look different on a different power grid or supplier. But it still seems like the end user could identify devices faster and upload the info back to the Sense master DB. It would be accomplishing the same goal. But I think Sense as a company is looking at the bigger picture, as they should. I’ll keep trying to label devices with as much info as possible.
I don’t want user identifications, I want AI to locate the devices for me.
Guys we all have slightly different hertz and voltage depending on your location on the grid and it’s relative to your supplier’s grid distribution equipment. For example, I actually receive 124V per leg in my breakerbox… My farther in law receives 119V.
Also, regarding the solution too turning things off and on to be identified… I struggled with my humidifier being detected as “Unknown” …by turning it off and on at the humidifier soft button or plug and unpluging it, Sense was undeniably detecting when it was off and on. The humidifier starts and stops differently when I allow it to operate on its own humidity sensor… It’s possibly a gradual rise in power and shutdown.
Shutting things off and on isn’t black and white… Unless maybe it’s incandescent light bulbs…
… … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … V heart V
Another possibility would be to use a WiFi enabled sense device that goes between a user selected device and the outlet. It could then send data remotely and get a baseline for that device and what it looks within your home. I’m new here but just brainstorming how you could shorten or improve the detection process.
Ps, I’m new here, installed my Sense yesterday, with a couple of devices already detected.
You are hitting close to best way to accelerate learning of straightforward devices. But the value of the remote is not so much in establishing a baseline, but rather appending whether a specific device is on/off to the collected waveform data. That extra feedback on the device state enables “supervised learning” using the “ground truth” from the probe attached directly to the device. Sure, humans could try to supply similar training by switching a device on or off in coordination with the Sense UI, but that would typically be far less accurate (@MachoDrone gives one good example) , plus only capture data against a limited set of household conditions.
Much of the precursor research for Sense-like machine learning was done with plug-in type monitors to supervise the machine learning. Take a look at the pictures in this research paper.
@kevin1 pretty much pre-empted my response here.
This is definitely something we’re looking into, but there is a line to straddle between unwieldiness and ease of install. Sense is great — and I love it as a consumer as well — partially because it’s easy to install. Once you introduce 20 Wi-Fi probes into the mix, it can get pretty cumbersome. There’s definitely a use case here for tough-to-detect devices, like variable speed motors, though.