We put together an explainer video on machine learning that you might want to check out. For the DS and ML experts here, it’s probably a bit elementary, but it’s a nice overview on what machine learning is for those unfamiliar…and for why it can be so difficult.
I would just like to point out that even though Skynet only took 25 days to become artificially conscious, it took a lot longer to figure out how to stop people trying to unplug it.
@RyanAtSense Thanks for the video, it helps understand some of the quirky things that I see in my system.
Question - Does all this data crunching happen in my device or just some of it, or none of it?
I had a Cyberdyne Systems t-shirt in high school. Whether or not that means you should be afraid is up to you
Some of it. Much of the real heavy “learning” happens in the cloud, but a lot of the basic signal processing/filtering happens on the device prior to going to our servers.
Why not have your customers pay you to do the supervised learning? If you had a Wi-Fi CT clamp that acted like the HS 100 where we would hang it off items that sense is not picking up. I could hang it off my furnace and even thought its an ecm motor I could tell sense that all random ramping of the motor is all coming from my furnace.
If you could separate the data into layers where layer 1 is all raw data that sense is collecting currently and layer 2 is the customer supervised learning; then you can still let sense ML learn on its own without being skewed by layer 2 customer data. This would also give the customer the ability to tag every thing in their homes if they choose (ie happier customers).
It could also help with trouble shooting and confirming found devices. The Wi-Fi CT clamp could be put on an item that is impeding sense from making other discoveries; the customer could then tag that item on his/her end and tell sense in the back end to ignore the noise.
I think a lot of customers would love the opportunity to tag devices. Some items would not even need a Wi-Fi CT clamp to tag, lights would be very easy to tag with in the app. If that tag is tagged under the customer layer 2 data then sense’s raw data would not be affected by the customer.
I think some customers feel a little misled when they see images like this:
And find that they are not getting the same results.
I’m not sure I fully follow how what you’re suggesting would be different than the current smart plug integration, other than just 240V support? With a smart plug on everything in your home, theoretically you’d see perfect whole-home device detection (minus the 240V devices that an HS110 won’t technically work with). How does that differ from the panel-level “tagging” feature that you suggest?
Keep in mind, the smart plugs aren’t just for local device detection or remote control; they feed us fantastic ground truth data. It will take time, but the goal here is that you’ll eventually be able to put a smart plug on a device for a couple days, get it detected, and then move the smart plug to another device and repeat the process.
But even with another layer of data, without a smart plug or your WiFi-connected CT, how would layer 2 learn the signature of said lights? Or are you suggesting purely historical, i.e., scroll back in the power meter, tagging?
As for the app image, that’s from my home and is a real screen grab and not “manufactured” in any way. Though, I’m seeing your point that it actually doesn’t reflect the ML aspects of Sense, since all of those devices are on smart plugs. Fair point that I did not think of at the time! Still, I don’t feel that it’s misleading in any way; it showcases our integration with smart plugs, which is another core feature of Sense.
Does that work today? I saw some of the smart plug integrations but mostly wrote them off because I have no interest in connecting everything in my home to a smart plug just to get them to show up in Sense. But if I can do what you described then it’s much more reasonable.
It is not, thus my “eventually” language. Doing so takes a lot of data, which means smart plugs installed in the field. We’re starting to get good stuff though and hope to have something to share soon.
Thanks for the quick reply,
Yes the ability to put it on a 240v would be huge. I also think this would be handy for people with noisy devices in their homes. From what I understand the devices creating the nose are more often than not hardwired.
This is great that you guys are working on this. Like ziebelje said I thought that I would need a smart plug on every device in the house. I also have no interest in connecting everything just to get it to show up in sense. The ability to be able to move one around the house would be 90% of what I was hoping for with a Wi-Fi CT clamp.
That’s right the smart plug, the WiFi-connected CT or historical tagging would populate layer 2. That way senses machine learning would not get disrupted by supervised learning or the tagging. As for tagging, it would happen by scrolling back i.e. flicking the hall lights on and off then like you said scroll back 1 min and tag the bump up and down. The information from tagging of lights would be useless to sense and just cause more confusion for it. Nobody is going to have the same light bulbs or number of lights in there hallway. I guess in reality only historical app tagging would only need to populate layer 2 because of the inaccuracies/differences in people’s homes.
I figured that they were smart plugs or wifi connected devices. I do think that a few people watch the this old house episode and then watch what is machine learning and might assume incorrectly that all the devices were ML. You guys have done a great job with sense and I think everyone wants to see it be the best it can be.
Definitely looking forward to what you have to share!
Thanks for the detailed reply. I absolutely agree (and many others here do as well) about a 240V “smart plug” or similar device. This is absolutely not an official recommendation, but a few people here have had success using the HS110 with 240V devices. I’ll let them chime in on details for that. Again, not telling you to do so! That said, new hardware is a long, incredibly resource intensive process and we’re focused on the core Sense monitor at this time. New hardware (or significant mods to existing hardware) isn’t likely in the immediate future. But, if a 240V plug comes to market, we’ll definitely look into a possible integration.
The historical tagging point has been brought up prior but I can’t seem to find it in Product Wishlist. It’s a good idea and I personally could see myself using it. You’re welcome to create a thread focused on that bit in Product Wishlist and see what others think.
Wouldn’t even need new hardware. Just the ability to add multiple sense units to 240V circuits roughly similar to what is needed for 400V systems. In other words the ability to tell the backend to use one or more sense units as a smart plug.
That’s a very bulky and expensive smart plug!
Ha! It’ll be a fancy smart plug for the rich. Kidding aside it would go great with a bulky and expensive piece of equipment like a heat pump (central or even ductless). Since it would sample a million times a second it could be used to track equipment health that you couldn’t do with a normal smart plug. Might be too expensive for a hot water tank but if it could predict tank failure it could prove it’s worth even there. If I had a heat pump I wouldn’t mind having one attached if it could do diagnostics and maybe even if it couldn’t. Up here a Mitsubishi heat pump costs about $4000.
Paging @ixu since he actually has a separate Sense monitor set up in the way you describe. It’s not a supported or tested use case by us, but he can at least tell you how it’s working for him.
That’s alright. No need for one now. I was just making a point about not necessarily needing separate hardware.
Yes, we had this conversation before …
And just to go over something I also pointed out: A second Sense (Solar) gives you 2 Smart Plugs so works out at $175/plug. When you factor in that the likely monitoring would be on your biggest energy suckers, the precise tracking and notifications would probably pay off in the long run.
Getting back to the OP, the Machine Learning aspect of the Two Senses equation is an interesting one. I feel like things will inevitably move toward more ML being done on the Solar input and also potentially that the Solar input could be retooled (User-Defined) for some other input. Let us know one way or the other @RyanAtSense when that happens … I don’t want to be corrupting the algorithm!
Not sure if this is the correct place to ask this question. Sense has discovered my microwave and a ‘Light 1’. It looks like this light is the light that is in the microwave. I see the merge option. I am not sure what this does and does it make sense to merge the microwave and the light? Are there any benefits to keeping them separate?
If you are really interested in knowing just how much electricity that light uses then sure, keep it separate. If you just don’t care then merge it. Merging the light into the microwave will just make it show up as part of the Microwave when it’s on. You can always unmerge it later if you really want.
The merge is not destructive - you can always unmerge them later and their histories will be preserved. You might want to be more careful with merges with devices on smart-plugs, because they behave differently.
Merge enables unified accounting so you can see all the power associated with a single microwave in one device display and via data export. Sense might also use the info long term to relate the two.
I keep mine separate today, partially because the places I might consider using it - linking my furnace fan to my AC, are more complicated than I would like.
Sorry to be late to the party on this thread, but I wonder: rather than using new smart plugs, etc., how about re-purposing the “Solar” CTs, which I’m guessing currently assume energy production, as a “single-device”, or “single-circuit” CT? If you want to know the exact power usage of a given circuit, you clamp the extra pair of CTs on that circuit. Sense could be told “this pair of CTs is not power production – it is Device X”, and could correlate / subtract that device’s signature from the “whole house” power consumption.