@jamieeburgess,
All constructive… But I think you are probably wrong on a few assumptions
“You’re now at a point where only new purchasers will help untangle the data from existing users homes” - There’s no big, magic “untangling” that’s going to come from more users. I believe that the big opportunity for improvement will come from more thoroughly labeled datasets, not just more data. That means incorporating data from smartplugs and other ground truth sources. Massive labeled datasets is what eventually enabled the big machine learning breakthroughs for vision. Labeled datasets and man hours are the key.
Phase is a measurement between the current waveform and voltage waveform on each leg, not a measurement between -120 and 120.
TP-Link Kasa and Wemo smartplugs are that “separate device that could be plugged between a device and the wall for a period of time to capture more data points, and add to your device pool“. Maybe they are a little too permanent for your liking today, but I’m surprised that you don’t recognize them as such. Sense Connects with Smart Plugs - Sense Blog
On 1. It’s about getting more data points on the same data with a new or different perspective (rather than just more data)
And 3. Now in all honesty THIS is the most valuable thing I’ve seen in the discussion. This is the link the mods should send to folks bothered about the time it takes!!! however as a customer I’m pissed I have to buy MORE stuff to help the detection. So, knowing this helps sense and the integration is well crafted, why don’t sense either provide a discount (buy in bulk and resell) as the data they will get from it is worth the discount they offer… OR make their own while label solution with a ESP8266 or Esp32 (like sonoff for example) that they could sell for say $15 and you could plug in a device for a week, once detected, change the device.
I agree, that the smartplug capabilities are underplayed. FYI, I bought most of my HS110s on a Father’s Day Amazon special where they offered a 2-Pack for 25$. But now the best price seems to hover around 20$ per unit.
I was a little tweaked about the added cost, but found it worth it when I started to fill in some of the big unknowns in my house, especially stuff like my networking rack and AV “rack”.
Honestly, this is good feedback to hear. I can be a bit conservative in mentioning them as a stopgap solution to some detection problems, partially because of the added cost. We do have some plans along those lines though that I can’t discuss right now but hope to share soon.
Sharing this, just so you can see where the state of the art was about 5 years before Sense came along. Lots of very expensive customized hardware that has far more limited sample rates. Everything looked great from the research work (high coverage and accuracy), but taking the technology into the real world is much harder.
Now we have simple, relatively inexpensive monitor hardware and consumer cost probes (smartplugs) that have much better sample rates.
I think you can still get those special networked power plugs, but they now only come connected to the cloud with machine learning behind them, for 125$ for 4 outlets, plus 12$ per year subscription…
Given the intersection between tinkerers/makers and energy monitor users I would strongly recommend offering an API to write data from existing devices (Wemo D1 mini, sonoff, ESP8266, NodeMCU and ESP32) so that existing home solutions could be added and data sent (imagine Home Assistant, Hubitat, SmartThings etc)
I’ll leave it to your big brain scientists to figure out calibration
The smart plugs won’t aid in detection. The plugs will be immediately detected and you will tell Sense what is plugged into them. Let’s say you use for a refrigerator, as soon as you remove the smart plug, the detection for that refrigerator will be lost.
I blend the data from my Tesla chargers with the my Sense data for a more complete picture, using the data export. Sense only intermittently finds the charging of my Model 3 (ever Tesla firmware release seems to ramble things) and hasn’t detected my Model S for a while. But export the Sense data, filter out the partial Tesla data in that, then add in my Tesla charger data and I have only around 15-20% Other.
Harder to use data from smart hub sources that only provide on/off/dimmer setting. Except for Philips Hue with actually converts that data to calibrated power info. I’m guessing that’s why that hub is already integrated with Sense.
ps: Actually some other smart hubs do report power for some select device types. But the vast majority only include on/off/dimmer level, plus the ones that do report power, typically report in every 30 secs or so, unlike the Wemo Insight or Kasa plugs that report every 2 seconds.
I actually get the impression after reading their blog post and replies; that this is an issue with their technology. Their machine learning does not take such input well if at all.
Manually training the system also seems to not be a goal of their company. I also get the impression that the company’s primary goal is to develop a completely unsupervised machine learning system for more commercial/big utility use in the future and that this was just a stepping stone.
They could of course just put in a “train function” that does not work(eg. dummy button in elevators) to shut everyone up
Machine learning does just fine with input. Supervised learning is very effective, but only if the dataset is very “complete” and the labeling has sufficient spacial accuracy in comparison to the detection features. The problem is that humans aren’t particularly good at either of those given the sub-second on/off signature detection windows that Sense uses. But smartplugs are pretty effective at both.
But as you suggest @andrewchong, perhaps a placebo train mode might make some people happier. Personally, I’m still betting on Sense’s approach.
ps: Both Neurio and Smappee have a train mode, but from everything I have heard from users, the train mode is worse than useless.
I have my share of complaints but “training” is not on the list.
Nowhere in their materials have they mentioned it was a feature or that it would ever be added.
I know it’s hard to be patient but it really is the best way to enjoy the Sense experience. The mix ups between appliances are normal and most do improve.
This is also where Sense does have a type of training. When you have two devices triggering as one detection, use the “this device is not on” button for one of them a few times.
You can’t tell it that it is correct about a detection but you can tell it that it is wrong. Kind of like my wife does me.
Customers always want lots of things. Customers want Nest to support more than three homes or integrate with HomeKit, but they don’t. Customers want Apple to let them merge apple ID’s, but they won’t.
Not sure if its the beginning of this thread or one of the many others of people insisting they know better than Sense does, but Sense had a training mode in their early pre-release. They turned it off because it was not helping anything and I suspect, actually created more problems. Now we have people who think they should be able to train it, but they don’t actually know how to implement it other than to suggest it shoudn’t be hard. Imagine if there was a training feature, but it didn’t actually do much. THen the forums would be full of “I don’t understand, I just turned on light switch on and off 40 times and told Sense I was doing it, why is it still not working!”
I don’t believe that Sense has ever said that training is not something they will ever support, but it is not something they are supporting now or in the near future.
@grovehaven Just to be super super clear, I am in no way affiliated with Sense. I am a customer myself. I found out recently that I do actually know someone who works at the company, but its someone I haven’t spoken with in 18 years and is merely a “oh, thats neat to know what they are doing” kind of thing.
