This is what sense is supposed to be able to do…for many of us the reality is a little different…
What is an acceptable length of time for “a while”?
Example…My pool filter has never been detected. I waited a year and then had reason to move the clamps and peform a data reset. This has meant that I now need to wait “a while” longer…
Yes,
Their marketing is terrible.
That shot you show makes it sound like Sense is going to tell you something when the reality is, you’ll have to use notifications, look at your timeline and use your own reasoning to pit together a picture of what’s happening.
As far as how long to wait for the pool pump, if it’s variable speed then possibly NEVER.
I feel your pain !
I have shared my Time2Detection chart before, but here’s an update. Many detections in the first week or so, but they continue trickling out. Some of the multiple vacuum cleaners are either other power tools, or our cleaning person’s different vacuums plugged into different legs at our house.
In theory, over time, the Sense-sentience time has decreased. i.e. The delay between install & a good number of detections, when Sense “gets a feel” for your house, has gone down. There is a limit. In the ideal case your devices would all have to go on and off in there different modes at least one. In practice, the number of on-off cycles required for detection can be very large, depending upon the device. That said, there is nothing to stop Sense trending toward the ideal … which, actually, it seems to be doing. The dataset has grown and the “whiles” get shorter, especially after you do a reset … there’s a kind of leapfrogging that happens akin to countries that didn’t have any communications services suddenly having fiber optic or 5G and bypassing all that expensive and laborious copper stuff.
I’m two days post resets and got 3 detections this morning. Two I had before, a heat pump and a refrigerator. The third was new, a refrigerator I added but Sense didn’t detect in the couple of months before the reset.
These fast detections makes me feel like not all DATA is lost from a reset. If machine learning takes hundreds or more cycles then how would I ha e such fast detections?
I could see where they ha e DATA for the same identical product across many users but I ran into the water heater replacement where the same exact model wasn’t detected the same. I’m thinking this new fridge was probably on the edge of Being detected before I decided to reset.
Detections are probably coming faster because:
I would say it’s more misleading than terrible per-se. If the marketing had been more accurate, I would not have purchased it in the first place. Now that I have it, although it doesn’t do quite what I expected, it is useful and I enjoy having it. It has probably already saved me some money by the early identification of a potential AC Fan or Compressor failure.
Guess misleading would be a better word choice.
I believe because it is misleading, a lot of people have expectations that are higher than the real world capabilities.
I see what is marketed as what the future goal is or what the hope for it is. I haven’t really read here where anyone has that “perfect” home setup, of such an animal exists.
I’m reposting a link to this Sense video that gives some insights into training and crowdsourced device classification. This is what “training” looks like - a bunch of color coded clusters of on and off events in a 17+ dimension feature space. We’re only looking at a 2 dimensional (power and phase) slice.
I’m back after a number of months and sense has detected a bunch of things (and even merged them) - its clever, but the issue is its not clever in the way I need it to be clever.
Its still very tough to just wait for the device to be detected!
Annoyingly the sense has classified my cleaners vacuum, but hasn’t classified my washing machine or dryer (but it has my dishwasher, 2 TVs and coffee machine).
The issue with the lack of ability to ‘train’ is still likely to be an ongoing problem in that the devices I WANT to show up, i’m supposed to wait an unknown amount of time for.
Whereas I can highlight here is my system with the thing off, now its on. Repeat the next day, and maybe a different time of day. Any good AI in the back end could extrapolate that, and if more info is needed then declare that (if its a niche piece of hardware).
I’m still waiting for it to detect my Tesla Model 3 (6 months later).
The example of ResNet also isn’t overly accurate as AI can be trained. Language Models, forms understanding, Object Vision, Face recognition - they all require a TRAINING phase - here’s a 1000 pictures of an orange, AI now knows what an orange is… A more accurate analogy is the Human Biome project - you can detect bacteria in the gut, some you know, but MILLIONS we don’t (in a single person). So require many humans with the same bacteria to identify them (i guess the human genome project is similar, but that’s been cracked) - its impossible to train a model here as you don’t know what you’re detecting.
Equally if you have detected a device, that doesn’t exist in the database, maybe unlock a beta function that shows ALL detected devices in a users house, that the user can self name. The user simply sets an alert on that device when it turns on, they can narrow down what that device is, and provide a device type (and confidence score). This is crowd sourced learning and very efficient (but you will need a different AI model to interpret the results).
All of a sudden I get unknown device show up in my app. Set the alert to tell me when its on and off, and over a week i’ll probably figure out what it is and can submit it.
It seems you’re taking a purist approach which relies on the device being correctly identified (and even then its not always 100% correct).
As discussed at length I n other threads, if we were able to train or rename devices, our inputs would become part of the database and shared with all users. A single mistake would corrupt the entire database.
When I first installed Sense, I also thought of it as AI. It’s actually much different and is classified as “machine learning”. While the two share some traits, they are much different.
It makes it difficult for every one of us to understand when we recognize patterns from our devices and think, “Why can’t Sense see this?”.
The team at Sense has made great improvements with detections in the time I’ve had it (since January).
You mentioned your Tesla, what charger are you using? If your charging with the 16 amp instead of the 32, that’s likely why your missing the detection.
Welcome back,
Understand your frustration and I wish there was a magic way to train Sense. I would love it too. But I do accept their answers vs. your assumptions for a few reasons. Check this out, then see my reasons:
I think it would be fair to say that Sense hasn’t focused as much on devices that show long-term (30 sec to 3 hour on and off signatures) as they have on the short stuff, but I think that’s changing. Take a look:
BTW - I have had my Model 3 and Model S picked up on and off at different points in time along the way, but Tesla software updates seem to regularly change the charging profiles, which affects recognition.
I haven’t been on in a while and just had the small solar system on my house installed last week but still waiting for the power company to come out and change out the meter so i can turn the solar system on and take advantage of it but while this was happening i was also wondering why Sense doesn’t add a feature that would allow someone that doesn’t have solar to use the solar clamps to detect a circuit that is by itself on a single breaker like the many single device 220V breakers that i have. i.e. electric vehicle, dryer, A/C unit and stove. just a thought and obviously if someone wanted to do this then they would have to buy the optional solar adapter.
It’s being explored
Support is working on my case.
I don’t see it as a Sense priority but I can imagine the (business) logic to eventually get there.
This is a product that isn’t cheap, and comes with a LOT of caveats.
I’m happy to provide some input, but it is not Seeing devices - including the network detection (not finding hue or ESP bulbs)
As for a corrupt database, that’s assuming that there isn’t a tech team that can review the submissions.
I’m sorry I completely disagree (regardless of how many threads this has been discussed on) that manual training isn’t possible. There must be a tech team reviewing and comparing device signatures as right now it’s crowdsourced and I’m sending a bucket load of data - so helping sense build a better device.
I would like an in depth reason why training is not possible. AI, well ML is used to identify delta changes. But it’s like unraveling a ball of string. With each device detected the tangle is simplified.
However there will be a threshold point that most houses will reach when as many known devices (tangles) that can be detected have been - and now there are 2 options (remembering that these devices are detected as we, the paying users are sending data to sense)
Hi @jamieeburgess,
A question and a couple of thoughts.
First, you mention that Sense isn’t finding your Hue ? My integration with my Hue hub worked just fine until I decided to put my Sense and smartplugs on a separate subnet. Then my Sense stopped talking to my Hue because it was on a different subnet, and due to my home networking, I can’t move the hardwired Hue over to the other subnet. So I’m living in a self-imposed non-integrated situation. Maybe you have a similar situation ?
Second, you ask why training is not possible - There’s a well written technical article here, that explains the challenges here, plus some of the ways you can actually help:
There’s a whole paragraph in there that deals with why your suggestion isn’t as easy as you make it seem.
Third, I have been seeing progress. My AC system detections are much better than they were during the last cooling season even though I installed new compressors in July. You might want to read this article as well:
Mike: I am a retired ATCS from the USCG. I have to agree with you about fixing the app and allowing us to train & have a little more control in determining and assigning electrical usage devices showing in the SENSE APP. I have an electrical use item that is named “AC2” for the lack of a better name, but I cannot figure out what device it is for. It’s using an average of 402 watts for an average of 9.2 minutes when on. I have popped every circuit breaker in my electrical panel and cannot figure out what the mystery device is.
I’m still trying to figure this thing out.
Barry Philippy
Welcome to the forum, Barry. If you’re trying to reply to someone directly, you can either hit the “Reply” button in the bottom right corner of their post or direct your post at them by using @ followed by their name on here. Guessing you’re replying to @mike_gessner?
@kevin1 Thanks for sharing those. And glad you found that article both “well written” and “technical.” That means a lot coming from you! I’ll tell my mom she did a good job teaching me how to write 
@jamieeburgess I hope kevin’s post lends some clarity to the difficulty. We have some great minds in data science and machine learning here. If a training mode was possible to do right now, there would be no reason not to do it. It’s probably the most asked for feature!
Let’s enshrine this as the auto-response moving forward! Your mom trained you well!
[If I could relive my childhood I would use that line when being forced to do things I didn’t want to do. Keep it safe from the kids!]
The issue isn’t about the big brains on the team (although I’d argue if it’s the most requested feature why they aren’t focussing ways to achieve it, and documenting the journey); the issue is that the current model (not AI model) but structure of how you discover devices seems to have hit a plateau due to the number of people buying the sense device. You’re now at a point where only new purchasers will help untangle the data from existing users homes - that will slow as users are sharing feedback that it may not be worth buying a sense. You need to keep forward progress by keeping existing users happy, through device discovery - otherwise the whole thing grinds to a halt.
This means it will cost money, and resource to change that model which it seems the company is not willing to do based on the party line above.
I read the article that continually gets pushed, but it doesn’t actually answer the actual question.
There are voice AI models that can hear people, identify them, and even transcribe them (that was at first simplified with a camera, but you know what made it easier? Add another microphone to make an array - increase the number of data points for triangulation and use a good voice model) - if you are currently stalled on the number of data points you need to find a new way to get more data points with the same number of sense hardware out there.
There are a few ways to do this in my small brain:
