High energy items not being identified

Excellent idea. Would be great to see!

@goalieman86, you are thinking the right way, it’s all about event detection and classification, both on and off. A few things to think about:

  • There are events you see in the Power Meter, then there are the events Sense “sees”. Except for special cases, like EV charging ramps and some heat pumps/mini splits, Sense only sees very short (like 1 second) on and off transitions. I believe that if you look at the Power Meter in phone/tablet app, you will see Sense tag every events/transitions it “sees” with a power number. Look at the transitions tagged in the Power Meter as the only thing Sense “sees” with its primary eyes.

  • Seeing / tagging one event is not enough - Recognition is needed. It is not useful to tag a single event that Sense “sees”. Machine learning has be able to be able to “recognize” a bunch of similar on or off events as distinct and unique, vs other transitions, before it can begin recognizing something. I often use the photo program face recognition analogy. It doesn’t help with face recognition, to tag a face, until the program has pulled together a bunch of faces. There is a school of thought (supervised learning) that says if you could label a whole bunch of these things that Sense sees as what they are, that could also work, but supervised learning tends to take hundreds to tens of thousands of labeled examples, something most users can’t reliably deliver on. A good discussion in this video about how recognition is done and why Sense chose to go the “unsupervised learning” route.
  • There’s also the challenge of pairing up on and off transitions.

I’m all for machine learning, but I’ll wager a quarter that the current owners of Sense are VERY, VERY interested in understanding ALL of their devices and are WILLING to HELP SENSE as much as they can. I get the feeling that the developers wanted to accomplish something VERY COOL, but eventually found it to be a LOT HARDER than they original expected. And from that original design platform, my guess is there was no thought on how the machine learning (ML) could have a human “tutor” it in the lesson plan!

As mentioned earlier, there is already something in the app where we can enter data on devices. Frankly, this could be thought of “cheating” (if we continue the “learning” references) but others might calling it “tutoring”. Here’s how it could work:

  1. User enters ALL the devices they can think of for their house, entering as much data as possible for each device, especially the heavy hitters. This includes model numbers, etc. During this entry process, the SENSE app could look into its database of devices and see if there is a match or something close. If there is a match, then the app would fill it in and note it as a match. If no match, then free-form entry would occur. (but it still would be categorized!)
  2. Whenever the user notes that any of the entered devices “just turned on”, they could tap the meter display and find some pull-down that would allow them to identify that change in wattage with a labeled device. Is that enough, most likely NO. But they can do the same for when it goes “Off”. And everytime this happens, if the human makes this identification, it would seem to me that SENSE should use this information (MY toaster just went on, just went off, just turned on, just turned off. ) that the tuturing should inform it that, hey. LIMIT the SENSE ML during each of the identified actions to TOASTERS!!!
  3. When SENSE decides it has detected a device, and it brings up the display of the choices, it should start with indicating it most likely is a Fill in the category and allow the user to find a drop-down menu of THAT category for which the human filled in the product(s) .
  4. If none of the devices appear correct, there should an option to “Do more homework”.

For my experience so far, perhaps I’m typical, maybe not, but my first detected device was the “Stove”. Two days before, I had entered the EXACT model stove I have, and since I had filled in that in devices, SENSE should have known the type of Stove I have, and if it had any signatures of this stove/range in its database, it should have compared them, in this case, rejecting the possibility of the detected device being a “stove”. In my case, the device it detected was my EV (M3/AWD using TMC). And not only was it wrong, but soon I found that even though it would report the EV as “turning on”, it was only showing the draw to be around 2KWh when in fact, it was clearly more like 7.5K and the text alerts happened precisely when charging started and stopped. In my case, perhaps out of naivety, I decided to delete the device and have SENSE go back to the books.

From reading many others experiences, It would wonderful to hear the SENSE product management and engineering teams fill us all in on how we can IMPROVE detection, rather than waiting for ML to magically do it’s stuff. It might be super-cool if it always worked, and worked in a month or less, but having the ability for the humans to “tutor” SENSE, even if it requires several “lessons”, I think we would all become much happier with the product.

1 Like

@ATechGuy, I’m betting the percentage of users that would love to “tutor” Sense on their specific home devices is even larger than 25%. And you are right, there should be tighter integration between the home electrics inventory one enters and the device info that gets filled in when a device is initially “detected”. FYI - the history is that the home device inventory came much later in time than the detected device form, and is there for a a slightly different reason (understanding the device population of the customer base - my end-user understanding) so there’s still some integration to do.

That said, I think your thoughts and intentions are right on, but you make the same tricky assumption many users make about how Sense does its job and what is technically possible to infer based on the just the power change during a transition. Watch the video I posted a couple times - that’s the Sense CEO explaining why they took the approach they did, especially unsupervised learning, vs supervised (using tagged examples as you suggest). And Sense do offer some pointers on how you can help.

Thanks @Kevin1. Despite trying to find more forum posts on this subject, I am finding some now from you and @RyanAtSense. I know that I am not just asking Sense to pick out a person in a room found in a home in a party with hundreds of other party-goers. I am looking for the ELEPHANT in the room. Everyone at the party knows who I am talking about when S/He talks–you can hear her/him from any other room. You KNOW when they start talking and you KNOW when they stop. (It’s not my analogy, it was Ryan’s.) And, even if they start slowly, the amplitude of their “voice” is significant and as a result, all the other conversations are drowned out. That’s the guy I want to start with for identification.

My apologies if I’m simply repeating what the last 100 new owners have requested, but perhaps Sense needs to hear this 101 times? If finding the elephant requires turning down the sampling frequency or the sensitivity–so be it. For nnyone who has read enough of the Sense articles, we GET that the leading and trailing edge of the power spike is not enough to identify a device, but toss a human in and perhaps we can get some synergy. Find me the top five SPIKES that last more than 30 seconds and we have something to talk about. My home? 1) EV, 2) Hot Tub, 3) Living Room space heater, 4) maybe the stove, and 5) The 1.1Kw consistently on “other devices”.

Ultimately, I’m hoping ML will dissect #5. But I can be patient. The first four? C’mon. Play dumb and out of the list of the top 20 domestic power suckers, GUESS! It’s OK to be wrong and the human can “suggest” the correct device (type) from that top-20 list. And most likely, it can even tell Sense the Make and Model #. Then it’s time for sense to “subtract” #1 from the party and move on to #2. and so on.

In my case, I consume just under 60 KWh per day and from the looks of things from Sense, my “floor” is about 27KWh/day for the “always on” stuff as I have yet to see my home under 1.1Kw. That leaves the remaining 55% for the “top-four”. It’s not trivial in $$, and when you’re dealing with 240v devices and decent-sized circuits (#3 is actually 110v), your power costs can get up there. Finding those devices with or without a human should be job #1 and those devices are part of my daily life so correctly identifying them and separating them should be accomplished in less than a week, if not a couple of days. Then, I can chill out and watch in awe as the ML slowly pulls the other devices out of the “other”. To get a “sense” of my frustration, SENSE seems to keep detecting what it believes to be #4. Problem is we can go days without turning on #4 (the stove). And coincidentally, it seems to detect this device whenever I am using my Tesla Mobile Charger. I guess I have to keep telling Sense “not on”. Wrong Elephant.


Hi @ATechGuy - the resources you read referenced @RyanAtSense, who has since moved to our product team, so I’ll be your contact in the Community moving forward.

The first thing worth watching through would be the most recent video we did with Data Science. It mentions quite a few things that shed more light on device detection and what we hope to do moving forward. This includes parts on your comment regarding stove detection and the role we hope home inventory will play in the future.

We released Dedicated Circuit Monitoring last year to enable instant detection for up to 2 120V or 240V circuits. While that doesn’t take all 5 of these off your list, it can allow you to focus on the two highest consumption devices. I know this doesn’t completely solve your issue, but between a couple smart plugs and dedicated circuit monitoring you could probably take a serious chunk out of your Other bubble. I reduced my Other bubble with 1 Kasa HS-300 smart strip by 250W immediately by putting it on my office/desk set-up.

No apology necessary, but maybe add a couple extra digits here :slight_smile:

We talk about this in the video above, towards the end regarding Progressive Device Detection. There are other releases planned for this year focused on helping users understand their energy usage.

In summary, we’re definitely aware of some of the obstacles for users re: device detection as it is today. The one thing that I think every Community member would agree on is that we’re constantly working to improve Sense. Behind-the-scenes, we have quite a few thing that I’m excited to share with you all once they’re set in stone :slight_smile: .

1 Like

Your “device monitoring” isn’t helpful for those of us using that circuit for solar (as originally intended). And it’s pretty clear that some of these other devices aren’t going to work with the Sense architecture.

What you need to do is to partner with someone to make an in-line (rather than plug-in) 240v monitor that works with Sense like the HS110, HS300, et al do. I’d gladly buy a few and get them wired into things like dryer, hot water heater, stove, heat pump, etc. That would allow Sense to accurately monitor all the “big users” in my home. Since Sense already has most of the software in place (all of it if the new device is HS- compatible), this should be a great solution.

1 Like

We’re aware of this gap for Solar users right now. Dedicated Circuit Monitoring was a step towards figuring out how we could approach this with our current hardware. We hope to do more with this, and our long-term vision is aligned with the understanding that future homes will more frequently have things like an EV Charger, Generator back-up and/or Solar packaged together in one home.

Our Product team is always looking for more third-party integrations though, so if a product comes to market that fits our integration requirements and can report energy from 240V loads, we’d definitely explore it.

1 Like

@ATechGuy, appreciate your discussion and really do feel your pain. I have 3 EVs (2 Tesla’s) plus some Sense-resistant (or inconsistent) 240V heating loads in my house, and for long while the detection of my 2 AC compressors was fairly dodgey. I have gotten a much better handle on most of these as Sense has improved AC detection (I give the Ecobee historic integration some credit), plus my addition of second Sense solely for DCM (I have Sense solar, so I had to commit an entire separate Sense unit to just 2 circuits of DCM) to monitor my floor heating subpanel plus the Model 3 on HPWC.

But I have also spent a fair amount of time poking my nose into trying to build my own EV detector, because it seems like it truly is the big elephant in the room. But when you get inside even that simple problem it is harder than it looks for a few reasons, especially when dealing with non-machine-learning algorithms.

  • the viewing window - searching for transitions requires a defined analysis window and some notion of a trigger delta. EV charging ramps don’t register as anything within Sense’s 1 second or so standard viewing window. I used a 15 min window for my (rough) detector, because I have that data window available to me. Our brains are “wired” to be more flexible so we can “see” 1 sec and 15 min elephants in the same go, but they have very different visibility to a machine, depending on window.

  • overlap - as your detection window widens, detection becomes more sensitive to overlapping transitions. Multiple on transitions and off transitions can get blended together. I had cases where two heaters turning on in the same 15 min period looked like the Model 3 charging.

  • differentiation between on and off patterns - I got my detector partially working, but then had to differentiate between the 3 and the S, because eventually detection has to match up ons with corresponding offs. Even with only two patterns, it was more difficult than it seems with a 15 min window, for several reasons. How do you tell a short charge (less than 15 min) apart between the two, or both ? What happens when you miss an on or off ? And the permutations go up exponentially.

  • EV charging profiles change over time - Sure, if I was Sense I could have build a more “sensitive” detector that used multiple narrower time windows and unique aspects of the Model 3 and S charging curves, but all of that gets thrown out the window when the charging profile changes, which seems to be every major software release for Tesla.

Based on my experiences, I sympathetic to Sense’s challenges, even for the detecting the elephants. If you want more details, here you go:

And here’s a test to highlight challenges, even for human detection of elephants.

If I ever get past the “needing approval” point, I would have edited my previous post on the SENSE not having the jack. I read further on the dedicated circuit monitoring page found it here in in step 8 I found my answer. So, no need to reply to that part, but Justin, please do still comment on MOVING the dedicated sensor around once the devices are “discovered”. Or, I think I get this, perhaps you don’t discover ANYTHING on those dedicated sensor, but simply report them as “Ded-1” and “Ded-2”. If that’s all you do, then I can easily expect others to chime in on “Why didn’t I just buy a dedicated circuit monitor like eGuage…?”

Hi Justin,

I super appreciate your comments. My SENSE is brand new. It reports version 35.0-f8afe9db, build 1531. I don’t recall my unit having the “center” wire/socket. Did I get shipped an older version? I like your Dedicated Circuit idea, especially if I could use that for 1-2 circuits, and then possibly move it to another circuit.

For me, I didn’t have a free 240 breaker, but I did have one that wasn’t using a lot of power and chose that one to tap onto. The next time I go into the breaker box, I can verify the SENSE, but presumably, if I HAD the option for the “middle wire”, then not only would there have been a jack, but also a wire included with my SENSE. Or, perhaps the socket is there, but it requires the “flex add-on sensors?” Certainly, if that’s all I need, please talk to us about moving around those add-on sensors once detection is made. One never likes opening the breaker panel often in any case, and for me, I have five 240 circuits (in use). (EV, Stove, Sub panel for my office area, Hot Tub, and breaker for an X-10 bridge, from which one half powers an Eaton UPS/power filtering circuit for servers and the other half is not used (other than for the bridge). My 240v dryer circuit is unused as I have 110v dyer (with gas).

1 Like

Great reply. I know I’m trying to trivialize this and i need to stop that. Of course, like any new user, I’m only trying to help improve the neural network.

Sticking with the elephants I understand that informing SENSE that there is an M3/AWD hung off a TMC is only useful if it can detect there might be an EV lurking on some circuit. That’s the part that is really hard for us noobs to comprehend. Instead, given the info I’ve provided I want SENSE to go elephant hunting! And not just blindly but it should pull out the M3/AWD/TMC signature info and adjust the scope appropriately. That’s all I’m asking. I know it still doesn’t make it easy but I would hope it could make it less difficult to find.

1 Like

That is correct, the HS-110 and its relatives don’t “discover” anything, and what they report to Sense is what you label the power they monitor. They do not teach Sense, for a number of reasons that are discussed throughout this thread. If you disconnect and move those smart outlets, the name you assigned goes with them unless you go thru a fairly difficult un-assign and re-assign to the new device process.

That said, they are VERY useful (if costly) for all those devices Sense can’t detect. I just wish they had a similar device for our large 240v consumers. That would be just like the HS- series but would need to be installed in the lines because so many large 240v devices do not plug in.


Can someone suggest the best method to respond to SENSE when it incorrectly identifies a device? Mine keeps identifying the “Stove” when I haven’t been near the stove in days.

Hi @ATechGuy, I wrote up the below response to this question in a related thread.

1 Like

Newbie question probably but I have had Sense for a couple of months now. I understand it takes time to discover devices but I have a couple of big ticket items it still hasn’t discovered. I charge my Tesla each evening at the same time and I also have a steam room. When I look at the meter it jumps up 12-13,000 watts when either are in use. Why would it take so long to identify? These are the only 2 items in my home with that kind of power requirement. Would love to be able to tell Sense what these items are. Overall, however, I am very happy with the product and it does give me great insight into my home.

1 Like

One issue with some big ticket items, like EV chargers, is that they ramp up and down slowly in a way that can be very different depending on model, charger type, charge limit, battery size, and even EV software version. Sense was originally designed to detect on and off of devices immediately (bubble pops up right at the start), but EV ramps are better detected after a few second to minutes. And just the size / magnitude of a ramp doesn’t tell what specific device it is - Sense has to sort through literally thousands of possibilities. Specific recognition is not as obvious as most people think.

More on coming enhancements in “Progressive Device Detection”, here:

1 Like

First and foremost, I am still very bullish on the POTENTIAL abilities of Sense, from the device, to the Data Science team, but I think these points need to get back to product management QUICKLY:

  1. Decide who your market is, and please, POST IT HERE, so we KNOW what you believe to be your market. If an existing customer is not aligned with that market, then we know we should accept that and walk away from these discussions.

  2. Decide if “really cool Data Science” is your goal, or if PRODUCT detection is. Of course, it would be very nice if both could be true, but your current product simply does NOT have the “crowd sourced” feedback loop for device identification in any usable state. To me this implies the FORMER is your goal and not the latter. I really think this is KEY. It will be a company-wide strategy that needs to be decided and executed. Personally, IMHO, the latter is what will attract new customers, and with more customers, you’ll eventually be able to do more with the former.

  3. I believe you have a very potentially helpful community and you’re not using them to the fullest. I’m sure I’m far from alone on saying I really want to see “Always On” eliminated. I want every item that makes up “Always On” to be listed! And, I want to method to divide “Other”. Do I care if that happens “automatically” NO! Would I find it very cool if it did, Of course!

  4. I’ll plainly state that I purchased Sense to determine the source of “Always On”. I will repeat that: I purchased Sense to determine which devices make up my 850w of “Always On”. I have owned sense for over two months and this collection of devices are still just that, a collection of unidentified devices.

  5. I did watch the 2021 state of Data Science video on YouTube and while I understand you have a VERY, VERY difficult task to identify appliances, I continue to feel you are missing your audience/market. Let me provide two examples where I strongly believe you need human intervention:

A. Manual device identification. Yes, it’s possible. And, today I ended up doing just that in my home. I shut off everything and then, one by one turned thigs back on. In some cases, I’ll admit the circuit would have powered a few devices, but I turned off well over 98% of the home for me to FIND THE ALWAYS ON devices!!!

B. In “Manual device identification”, Sense should start with a complete list of appliances and devices that are entered by the user. That gives the program a very, very small subset of devices from which to choose. No need to figure out WHICH EV I have, I AM TELLING YOU! Fridge, Over, Range, Yep. All there. Did you use any of that data? Nope.

C. START WITH THE BIG ITEMS–they mean the most to the home owner with respect to $$. Look at EVERY spike that happens with more than 1500w and you’ll know (in the US) that it’s a 220/240v device. How many did you have to choose from? Four in my home. Probably fewer than eight in 99.44% of the homes. So GUESS!!! You’ll be right 12.5% of the time in the worst case–then let the homeowner CORRECT you. In less than an hour, you should have ALL the 240v devices identified. As with your video, if ANY of these devices doesn’t match your database, then ASK THE HUMAN for more info. You might not get it, but it’s OK to ask! Then use the observed data for a week to really get to know these devices and understand how to isolate them as all the other devices turn on and off.

D. Next come the 10-15A (120v) spikes. Just think how few devices can cause this kind of load. Big motors and resistance heating. Let’s consult with our nearby human again! Whenever you get a semi “clean” spike you have seen a few times, ASK!!! Does the human reply quickly? Perhaps not, but ask the human if you can set a notification whenever that spike happens again. I know I would answer affirmatively! Since I already would have registered ALL my energy consuming devices, go ahead and take a guess when that 1400w spike happens next. Computer - Don’t be serious. Light bulb. C’mon. space heater, Possibly… Coffee machine? Microwave, toaster? OK, now we’re getting there! If the item isn’t on the list, I’ll ADD IT!!! Not in your database? No worries, ask for more info as time goes on. You get the idea.

  1. The bottom line is I really feel you need FAR more data, and better ML algorithms to auto detect devices that are the big consumers. And until you have that, you are going to frustrate users who could be your best friend for gathering that data. You also will generate poor reviews and ultimately kill your chances for developing those great ML algorithms when product sales decline as competitors enter the market with a SMART UI that uses the homeowner for identification. That kind of competitor need not concern itself with data science or smart engineers if 95% of all devices capable of consuming 1A or more are identified. Go ahead, run a poll if you don’t believe it, and ask your community if human-aided identification or “Cool Data Science to auto identify devices (that isn’t working)” is more important to spend R&D money on. I believe you know the answer. You don’t have time on your side.

I completely agree with ATechguy. While it’s cool to have ML algorithms figure things out for you, it’s just not feasible to compel your core audience to stay on board should something better come along that does incorporate the help of the community to advance the learning.

I commend Sense for wanting to stay to its core value of trying to get ML to do the heavy lifting, but I think we’ve all learned it’s just not as advanced to keep up with the demand of the community. Sense has an opportunity to cater to the core audience that help with ideas and learning opportunities to better develop what they’ve designed. It’ll come down to WHEN, not IF another company pulls resources from their community to help learning so end users can get the big picture idea of their energy usage. Because in the end, that’s what we want.