Ongoing device detection issues

I know exactly what the item is BUT it won’t let me name it

I’ve completely given up. Clearly this is a product that has promised something that just is not feasible. I understand everyone saying that it is an extremely challenging feat and that they are doing their best, and I’m sure they are, but the fact of the matter is that the product is advertised of being capable to doing such things.

On a note about support, I have a TV that is supposedly detecting via Network Identification. It has recently started showing as on when it is not. I submitted a support ticket and their response was that it was in fact on when Sense said it was. I was literally staring at the TV as it was off and Sense said it was. Apparently their support is much better than me at looking at a TV and determining if it is on or not.

1 Like

I would encourage the use of more of the TPLINK devices => They eliminate the noise from the neural network training required by Sense’s algorithms. Once I added about 6 of them (particularly in places where power strips are plugged into generic devices), Sense started detecting the “heavy hitters” around the rest of my house very quickly. My Nissan Leaf and my Dryer (both 220v) did not get detected until after the TPLinks were installed and running for a couple months. I’m pretty convinced that went a long way towards making the machine learning logic work better for Sense.

1 Like

Is there a way to detect a device by taking a smart plug around and tell sense to snapshot that device when plugged in to teach it? Then move the smart plug to another device?

1 Like

At this time, no, but that is the eventual goal. We talk a bit more about the integration here:

Nope, there isn’t sorry. Among other reasons is the fact that sense and Smart plugs use totally different power monitoring techniques.

That’s a good question, but there are several reasons it can’t quite work like that.

  • As @andy suggests, smartplugs only provide data on roughly once per second basis, not the million samples / sec that the Sense monitor does its work at.
  • A few “snapshots” isn’t nearly enough patterns for doing machine learning training, even if the Smartplug data was similar to the directly collected Sense data. Hundreds to thousands of on-off cycles under a variety of conditions are required.

Thanks for the suggestion, but I don’t really need smart plugs. I’ve only got 4-5 things plugged into an outlet 24/7. router, modem, tv, apple tv, desktop.

I’m 2+ years in and except for looking at the solar I get no value as the devices it’s detected don’t make sense. The TV I replaced, still detected one and awhile. Some work lights, sometimes sometimes not. Garage door opener- not any more. I get how hard it could be to detect devices-the algorithms involved. But my guess it won’t get better as the source of data isn’t great. Perhaps with totally new data analytics team there might be a different approach taken but other than that…

I agree with you and relate to your pain with Sense detection.
This would be a great product if the owner/user had some control over detection.
It reminds me of the old style server based systems that only IT people could work with.
For me, it’s a very accurate power meter. That’s it.

I’m 8 months into using sense, and I’d say I gave up after the 3rd toaster was detected. I mainly use it as a shaming device to remind me how much energy we waste. There aren’t many things Sense has detected, and besides water pump and fridge many devices show up as on although they are not.

Until they can figure out how to incorporate user on/off device feedback into their ML model nothing will improve. Maybe someone like google buys Sense.

That seems to be a common theme, but the power company provides very accurate metering and in many areas they are even accessible from the web. Up here in the wilds of NH, we’re more primitive, so some of us purchased alternative systems (we have welserver), some quite sophisticated.

I bought Sense (and experienced quite a bit of install, then re-install cost) to provide detailed consumption device by device, room by room (we have over 145 “devices”), so I could tune our power usage. I already had systems in place to monitor both overall consumption and solar, including detailed geothermal HVAC systems on/off and temperature monitoring.

Unfortunately, Sense can’t handle our environment, because its detection is seriously distracted by one of our largest power users, our deep well variable frequency constant pressure pump.

I am also frustrated. I would rather turn a specific device on/off manually multiple times while Sense learns which one it is. If only to make sure that is getting the right device rather than sensing a new device. I know there is a long winded article about this but as a programmer myself, I just can’t help but think there is a way to make it work. A human feeding “clues” to the ML algorithm should work. Even for more complicated devices that have multiple motors like an espresso machine (grinder, pump, etc…) or a multi-stage heat pump, I would like Sense to know whether it’s on or off. Perhaps Sense should have a mobile team that comes to customer homes with specialized equipment to help learn about various devices and build a large database of them to feed the ML algorithm.

Can understand your frustration. You should probably take a couple intro courses into machine learning and read the Sense blogs to understand why human-supplied training data (not hints) wouldn’t be suitable for their short-window detections. I do like the idea of a mobile lab for investigating devices that present difficulties, but to some degree, smartplugs do something very similar (accurate, consistent feedback). And a mobile lab would have some of the same difficulties we have - no easy to use 240V smartplugs, plus wired 120V devices where it is difficult to separate the two supply wires (putting a current clamp around both wires cancels everything out)

I have read the blog posts (although not recently) and have a basic understanding of ML. The Sense algorithm is monitoring electrical signatures from a noisy network of devices over a long duration. Each device has different modes (startup, power down and multiple stages during use) which are “ML features” used to predict a unique device (ML label). Those unique devices are matched against a database of devices. The HITL (human in the loop) can help the ML algorithm know which unique patterns match which device. For example, my Sense device list is “Always On”, “Motor 4”, “Other”, “Coffee grinder”, “Coffee Maker”, “Coffee Maker” (yes, a duplicate!), “Coffee Maker 2”, “Dehumidifier”, “Dishwasher”, “Dryer”, “Espresso Machine”, “Fridge”, “Furnace”, “Garage door”, “GE Clothes Washer”, “Greenhouse”, “Hair dryer”, “Heat 1”, “Heat pump”, “Oven”, “Oven”, “Rice cooker”, “Stove”, “Stove 3”, “Toaster oven”, “Vacuum”, “Vacuum 2”, “Water dispenser”, “Water heater”, “Water heater” (duplicate), “Water heater (?)”. But what Sense doesn’t know is that I have only one coffee maker (Delonghi Espresso superautomatic machine), I have two furnaces each paired with a heat pump, etc… I also don’t know which device “Stove” or “Stove 3”. I have a heat pump water heater and also an Insinkerator Instant Hot water heater. Which one is it referring to? I only have one Oven but I also have a toaster oven. I have two garage door openers but it lists only one. Sense gets an “D” as a letter grade. The only thing it’s useful for is daily consumption/solar production and even there Sense can’t seem to put BOTH numbers on the same screen. Nor can they add a start/end date for reports. I’m not impressed with their management team so far. They can and should do better. Perhaps, they need to read some books.

1 Like

Good overview, though there are a few technical points that you oversimplified in the following statement, that might lead you to some incorrect assumptions:

  • There’s not really a database per se. Machine learning is stored as hundreds to thousands of parameters, none of which has any deterministic meaning to the humans outside.
  • Once Sense has “clustered” / detected a particular repeated pattern, it becomes namable and classifiable. That naming and classification data does roll back into Sense crowdsourcing and logistic regression.
  • Much of the Sense pattern detection is based on a very short sub-second analysis window, not really amenable for human training, because humans just aren’t able to deliver the timing accuracy and repetition.
  • There probably is a place for using some human data for constrained training (I only have two fridges in my house so don’t find more than two), but that would likely delay detection of devices longer than the current wait, because Sense would have to use the “two many fridges” as feedback once it detected the third. It’s very hard to inject this kind of end-result logic for training into the inner loop of ML training - you can only really use the logic once the ML network is churning out final results.

What the Sense world really really needs is a 240v wired in smart plug equivalent. That would help tons for all those devices Sense isn’t smart enough to handle.


After a year or so, I too find myself in the giving-up frame. I have several Heat 1, 2, 3, Vacuum 1, 2, 3 devices that I have no idea what they are. I’ve tried to isolate them, and in the end have just spent so much time, it’s totally in the realm of diminished return. That said, this is not what gripes me the most. no, that spot is reserved for the items Sense has found only to loose at a later date.

My NurtiNinja blender was identified and accurately reported on for months, and then suddenly stopped showing up. It was, however, replaced by a NEW device, Vacuum X. My pool pump is a bit different. It runs 7 hours per day and was found rather easily and accurately by Sense, but then after weeks of consistent reporting, my Pool Pump device started showing up for only an hour and a half of operation each day. It shows on at the correct time, but even though the pump still runs for hours longer, Sense reports it as turned off shortly after it starts. Seriously WTF?

What am I to believe? What’s accurate and what’s not? The only thing Sense seems to do reliably is track my smart plugs, but who needs Sense for that task? And don’t even get me started on the machine learning excuse for not even bothering to consider other data points that could be given my we humans. The fact of the matter is that the folk at sense just don’t want to tackle the difficulty of overlaying additional data points with their ML algorithms.

Anyway, the whole Sense experience is at best frustrating.

1 Like

For the devices you do t know about, stove 1 and stove 2, they may be the same appliance. My Frigidaire gallery range with convection oven is detected by sense. I currently have stove through stove 7. They are merged into “range”. Each burner is detected separately (5 burners) and the oven broiler element and the oven bottom elements. The only thing not detected is the light and the convection fan.
Maybe both your stove devices are separate burners.

I’m still within my return window (Sept 3) and I’m seriously weighing whether I should keep it or not. I get pretty accurate (though next day) updates on how much power I used from my power company and the solar is nice but I can get that from my solar application.

What I was really hoping Sense could help me with is my demand during the high times of our Solar Generation Plan for SRP. Demand charges for 4.5kWh are $55 but if I hit 7kWh it jumps to $99. We try to keep our usage at the 4.5 via the load controller but that means I generally can’t run my AC upstairs after solar dips to 3kWh because it pulls too much with the other stuff that is on. My Always On is at almost 2.5kWh right now (it was 800-900W a couple weeks ago) so there appears to be something that has changed but I have zero knowledge of what that might be since it’s just a Min over a given time period. Unplugging and plugging in things won’t help me at all to determine anything.

So far it has identified a Fridge (1 of 2), a garage door (1 of 2), a hair dryer, a microwave that are reliable. It’s identified 2 ACs that are not reliable (we have 2 units), the dryer (not reliable) and a Heat 2 which I don’t know what it is yet. Also I have one HS200 that is plugged into an instant hot water but it’s not “tracking” that itself as it seems it gets that info from that device which I have access to already. It seems like the only answer is to go buy hundreds of dollars of those plugs and plug everything into those (which itself adds energy consumption) then Sense can tell me what is using what by asking the plugs but that just isn’t feasible.

I have another month before I need to determine if it’s worth keeping or not but with the flood of advertising I saw I was hoping it would be a little less wet behind the ears than it was. I’m not sure how people have had this for 2 years and not regretted it unless they had no visibility into their power consumption at all.

1 Like