I know i am using the same title as the pinned post but that post was created and locked for future Q and A so i will post this myself for others to post in and so i can ask the question as well since i think your answer to the question isn’t correct. I’ve already introduced myself but i will post my background here so everyone can understand where i am coming from. I am a Chief Electronics Technician in the United States Coast Guard and have been an ET for the past 18.5 years with a background in LORAN, Radar technician/instructor, communications technician/instructor, fire control technician, 2M technician/instructor and MTR technician/instructor. With my background I understand why power and RF signals look different BUT within a specific known good range across multiple devices of the same make and model which is why it is possible to add a learning or training mode to the sense device. If you gave us as the end user the ability to turn on a learning or training mode and allowed us to tell sense what devices just turned on and off it would help sense learn the power fingerprint of each device that much quicker and would also give us the ability to train it with minimal devices on and also with multiple devices on.
I agree with this statement 100%. Sense should allow users to identify and help Sense train the device. I have two heaters in my house that use electricity; The Trane Furnace with a ECM blower and draft inducer, a Lenox pellet stove with an ignitor, draft inducer and blower motor. I also have a heated matters pad and an engine block heater. Sense has identified 7 heaters in my house, none of which I believe are correct. Therefore I agree with Mike Gessner
Now that sense has detected some heaters in your hone, you can help “train” it in the sense that you can identify the detections. I know it’s not training the way you or @mike_gessner are using the term.
I’d sense has detected something like heat 1 or heat 2, that doesn’t necessarily mean it is heat, it could be something else. It’s just likely to be a heat source of some kind.
There are heaters in refrigerators, dishwashers, dryers, clothes washers and many other devices that we don’t immediately think of as heaters.
I’ll make two comments here…
-
You are right about about the first part @mike_gessner, waveforms from the same make and model of a device, in the same mode/dimmer setting, etc. (very important here), should have similar waveforms within some set of parameters. But all the other devices in your house also play a role in identification when they generate background noise that is mixed into your device’s unique signatures.
-
Because of background noise and parametric variation in devices, you can’t just teach Sense what a best case waveform looks like, and hope for the best. The training used for this is much different than you imagine, because it uses machine learning. If you wanted to train it, you would have to expose Sense to thousands of normal (not circuit breaker) on and off cycles of a device under a broad range of conditions in your house.
The best things you can do to train Sense on specific devices today would be to:
-
goose up the number of cycles that an infrequent device runs. I used my thermostat to turn on my furnace fan for a minimum of 5 minutes every hour. That helps with circulation plus I believe it helped with faster recognition of the furnace/AC fan.
-
add smartplugs for devices that seem not to get picked up after a while. The supported smartplugs will give you the usage data you are looking for, plus the “Ground Truth” data from the smartplug can be used by Sense to speed training for that specific device type.
That’s not training and its just labeling what you know or think you know is a device. my sense has only detected two things in the past three days which come on quite a bit it seems and one of them i have no idea what it is and the other i am about 80% confident that its my daughters aquarium heater. I thought heat two was the electric tea kettle because it popped up twice when my wife was heating up water but i latter discovered that it was just a coincidence. it would be nice if they would integrate a training mode so i could turn it on and walk around and turn things on and off and it could ask me what just happened when it see’s these spikes in power and then again when it see’s something turn off so it can learn what a device looks like when it starts up and shuts off being that they show different power spikes and drops.
blue arrows show my furnace/AC fan, red arrow shows a spike when i plugged in my BMW i3 when i got home (set to charge between 1am and 6am super off-peak hours) I would have to disagree to a degree with you on number two. I have my ecobee3 set to 72F and it cycles on and off quite frequently because of the cold weather right now so my furnace/AC fan is running at least once an hour as you can see from the picture. as things are running the line noise isn’t as distinguishable but as you turn things on and off the positive and negative going cycle is where you can best see its distinct power signatures. If sense had the option to do a training mode that would ask you what an event is then it should help it learn faster. For instance when i was teaching my phone what pictures had who’s face in it to sort pictures by person. the more pictures i confirmed of different people the more accurate the phone got to sorting photos.
Did you turn on the notification for what your previous 24 always on usage is somewhere?
I’ve never seen that and it appears we use the same device.
@kevin1
In point 1 above, you were talking about similar or same models of products and their waveforms.
Something I just had happen was a replaced my water heater a month ago. It was defective and I replaced it again last week. The first was detected by sense and tagged properly every time. The replacement was the exact same brand and model but when I hooked it up, it was not noticed as the detected water heater. I ended up deleting and letting it detect it again, it did so within 2 days.
What I noticed was the first water heater had a draw of 4640 watts while the replacement is just over 4700 watts. My question is what the threshold is before sense won’t seee two identical devices as the same? These two water heaters are within 3% of each other.
I have a space heater and stove element that are further off percentage wise than that and they get mixed up (1500 vs 1650).
my water heater is 100% gas and doesn’t have power going to it at all and so far my sense has detected heat 1 and heat 2 and i have determined that heat 1 is the aquarium heater but i am still trying to figure out what heat 2 is now that i have determined that it isn’t the electric tea kettle. I am thinking that heat 2 could possibly be my owl security camera system with the amount of times its detected it. could be the security system recording video to the HDD but again this is just a guess. I would say its a good thing that your sense didn’t associate your new water heater to the old one because they would have different power signatures and if you had two identical water heaters you would want your sense to be able to distinguish between the two.
I was just able to identify a detection and associate it with the correct device. I had a “heat 4” detection a couple weeks ago. I’ve been watching for it to turn on or off and checking th usage history daily. There was only one day and a single use of history for the device with a fraction of a KWH. I debated deleting it many times until my son decided to make rice tonight.
I’m surprised the history showed it coming on only one time because when he used it tonight, that thing cycled at least thirty times.
@mike_gessner,
Glad you used the face analogy, because you just skipped the biggest part of the process that you never helped train, but many thousands have - identifying human faces. It seem simple when the bounding box for most faces just magically pops up. But that simple development took many years with hundreds of people working on it, along with hundreds of thousands of labeled images. More info on facial and other recognition history here:
Once the software knows how to find a face under all conditions, the next step, classifying faces by name can be trained. Sense does something similar - once it finds a device pattern it gives you crowd-sourced options for what that device might be based on how people have labeled devices with similar patterns.
There are two more issues with human training today. The first is that Sense only looks at patterns of a few sizes today. If you read through the technical blogs, you’ll see that most of the recognition is based on less than 1/2 second of samples. A few device detectors do look at longer periods, like the detectors used for EVs. So unless you are extraordinarily fast, and I mean like the Flash, there’s no way you could accurately label the waveforms that Sense is currently looking for.
Second, the key is consistency and volume - labeling hundreds or even thousands of different waveforms from the same device under slightly different house conditions. The guys doing facial recognition used hundreds of people to manually label images via Amazon’s mechanical Turk service. You really wouldn’t want to do that thankless task.
This is my past 24 hours screenshot taken from my iPhone. the first green arrow shows where my car was charging early in the morning and the last two green arrows show when i plugged it in but it didn’t start charging and turned off the charger immediately. the two red arrows show when my drier was running. i have an LG true steam drier that senses the dryness level of the laundry and it ran longer with a larger load of laundry during that first event and much shorter cycle with a much smaller load on the second event.
yes i understand what your saying but i was just using that as an example even though they are looking for different things BUT i can tell you that what sense is measuring is just an AC waveform which will look different from device to device because of different components inside each device i.e. motor, heating element, capacitors, resistors, transformers and so on and within a single device you can have a mixture of all or some of these devices which will change the characteristics of the power usage of each one. As I’ve said before i understand how this works and how fast this happens being that teach this stuff and i currently teach a Navy class for MTR which analyzes power signatures within systems and all the way down to component level to detect trends and isolate faulty components within a circuit. Measuring something down to the half second isn’t that hard and when i worked in LORAN we had to adjust our signal down to the femtosecond. when everything is running the different power signatures are there and can be distinguished and again i will say that sense could learn to distinguish between different devices quicker if we were given the ability to tell it what certain events were if we already know for a fact that a specific even was linked to a specific device.
Maybe you should get a job at Sense then?
With all do respect, what you did in the Coast Guarg was likely with special gear on specific equipment. Knowing what you know, how do you expect a little generic device to do the same on a much broader scale?
I hear your frustration, thank you for your service 
I’m thinking the same thing as @samheidie. I’m interested in the work you do as an example of automated identification. How do you detect failure trends from signature analysis of signals in the systems you deal with ? How is the software trained to recognize failure events in specific components ?
Hearing the techniques you are familiar with and you think highlight how training works well today, might help me compare and contrast. Appreciate your service as well.
BTW - Here’s what a day at my house looks like… One can tell when one of the EVs was charging, but not much else…
Hey Sam,
How are you? I’ve been having all these heaters show up all winter long. I think I am on heater number 8 up here in upstate Albany,
Hey there @mark.dechiro
I’ve only had sense installed since January 15th and have had a long list of heaters.
Some correctly identified and some seem to be multiple devices.
I’ve got 3 heat devices now that all trigger the stovetop(several burners), dryer, washing machine and dishwasher randomly.
Whenever I get a notification when away from home I get excited thinking the house will be clean and laundry and dishes done when I get home. No such luck yet.
@kevin1
May I ask what the yellow overlay is?
I’m assuming it’s solar by the time of day and the waveform produced.
You’re right… That’s Sense Solar in action. We had a mixture of sun and rain/clouds today in the Bay Area.
Something important that comes to mind for me every single time I see one of these threads about “training” (versus learning) with Sense is…
If it was really that easy, do you not think the Sense engineers would have instituted it long ago?
And secondly…
No other competing product offers said functionality?
Accordingly, I always come to the same conclusion - despite sounding easy to the layperson, it’s really not.