Has the Sense team ever considered


#1

Allowing users to input major appliance info at install as an option? I would happily put in info about all my major appliances if that would aid detection in any way. I have 2 Refrigerators and 2 deep freezers, 2 separate ovens, 2 separate microwaves a well pump and a sump pump along with the other host of electronics in a modern house.

I understand the AI is figuring it out, but giving it a cheat sheet might speed device detection along.


#2

I provided all of that information to support at their request as part of a support case about poor auto discovery but nothing ever came of it, and it never went anywhere. It doesn’t seem that it does them any good despite requesting that I provide it.


#3

Not super useful to machine learning until it “learns” to find the signatures of each device. Kind of like working with a photo program that identifies faces. It doesn’t help to give the program the names of all your friends that it will find in your entire photo collection. You have to wait until it finds the faces.


#4

Facial recognition systems do get information about the names, addresses, etc for people passing thru specific airports. Knowing that I’ve been thru 23 airports (and when), but never have been in the thousands and thousands of others, helps those systems narrow down the facial signatures.

It seems to me that knowing a friend was blonde, wears glasses, and would be found in photos in 500 houses (but not the other 50,000) would provide valuable data to the algorithm generation process. Or, knowing that hundreds and hundreds of homes have a specific make and model oven should help Sense determine the pattern for that oven more quickly. MUCH bigger data sample and far more events to work against.

As others have noted, Sense Support (and engineering) did ask me about specific devices manufacturers and model numbers. That was part of how they determined that my “noisy” device adversely affecting all device detection in my home was the Franklin Electric SubDrive75 variable frequency 3 phase synchronous motor in my constant pressure geothermal deep well pump. So, now they know what one looks like, even though they haven’t been able to teach Sense how to deal with it.


#5

I guess we can just loop back to this…

The list doesn’t help with machine learning, but a survey of the user base could help long term targeting of models:

“While such a survey would provide us with some useful information, and we’re looking for ways to gather this information with your help, it would benefit detection in the long term and not result in immediate detection benefits.”

@jacksons, @don1,
Knowing make and model of specific device might, or might not help with recognition when Sense is debugging a detection issue. It might not help when the device is so far outside the current models they have developed, or is a device that is resistant to detection (i.e. plasma TVs power consumption is big and highly dependent on what is playing on them)


#6

My only point in even raising it is tied to the terrible auto discovery performance I’ve experienced from the start. I opened another support case in early October and support said they could see records of my main refrigerator and many other major appliances that had never been discovered and they couldn’t understand why they were never ‘found’. They expressly asked me for all the model numbers of the devices but nothing ever happened with it. I provided all of that back in early October and the support ticket seems to be orphaned with no updates since late October so it seems I just have to lower my exceptions on Auto Discovery and accept this is what it is.


#7

@kevin1 is pretty spot on with what he’s saying, so I won’t totally repeat. In short, yes it’s been considered and it’s a good idea because it would help to train models on our end (and it can help troubleshoot specific detection issues). In that manner, it’s something we plan to implement in the future. However, it would be less useful in the form of some immediate training feature. We wish it was that simple.

This has also been discussed extensively here, I would encourage you to make your opinions known in that thread since it’s the actual “feature request” thread: Allow us to provide a list of devices before they're detected


#8

If the analysts have a database and can correlate user provided data with signals that the AI is having difficulty discovering, how is having more information and datapoints a bad thing?


#9

To further the idea, once I had put in my major appliances, have a " I just turned it on" button.

Put in a load of wash, press start on the machine and click in the app that you know for a fact that you just turned on that device that you had input, Move the clothes to the dryer, press start select your entered dryer device info and click that you just turned it on. The device in question wouldn’t be the only load, but after doing that a few times the AI should have a much better “idea” as to which exact portion of the noise signature is that device, or maybe I am oversimplifying it…

Just an idea and it seems like it would build the known database of signature to device correlations.


#10

You have to understand the time element here. Sense looks at samples a million times a second and many of the signatures it identifies last less than a half second. If you could consistently hit the “I just started this” button within a half second of the actual signature, that kind of training data might work. But hitting it randomly within a 4 second window, which is likely the response window of a casual user, probably would be worse than helpful, actually creating bad learning. Equivalent of you putting a bounding box up and to the right of a face, when training machine learning on face recognition.


#11

Kevin linked this above. It’s really worth a read: Why can't you train Sense?

More datapoints are a good thing if they’re spot-on accurate. Otherwise, they can become a bad thing pretty quickly.


#12

I’ve had Sense 1 year. Have many water pumps and led lights going. 6500kw undiscovered always on. Some pumps small aquarium pumps, sometimes 500w water heater,
Have 4 zero to three hp sequence pumps including two sta-rite, one sequence pool pump, one performance pro. All sequence manufacturing. They have variable speed control on motor back and computerized. Motors ( I think) have 48-52-56?? Permanent magnets and controller converts 220v 1 phase to a different variant for speed control at high current. 3 phase? Pulse width modulation? They all are working perfect. They have Knick name "dial a flow " Here is the question based on forum questions:
Will sense just give always on for all as pump signatures are so similar? Like the guy with all the refrigerator/ freezers. I think auto detection is good for a light. One heater, individual comonents. Computer signature compared to data base. But it seems the mitors I use are minimal in use and never get a database like a tv. Too bad as mentioned before we ant force sense to learn a signature, and it will reorganize and send data to sense computers.