Refinement after device detection?

My clothes dryer has been successfully and correctly identified. But when my daughter dries her hair, Sense detects her blowdryer as the clothes dryer.

Curious if Sense keeps analyzing the devices that are already detected to help fix this sort of thing.

When she dries her hair and while the hair dryer is on, go to the dryer device page, click report a problem and then click this device is not on.

5 Likes

You’ll find more details here

And in particular here

https://blog.sense.com/articles/help-make-sense-smarter-device-is-not-on/

1 Like

Does that help specifically or just in a general way?

Both ways:

  • For your specific case it will help to weight the ML toward correct detection. Effects won’t be immediate.

  • In general, that weighting (your input) will contribute to other detection instances with similar signatures and potential matches for you and other users.

I think it has a pretty direct effect. It doesn’t usually correct itself immediately it after I’ve used “device not on” 3 times, it will get it right much more often.

Is there any technical roadmap for using crowdsourcing to tune your ML repository? For example, couldn’t you get a wealth of high-quality spectrum signals if you offered users a chance to capture a snapshot along with an indication of what they do in the electrical environment? For example, the user swipes out a spectrum region, pushes on the highlit range, chooses “Inform Sense”, types in “turned on and then turned off ceiling vent in upstairs bathroom” or “ground coffee in espresso grinder” etc.

There would still be noise - but with hundreds of thousands of quality reports coming in, wouldn’t Sense be orders of magnitude faster about recognizing the electrical world?

All of us would love to crowdsource our way to faster / better detection. There are ways to help, but manual entry of “on-windows” isn’t one of them. One constraint:

“Exactly how many noise-altered iterations of a device Sense needs to see varies, but it’s much more than you would be able to comfortably tag.”

https://blog.sense.com/can-i-train-sense

1 Like

I can’t speak to Sense’s motivations but I understand the challenges:

Adding to @kevin1’s explanation in the above I would extend thus:

The facial recognition analogy is a good one up to a point. A face has a bounding box in a similar way that you might think a Sense waveform has a bounding box for “Device X was on and off here”. As a human (in the facial analogy) you imagine a bunch of faces and you can pick them out visually and spatially (stereo vision and motion). Sense’s challenge though is more like, in my mind, imagining yourself doing facial recognition on movie where all the frames are superimposed over one another. In the extreme case, where things get effectively impossible … when time is also compressed … i.e. you convert a movie into a still frame. Where does my bounding box go? Tricky.

1 Like

As I understood it at the time of it’s inception, data gathered through the use of the Sense Smartplug integration could/may be used at some point in the future as well, although there’s challenges in that even - ensuring that customers are actually providing valid data (IE, Just a single appliance behind a smartplug, not a power bar with 3 different appliances plugged into it which would muddy the data) as well as getting users to input very specific data about each appliance including brands and model numbers so that the signature data being generated can be correlated to a specific device.

For example, I have our LG linear compressor fridge on an HS110 because linear compressors are very difficult for Sense to detect - unlike a traditional HVAC compressor, linear compressors can vary their speed up and down as required for cooling demand vs efficiency.

I have inputted all the data from my fridge into Sense so that some day perhaps the data my HS110 is collecting could be used to aid the ML algorithms.

1 Like