Seems like I lose winter devices every year, Furnace falls back to other being the common one. Now it’s re-found the central vac that would not go a week w/o use. And it had found our washer, but doesn’t see it anymore, and our dryer just got re-found. I am NOT working on our panel, but I did change the router in our house when I got 10Gig fiber. I have put perhaps the max number of KASA devices around our house to try to have it KNOW what power is being used, but this seems to only confuse the issue more. I’ll have it see KASA Dryer and the detected Dryer about 50% of the time.
At least we are not having all the connectivity issues we were having a few months ago, every week we’d have long outages. OK, that’s my rant
You’re raising a good question in there - why does Sense re-find already discovered devices as new devices ? I can’t claim to know the exact answer, but I do have my theories based on some knowledge of machine learning. Two thoughts:
Concept drift - when models change slightly over time, but then suddenly stop detecting the original source of the model.
Data drift - when the usage subtly change over time due to seasonal variation and other factors.
Both of these come into play with devices that have periods of non-use, especially when Sense is regularly using incoming data to update models. Long periods without incoming data from the original device might push Sense to use close data from other devices to update the model. And data from a seasonal device can certainly change between years. I saw that with my AC units a few years ago.
I literally just opened a support ticket on this topic. It seems that every “season” when I turn on existing devices for the first time, detection is very wonky (AC for the first time in the spring and it can only find one leg, dehumidifier I haven’t used all winter. It will almost never use the same remembered device
The recommendation from support was that there was a device detection update released last October and it was based on a survey of devices…I would need to reset my sense to take advantage of it. I have 4 years of data. I can export it, but would not be able to visualize it as well, and it’s annoying that I would not be able to see my 2024 usage from January…
You think it’s worth me starting over? I would like to be using the latest AI/ML models, so if anyone was suggested this, did you notice an improvement? Everything seems to be
I reset in Dec. I did have a couple fluorescent lights detected that had never been detected before. On the downside I’ve had quite a few issues. While I can’t say they are a result of an updated model, the reset caused them. @JamesDrewAtSense has been great at looking into, acknowledging, documenting and seeking resolve. Getting the backend to fix anything can be hit or miss. I’ve learned that Sense backend does have some built in checks and self resolves for some unexpected things. The engineers are pretty good about trying to resolve system wide bugs, but when things involve one-off issues it’s a wait and see that could take many months. Detection related issues on the ML side of things there’s not much to be done other than a reset, if the standard delete & redetect doesn’t work. Even after a reset it’s a guessing game as to what will be detected, some are usually quick to detect, some may never be detected again, some new devices may be detected. Detection is like box of chocolates.
IMO, I really do think a individual users data can get garbled up on the backend and in a way wrongly sway how the ML interprets things. I’ve had devices stuck in a state of Limbo, either not fully detected or not fully deleted. Self checks are supposed to correct things like that, but I’m somewhat convinced the checks makes them disappear from the UI and counting towards any total, but doesn’t actually remove the data. Which is one of the reasons why I’m thinking about another reset very soon.
Because I heavily depend on integrations for detection (probably safe to say I’m in the top couple users for number of Kasa plugs) resetting isn’t as big of a deal to me as it would be for others who depend more on native detection.
The “survey of devices” doesn’t have any effect whatsoever on ‘native detection’, instead it creates a couple new “estimated devices” based on your survey answers. For example, based on my answer I have a ‘phantom’ device called “Fridges” “5 units” that use an average 128w while on, which is somehow based on the “average fridge uses 180w wile on”. There is no power meter or bubbles for these devices. I also have a phantom device called Washer with no data and a phantom device called dehumidifier with no data along with a couple others. I’m really not sure why fridges have an estimate but not the other things. All fridges, washer, dehumidifier are on smart plugs but Sense does not use smart plug data for these estimated phantom devices. From what I understand these estimated devices turn into actual devices if they are detected. In the meantime the estimated usage reported for an estimated device is recalculated daily and subtracted from “other”, but are not supposed to have any impact on Always on. I can’t say for certain (haven’t had the time to really dig into it) if having these devices on smart plugs results in a double count with the estimated devices. @BeccaAtSense could probably clarify.
Thanks so much for the insight! This was a much better answer than the generic response I received from support. I really thought I’d get more details on the new detection method, but it was very generic.
Because the majority of my devices are fine, I am not going to reset. Kind of disappointed in support taking the lazy way out and telling me to “just start over” when last year someone went above and beyond to take a look at my device models and corrected them.
While I would like to be on the “latest and greatest” detection model or whatever it was that was improved, I feel like I should just stay where I am. If my monitor ever breaks and I have to replace the hardware, I’ll start over.
From my understanding (though I could be entirely wrong) any updates to device detection would be system wide, an update to the ML algorithm of sorts that would effect all users the same.
This is the reason why users can’t ‘train’ sense as it would require customized algorithms for each user, a LOT more ‘real time’ processing power not to mention expense and highly skilled human intervention when something goes astray.
One system processes raw data for all users, another that customizes the processed data for retention and the UI experience.
Hence why detection updates are such a rarity and huge deal as even the smallest of changes could have a massive and potentially catastrophic effect on the ML aspect for all users in addition to corrupting the retained processed data.