Since we have the ability to see in real time in the powermeter graph when we turn devices on and off. Why not give users the ability to “tag” a waveform? And Sense in the back end can use that tag to improve the device matching DL/ML models. It will have a 90% success rate. It already tells you the “delta” in wattage - if there is a way to tag that delta… we’d be set!!
A few things to think about:
- Humans don’t have the combination of timing accuracy and repetitive consistency to tag half second swatches of note sufficient for machine learning. The only thing worse than no feedback is bad feedback (even slightly off).
- I do thing the tags on spikes represent at least some significant patterns to Sense. But there are at least two problems associated with tagging them with device names, even if they are significant patterns.
- What good is tagging a lone pattern of Sense hasn’t made the connection between multiple occurrences under many different conditions ? The device tag doesn’t help Sense make associations between multiple on or off events. Maybe you could tag a a few dozen to couple hundred and Sense would have enough association to work with.
- There’s also the issue of associating on-events with off-events. Probably not useful tagging one without the other.
- From what I have seen, not all events are identified by Sense. We don’t see all the spikes for all the detected devices turning on or off. I haven’t look closely, but if we can only label what is flagged, we’re probably going to miss many transitions.
3 Likes
Another way to ponder this is:
Consider smartplugs.
You tell Sense “this is my toaster on a smartplug”
Wonderful: 100% detection … AND Sense can learn from all that known-to-be-a toaster-and-only-a-toaster data. Even better. It can correlate the on/off in the high-res samples it’s doing (vs what the smartplug is feeding Sense) and really, potentially, get to know your toaster.
Now you unplug the toaster from the smarplug and plug it back in without the smartplug.
That’s the lazy and much more reliable way of doing what you suggest.
And still there is no real guarantee (certainly not 90% at this stage) that the toaster will be recognized each and every time it gets used.
2 Likes
Aah - this is something I can get on board with. Yes its a bunch of work on my part. But its definitely something that can help me. So some ground rule verifications here:
- There can be only one device per smart plug. Do all smart plugs work Wemo, Amazon etc.
- How and where do you tell Sense that its a smartplug on deviceX?
- Once you remove it - will it remember? 90% of the time?
Thanks
Three good links that explain all:
https://help.sense.com/hc/en-us/articles/360012089393-What-smart-plugs-are-compatible-with-Sense-
I do think there’s some utility in tagging a waveform in the Power Meter for historical knowledge, but it wouldn’t be a major help in Sense’s learning process, as noted above.
This thread is worth a read: Why can't you train Sense?
@kevin1 I did install a WEMO plug but its not the specific “insight plug” - so am I SOL? I also have the amazon’s smart plug - guessing that wont work either. Since I did try it with the WEMO plug - byt the my-devices page has a “looking for plugs” but nothing shows up. its been about 10hours now…
And No NEW devices1
Thanks @RyanAtSense I will get to it…
Sorry, but you are SOL if you don’t own the specific WEMO Insight smart plug or TP-Link smart plugs specified. Those are the ones that actually measure power. Most brands and models of smartplugs do not measure power usage.
I’m partial to the TP-Link smart plugs for a few reasons - better App, they default to On when the power goes out and comes back on (don’t use Wemos on freezers), plus they offer both a single plug, plus the HS300 outlet strip. The HS110 usually runs around 22$, but the price varies online based on inventories.
https://www.walmart.com/ip/TP-Link-HS110-Smart-Plug-with-Energy-Monitoring-1-Pack/48695645?wmlspartner=wlpa&selectedSellerId=1631&adid=22222222227036250118&wl0=&wl1=g&wl2=c&wl3=57332447378&wl4=aud-534311367800:pla-89655176018&wl5=9031914&wl6=&wl7=&wl8=&wl9=pla&wl10=113509192&wl11=online&wl12=48695645&veh=sem&gclid=CjwKCAjwyqTqBRAyEiwA8K_4O7LEVe7zXGVnv0NO0_9feSqbF1_aQyPwQdZsPqeTUj4N3RwzYkEhQBoCZdoQAvD_BwE
Krishnan’s got the right idea. From what I’m reading you keep saying we can’t provide input because it’s all about the Sense algorithm. You don’t have to include my input in your algorithm if you don’t want it, but it helps me. I bought sense so that I can analyze my power consumption and make decisions, improving your algorithm is a bonus for both of us so I’m willing to give you access to my data. The ability to tag spikes helps me as a homeowner. I should be able to click on a 1000 watt spike tag right after I turn on my angle grinder and select a device from a list to identify it. Then run some reports to see really how often I use it. It’s some work on my end but that’s where we’re at right now until we start feeding device ID back onto the wire. It’s hard to tell the difference in the signals made by an angle grinder and a vacuum cleaner but if I tell Sense that every Saturday I run my angle grinder a bunch of times between the hours of 10am and 3PM it will learn the difference between this device and the vacuum on the same day. Regardless of what your algorithm learns, I did not buy the product to grow your algorithm, I bought it for my insight.
Please add two features:
1- Flag a time range on a graph and add a note so I can refer back to it later.
2- Select a demand spike tag so I can associate it with a device and add a note.
@user1,
The traditional Sense “transition detection” looks for very specific short transitions of interest (less than 1 sec transitions) and characterizes them with 20 or so parameters. It doesn’t begin to understand them as on/off until 3 things happen:
- It sees enough transitions with similar parameters to determine that there is a “cluster”
- The cluster passes a distinctiveness test - it has to be a cluster without intersecting neighbors, otherwise it really isn’t a cluster.
- It is able to find a matching cluster of off transitions, because finding an on without an off is useless, and vice versa.
I mention all this for two reasons. First it’s better to understand how Sense currently works when speculating on the creation of new features and how they might work. Today, there’s no value to “flagging” a waveform feature that Sense doesn’t already understand. And there are reasons that flagging a few spikes is useless for teaching Sense. Plus, right now, you can’t count on Sense to include time of day, day of week, week of month, in the prediction. Second, because Sense is working with broader ways to natively detect a broader swath of transitions.
1 Like