We’re thrilled to kick off something exciting with one of our longtime Sense users, Kevin Kranen, @kevin1. There are few people, if any, as capable of getting the most information and insight out of their Sense monitor as Kevin. He’s been a tremendous help to us and you, so we asked him if he’d go a step further to provide you with valuable insights, tips, and stories that will enhance your experience with Sense.
Kevin has graciously agreed to write a blog series of informative and engaging articles that dive deep into various aspects of Sense. His first article below is entitled “What to Expect When You’re Expecting (Detections).”
Have a read and let us know what you think. Do you have any additional questions after reading? Which explanations could use more details and elaboration? We want this to be an interactive experience. Kevin will be available to join the conversation and answer your questions.
We want to thank him for his continued support, of Sense and the Sense Community!
Here’s an excerpt from the post:
“We all do better when we know what to expect from a new technology. Sense can be a helpful tool for everyone, but every home is unique and device identification may vary depending on a variety of factors, which I will go into detail about here. Sense users should be aware up front that there are some types of home devices that Sense will intrinsically be good at learning and subsequently detecting, but other types of devices are far more challenging for Sense to learn using what I refer to as "native detection”, Sense’s main artificial intelligence (AI) approach to identifying devices. But that’s not the limit of devices Sense can “detect” - Sense offers a whole range of additional capabilities and integrations that enable users to “detect” many devices outside the ones that are “learned” by the current Sense AI approach. When set-up and managed with a little bit of intelligent methodology, Sense can deliver a very complete view of energy usage and management in your home.” (Continue reading here)
Thank you, Julia and the Sense team, for organizing a new way of interacting with users.
Thank you, Kevin, for setting down your thoughts in a new blog series.
You asked if there are any questions after reading the first episode. I have two questions for Kevin.
What might we expect in future blog episodes?
Could you clarify the Sense concepts of “see”, “learn,” and “detect”?
My current understanding is that the power meter summarizes what Sense sees; that all learning happens in the cloud and is invisible to the user; and that detections are when the cloud AI passes a new device signature/definition back to the local monitor for use in disaggreagating energy loads. Yet some of this doesn’t match what I read in the blog.
Thanks again for the new blog series! I always enjoy new content posted here and by Sense.
This section looks a little odd to me.
I thought Teslas could charge at a much higher rate than 7.7Kw. Something closer to 11KW when using US 220V. Maybe it was just limited to this rate because of wiring or other issues
I know they can charge in the 20-30Kw range in countries with 240V and three phase power(My brother has this setup at his home in Australia).
I can’t reveal everything, but the general tenor of future episodes is on helping users get the most out of their Sense through a combination of education and a consistent series of practical tips. The next installment will be on a top level methodology for finding and filling in the blanks where Sense AI hasn’t been able to help yet.
I’ll try to clarify, but that’s a bit tricky because I used the “see”, “learn”, and “detect” metaphors to try to simplify some very complex (and cool) stuff that Sense is doing between the Sense monitor and the cloud. And I don’t know all the proprietary details, so I might not give as exact an answer as some might like.
Seeing - Yes, the Power Meter is a nice, human-consumable summary of all the electrical information the Sense monitor is taking in or acquiring at up to 4 million samples per second. But I have tried to more narrowly use the term “see” for Sense’s ability to filter and spot on/off transition events in the incoming stream of data. To connect the analogy, the rod and cone photo receptors in our eyes produce millions of impulses, that would be meaningless if we were presented all of them as streams of data. Our nervous system processes them into objects that we can “see”, and then our brain can potentially identify or “detect” the object if we already have “learned” the object. The Sense monitor has most of smarts for “seeing” or discerning the on/off transitions in the incoming stream of data, then extracting the key information that could used to identify that transition. One more aside, the power value tagging in the phone/table app Power Meter I highlighted in the blog, is a reasonably good PROXY (not 100%) for when Sense AI actually “sees” a on/off transition.
Learning - All those on/off events, with the additional identifying information go to the Sense cloud, where they are (massively) collected and regularly analyzed. If enough similar on/off transitions meet the conditions I mentioned in the blog - they are similar enough, but the group of them is unique enough, PLUS there are offs that match the ons, that group of similar transitions is “learned” ! I used the term “enough” twice in the past sentence - that value judgement “enough” is part of the magic of machine learning, or learning in general. Suffice it to say that there are lots of mathematical ways make this call intelligently with a little human steering. That judgement call is made regularly in the Sense cloud and the detection criteria for the new device is then sent out to Sense monitors. Also note that new flavors of on/off events have the potential to upset the apple cart, by forcing reconsidering of what’s similar and unique. Once an on / off pair has been “learned”, the learning can be sent back out to the to the Sense monitor so it can “detect” new on/off transitions.
Detecting - Hopefully the previous section has helped differentiate between Learning and Detecting. Detection happens only after Sense has “learned” the on/off transition for a device (or set of similar devices).
One more detail - We don’t get to peek into the key identifying information is for each on/off transition, but Sense gives up some hints when they share technical charts like the one below - the vertical axis is Feature 1 - Power, and the horizontal axis if Feature 17 - p0, which I suspect is some measurement of the phase angle change from the on/off transition. The “17” indicates that there are at least 17 key pieces of identifying information associated with that on/off transition.
@jonhawkes, great observation. My 240V charging is set to the max for my Model 3 which is 48A. As you suggest, that should result in a 11.5kW charge rate. And I did see that in the past. Not sure when my car decided to start charging at 7.7kW (32A) ! I did do a double check on my most recent charge cycle to make sure this wasn’t a DCM measurement issue. The good news is that my Sense measured 2.7 hr charge cycle at 7.8kW early this AM matched up 20kWh charge on the Tesla app.
I’ll have to take to the Tesla forums to see if this charge rate reduction is a “thing”. But this also shows why Sense has to stay on their toes when it comes to any specialized detections for the slow EV ramps. Some of the key features of these ramps are subject to software change without notice.
Interesting. We’ve been using sense since 2019. We also have two EVs. In my scenario it never seems to ‘catch’ either vehicle charging and as of late its mostly due to the ‘charge on solar’ option.
But even before them it would catch it once in a while.
Ultimately after getting a pair of Gen3 wall connectors, I used HA to read the amps and voltage on both of them locally, create a sensor that calculates that out to watts. Then used a HA plug-in to send those two values as a Kasa emulated plug to Sense. And thats worked perfectly. Thou it would be nice if sense could do that directly.
@ymilord , thanks for your experiences. That makes me wonder if the “charge on solar” on and off ramps are substantially different from the timed charging ramps. Glad you are able to use Home Assistant with SenseLink work for you. That’s a great option for enterprising users who have power usage data sources beyond the ones supported by current Sense integrations.
If they are going to mention flex sensors as item 2, then another optional feature deserves mention: integrations. I regard the links with Kasa and Hue as a key strength of Sense. Yes, those are products by other companies so I can see why they don’t spend their marketing dollars to promote them, yet data from those other companies is only available in real time. Logging the data generated by those other companies is a valuable service that is, as far as I know, not available elsewhere.