First, thanks to all that replied. I DO understand, within limits, the nature of the complexity of what can be literally dozens of different devices singing together in disharmony. Given the array of posts up here, over several years, I am chagrined to admit I did not bother to delve into this forum prior to purchasing the device. I accepted the superficial advertisement BS, and had NO IDEA how lethargic and drawn out the recognition/learning process would be, nor that it would not be designed in a way that eliminated ME from the training process.
It IS useful, albeit in a limited fashion. That the company gets all the same comments as mine, and has NOT felt the need to create additional algorithms or processes to accelerate the Senseā learning process, really is disappointing. Iām willing to put up with this BS, and hope that I donāt have to wait a year as some have had to, in order to ID the majority of my houseās electrical loads/devices. But the proposed device identifications I already have (24 so far) have so many errors that Iām tempted to delete most of them. I have one whole-house vacuum system, a Kenmore in the garage, yet Sense has identified THREE vacuums so far, NONE of which is that monster motor in the garage. I have ONE submersible 220v, 1hp well pump, yet Sense says I have TWO pumps and two Motors, NONE of which are the lake pump. I do have a grinder pump, which it has correctly identified. I have one dryer, Sense thinks I have two. So hey, ONE correct major devices out of 10 indetified aināt bad, right??? right???
@texarc,
Understand your frustration and get your desire to take an active role in training. I think a bunch of us who have had Sense for a year or more have seen big improvements as well as a few setbacks over that time. Thereās still a lot Sense needs to do, but I see tangible progress. Now that you are āinā on the product, hoping you can celebrate the improvements as they come.
Personally, before I bought sense, or before I buy any product thatās more than say $100, I thoroughly researched it to see what it was capable of. I spent probably 2-3wks researching it before I bought.
Even two years ago, it was made pretty clear that sense did not have a training mode and the reasons were spelled out very clearly. Even two years ago, it was pretty clear that Sense was a work in progress, no matter what the marketing said. There are readily available reviews all over the internet.
This thread is BANANAS, and should probably be locked except for the fact that someone new will come along next month who hadnāt bothered researching the product. SMH as the kids say.
Helpful post, as always.
But I do think that we can continue to refine our messaging about training and device detection. Itās a tough issue to communicate because itās pretty complicated and really unintuitive if youāre not familiar with ML. Truly, it took me a while before I fully wrapped my head around why Sense canāt just be trained, and I work here! I donāt want to lock the thread as I do think there is value in being helpful to newcomers (Iām working on a āstickyā about the topic though) because, again, itās a pretty legit desire, despite being at odds with the core of Sense. You, @kevin1, @markhovis73, and @MachoDrone have been incredibly helpful here and Iām thankful (and I hope recent Sense converts are as well). Iāll likely borrow some of your language for the āstickyā, if thatās ok. 
@texarc This is definitely one of those areas where device detection varies drastically across homes. My own experience with Sense is similar to yours with a lot of duplicate devices and the like (even I donāt have a data scientist on speed dial to solve my problems). But I want to stress that we have incorporated new processes to accelerate device detection, like network identification and community naming, as well as a ton of non-public facing work. This last one needs emphasis: these algorithms arenāt just set and forget, and our large data science team sits back, sips their coffee, and watches them work. Theyāre constantly refining and building new detectors. Weāre working on ways to make some of this work more publicly facing, without giving away any secrets.
Maybe another way to look at the training request would be the ability for a person to identify a pattern, submit this pattern, make and model, select from a drop down the device type and this info is used within the sense database to identify future devices. Kind of a way to kick start or boost the database.
This may not fit into the Sense model, but if it does, you would have a lot of dedicated users shutting down their entire breaker panels to isolate loads to submit to you. That would be a lot of data coming your way.
Just a thoughtā¦
Iāll likely borrow some of your language for the āstickyā, if thatās ok.
Somehow use the word bananas in uppercase.
I just donāt know how you could stop someone from purposefully trying to break the learning system by training it incorrectly on purpose. Be it from jealous competitors, user error, or just plain trolls that just want to watch the world burn.
Very true. Unless this data is collected and reviewed before bein committed to the database, that would be a possibility. This would add human interaction. The sense learning model removes this
Ryan your answers have been very in depth and are appreciated but users are still beating this topic like a dead horse. Could this thread be locked and stickied?
Sure, just ignore the elephant in the room. Had I known what I was really buying, when I bought this NONSense several years ago I would have made quite a different decision. Just as had I known how lousy the 44 solar panels on my house would be I would never have invested in that either. The answers I get back from both the Solar company and the NONSense people are about the same. Their product sucks and, well, thatās life. At least I have the intestinal fortitude to sign my real and correct name to any post I make and not hide behind a moniker. My NO-Sense has been circling the drain since I installed it. Refinding heat pumps, and many other devices, and forgetting that just last year it was straight. It is of no use to me and I mostly ignore it. Ignoring anarchist non sense posts is becoming increasingly harder.
Spelled out B-A-N-A-N-A-S
You rocking one of the original prototype models?
No clue, 2016 I think.
@texarc,
I would suggest that any techie that is frustrated with Sense detection read some of the recent university papers on energy disaggregation the see how nascent a technology this really
is. We really are riding the early wave, even though many of us are seeing tangible and somewhat predictable results. MIT (REDD) and CMU (BLUED) efforts are both really coolā¦
My power bill last month was $25. That is both my house AND 80% of my transportation miles. All powered by a 6kW solar array. Thatās 24 250 watt panels. I did the installation myself so I saved the labor hours. Here in NC, I spent $1700 (after credits) on a 1kW addition to paying for 80% of my driving for 25 years (the life cycle of solar). 1 kW covers 10,000 miles. If you drive 15,000 add two more panels. I am very sorry you had a bad installation but it isnāt the technologies fault. It is fair to point out to folks that tree shading can compromise your installation. But it is has to be called out to say the technology doesnāt work Jack.
Jack, shoot me a PM and we can discuss how to make this right for you. Iāve kept this thread unlocked for the very purpose of not ignoring the possible issues that Sense has. While Sense works well for many of our customers, it does not for everybody.
@jasonemoyer That idea is complicated for the reasons that @oct42 mentions, and it truly wouldnāt give us that much more helpful data to justify that labor that it would involve on the behalf of our customers and DS team. Again, ML here doesnāt just rely on a database of existent signatures. Rather, it has to parse the unique signatures out in your home. That sort of backend database can help this process, but itās not a 1:1 link.
Late to the party here but jdtsmith, you ācouldā lay the bread on the floor and it might get toasted soā¦
Iāve said it in previous posts, but suppose we would give you a list of all/most of our electrical devices in the house with information like manufacturer, make and model, 220v vs 110v, item type (like, Car EVSE, heater, fridge etc) approximate watts, and maybe some other info like put in service, took out of service dates.
This would allow data scientists to know what users actually have, and work on those models that have the greatest penetration in the households.
If Sense needs to guess at what an item is, and it doesnāt match an item on the list, then it should not list as a specific detection, but as an unknown type.
Just my 2 cents.
@miraj,
I like your marketing angle, giving Sense data scientists a better picture of what the population of end-users devices consists of so they can target development appropriately. But a home-by-home list would likely be useless for recognition purposes. Thatās like providing a list of names of people who are in your photo collection to facial recognition software. Itās not useful until the software has identified faces in the collection, and most of the learning value comes from a human tagging identified faces with specific names.
@kevin1
Having the list would eliminate false positives. If the item is not on the list, then it canāt exist in that particular home. The data scientist(s) could then re-tailor the data model to remove the false positive.