Thanks. I’m betting the variable speed motors are the challenge. My old school Liftmaster garage doors are picked up by Sense like clockwork. They are one model year to early to be IoTable via SmartIQ.
I agree completely with you that Sense needs take a new tack, though on training, not detection. They need smarter feedback both from humans when detection goes askew, and from smart controllers when known devices are turned on/off. The trick with the former is to do it in a very structured way, so humans give Sense unambiguous training information. And the hard thing about the latter is building robust integrations with the dozens of automation APIs that users are likely to encounter.
I think we’ll see increased progress on the detection side for harder devices as the training feedback improves.
http://community.sense.com/t/we-should-be-able-to-train-this/1483/35 @jackfurr, is the pool pump literally always-on? Can you turn it off at night and then it back on every morning? …maybe even a timer for 3, 4 , 6, or 8 hours off? I’ve owned a few pools and that is what I did only to save energy costs, yet I could keep my pool levels balanced… Power cycling could help a lot with pool pump detection, but while the pump is off Sense may find other devices.
… … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … … V heart V
We did talk a bit in the webinar about the initiatives towards device detection, training possibilities, and integrations to support those efforts, which are absolutely the focus for us in 2018. Some of the work we’re doing to support integrations involves a pretty high development effort up front, but will mean that integrations with devices like smartplugs, connected thermostats, etc. will be much easier to implement down the road.
While the data science and engineering teams are knee deep in getting frameworks in place for these kinds of major detection improvements, there are smaller, parallel teams making sure the web and mobile apps are up-to-date and continuing to add new features, but know that the majority of our resources are focused on helping to make big strides in device detection and integrations this year.
Thanks to all of you for your patience and persistence as we keep working on things, and please keep the feedback coming!
Wouldn’t it be easier to train the model for device detection if everyone entered their appliances by make, model, approx age, etc without directly tagging them and then the Sense ML would know to look for the same energy signature across a subset of households to identify that common appliance?
I was thinking the same thing. I’m not a machine learning expert, but It seems the whole idea is gather as much data as possible to paint a picture. If the machines can continue learning their patterns, and humans provide basic input, can’t we match the two together? Seems like more data is better than less.
This question has come up many times over. Unfortunately, it’s just not that simple. Yes, we need data, but we need very particular types of data. It’s not a matter of Sense just knowing the signature of a device in isolation. Those signatures are unique, yes, but they’re unique in your particular home too. The common ML analogy for what we do is trying to pick out a particular voice in a room of people speaking. This is a pretty strong analogy, but it’s not perfect. That’s already a very hard problem, but what if that voice keeps changing in quality (in vocal range, in accent, in speed, and so on)? That’s essentially what we’re dealing with. The same device can look pretty different across homes based on other devices that are running concurrently as well as the general noisiness of your home power, which itself can vary over time. It’s tough to overstate how complicated and challenging this problem is.
Now, I will say that the points about make and model are solid because they do give us clues to work with. We allow you to do that in the Device Settings screen for detected devices we’d really love it you could do it for devices that you’re certain of. But still, they’re just clues and not 100% solutions.
We still do incorporate user input in other ways, just not in a manner where you can directly train Sense. Every time you utilize features like renaming, Community Names, marking devices as not on, and Network Identification, your input is directly helping us with device detection in your home and across our larger user base.
I agree completely that the uknown list should release the devices and allow the home owners to identify them (in a fraction of the time that it takes Sense to do so). It would still allow for the Sense DB to be built. I understand wanting the ‘unsupervised machine learning’ to do it’s job but right now it is way to slow. Let the end user help Sense and build up the DB quicker while Sense works on improving the code of it’s learning, context and electrical signature systems. Seriously, how long does it take to identify a heat pump or a refrigerator? Excellent product but just don’t understand the logic in preventing end users from marking devices themselves.
I guess I don’t really understand this. If the sense was able to recognize something enough to be able to release it to you to label, then wouldn’t it know what it is so you don’t have to label it?
This request makes no sense. There is no unknown list just a blob of noise. If something was identified it would be kicked back to the user as an unknown device.
I’ve had a Sense for about 3 weeks. Purchased to help me evaluate the effectiveness in a monetary sense for changing from a 22-yo. open-loop lakewater based heat pump HVAC system to a modern American Standard pair of compressors with new air handlers.
So my original system, for more than two weeks, consisted of a 220v, 1hp, submersible well pump in the lake that fired up when either the 2 or 2.5 ton compressors in my garage needed support. I have two refrigerators. Two split-unit AC Heat pumps on another floor. My home has been “automated” using StarGate since it was built in 1996, and I’ve had a long life in technology as well as pharmacy, I am very much the expert generalist. Beta for US Robotics, IBM OS2, Stacker, Citrix, and a host of long-gone softwares. Web designer for 15 years, and database programmer and systems analyst for more than 20. So I have some experience in software development and featureset origination.
So I am baffled to discover that Sense has been around for a couple of years. My initial judgment was that this was a very new technology, from the methods employed to accrue “devices”. Wrote tech support with questions and suggestions, didn’t make any real difference.
WHY iS THERE NO TRAINING MODE FOR THE DEVICE?
Logically, one can install the device properly, and then, with EVERYTHING POSSIBLE TURNED OFF, begin to enable devices, and have Sense respond to that new amperage draw/cyclic flavor with a query/guess, and allow me to immediately specify what that consumption is dedicated to.
Instead, every random few days, a guess is made, and with my sheer diversity of electrical components in the house that kick on and off for randomly determined durations, I have to guess as to what it was/is, quite often WRONG.
So 5 days ago I did the replacement of the two water-based heat pumps with 2 new air handlers which have variable speed motors, 2 new rooftop heat pump American Standard compressors, and altered the “on-demand” nature of the well pump to a simple breaker switch off/on setting. (Though I do have an Insteon 220v switch to deploy for control via my HomeSeer Hometroller S6 Pro, eventually) I am gradually decomissioning my StarGate as I employ a variety of technologies that will perhaps surpass the functionality of the 22 y.o. StarGate. ?
So WHY does Sense, after at least a year on the market, not have a teaching/learning mode available?
Certainly, the business model requires they develop this superb database of fingerprints for each electrical device in use in homes throughout their market. But without enabling more dialogue between the shareholder/owner of the device and the device, ALL data is supect, and the end-user’s experience is compromised.
So, all that consideration evoked, WTF?
Try and think of the electrical mapping like water. You can’t separate or tag the water molecules. That is why tech support will insist that you don’t turn devices off. If it can’t learn your water pool when all the pumps are pumping in normal usage then it never will. There is no finding or tagging a single device or electrical signature on its on. There is only the singular water pool. Hope that helps.
Also, it takes several months for most devices to show. Any slow start devices are tougher like LED lights, EVSE car chargers and two stage heat pumps.
Read a couple of the blogs below and you’ll understand why your ideas on “training mode” don’t really make sense (no pun intended) given the current approaches Sense uses for identification:
Sense currently looks at sub-second features to detect on and off events, often as short as 500msec., not at the whole power waveform envelops that humans are attuned to. I would be impressed if you could accurately mark the beginning and end of a 130msec on-event, especially with background noise. And yes, you do need the background noise - a baseline waveform is useless in the context of machine learning. You need thousands of slightly different noise altered variations of the same on/off signature to learn…
I would recommend reading through the rest of the tech blogs to get a better understanding of the existing technology. There is lots of room for improvement, but whole idea of a human controlled “training mode” for the current form of identification is a bit of a fool’s errand…
In short, it’s not feasible at this time. It’s something we’ve thought deeply about and experimented with a lot (and even had a training mode in an early iteration of the app) and it just doesn’t work as you’d hope. Yes, we collect device profiles from people’s homes, but those are only one small part of the device detection process. As Mark noted above, you can’t totally separate a device fingerprint from the context in which it is operating. Any sort of realistic training mode would thus be so involved and long-winded that it just wouldn’t be worth the cost of entry. You’d essentially need to mark a device as on so many times and in so many different combinations of other devices in your home, that it would make your fingers go numb for days. And it still wouldn’t work like you’d expect. Machine learning is a whole different ballgame. Two years may not sound new, but tech really still is. Device detection has improved drastically over the past two years and will continue to improve every day.
I recently installed my Sense and was expecting much faster detection (actually any detection at all) and am frustrated to learn there is no training mode. It makes absolutely no technical sense why there can’t be one. If an “AI”/algorithm can “learn” a device signature, then SURELY telling that AI/algorithm, “hey, here is an instance of the power signature of this device, it is 100% known to be this device” then that will help it discover it in the future. Furthermore, if the device signature is recorded when all other things are off, how can that not be the best possible sample for a specific device. The resistance to making a learn feature is confusing. Your app is great, the device is a great idea but your marketing hype is extremely overblown compared to the actual results thus far, I’m still hopeful though…
I’m going to quote two posts from earlier in this thread that explain why training is not technically feasible in an ML framework:
Trust me, were training feasible, we would do it. It would solve a ton of device detection issues. Any resistance is not because we don’t want to do it. We have spent a lot of time looking into training options and it’s just not viable. Machine learning doesn’t work like that, at least not in its relative infancy.
I’ll go the other way with this. I think Sense shouldn’t let us train the data. Any human intervention can corrupt the entire data model. Allowing people to even touch the data is terrifying. How would you stop someone from maliciously trying to train Sense the wrong data, to wrong devices on purpose. Those errors or propagate though the data set forever. Like a genetic disease that you can never cure. That fact that we are even aware that Sense is listening probably causes trouble. How many people trying to rush detection, have turned off their A/C or just messed around with their devices uncharacteristically. That right there is messing up the data. I assume Sense only accepts a pattern once they see it a “million times” to smooth out this behavior. Thanks Sense for being nice and not telling us what we really are…End Users and all that, that implies. Sorry people with PHDs in Quantum Mechanics, Thermal Flux Capacitors, and all the other cliche titles. Watch this video and in case your’re wondering we are NOT the Rick in this situation. P.S. Or Summer.Rick and Morty True Level