Sense’s detection algorithm needs help with training. We (customers) want Sense to find all major home appliances promptly (water pumps, refrigerators AC/Heat units). Sense’s current algorithm has been working with my data for three days and found only three devices: 1) Dehumidifier (cycling); Aquarium water heater; 3) One of my two refrigerators.
Looking at the main power meter in the Sense web application, I see clearly when my 240V water pump starts – Sense needs to enable supervised learning mode that can take the customers’ input and focus on learning the specific pattern identified visually by a human (and additional information provided by the human) and mark it as pump.
It seems that Sense is trying to use our data to experiment with unsupervised learning – this is a great scientific goal at our (customers) expense.
I’d like to help Sense with their scientific research and experiments, but I really need my water pump be detected, so I know when it is running for too long.
Three days is probably not enough time to gather data for a 240V device. My 240V devices have taken more time rather than less to detect. Possible reason here: AskSense - April 2019 - YouTube
I would love Sense to use more supervised learning, especially when smart devices offer subscribed access to the data via REST APIs and OAuth2. Two examples that come to mind for me are my NuHeat thermostats and my Teslas. But I’m leary of human “supervision”, mainly because humans are typically too sloppy for the accuracy and repetition required.
There definitely supervised learning taking place if you pay attention to the current integrations. Sense is certainly using the Ecobee Historic integration to improve HVAC models. And I’m sure Sense is also using data from all the integrated Kasa and Wemo smartplugs to enhance modeling for user devices attached to those.
Regarding point #2 about humans being too sloppy/slow/choose-your-adjective to get accurate timing around on and off transitions, I think having a “this device is on” like what we have today for “this device is off” could, over time, help the algorithm determine improvements to detection of already-detected devices.
For instance, mentioned (I think by you) in another thread is that having long AC run times can have Sense just assume that it missed the off transition after a while and lose the bubble. If I opened my app and saw that the AC wasn’t listed as being on, but I knew that it was on, I would like to be able to say “this device is on”. Even if that had no immediate effect on what I saw in the app, I believe that this data could be used to help learning. Hey, if nothing else, I guess they could just throw away that data but use it as one way to let people think they’re helping (shutting down some of the “training” questions that come up).
I like the fact that Sense is trying to find a specific device signature - helps a lot to know when it starts/stops, or something is going wrong with it over time.
However, for my benefit, I’d like catch large appliances by not a signature specifics, but by the characteristics at large: where, duration, frequency, number of legs, wattage etc. then continue to collect the detailed signature within the large profile. Don’t separate legs 1st… A specific large increase to wattage is a signature on its own… collect General associations 1st then details. Verify associations with a human converting them into rules. Yes - bad input from a human can spoil the analysis - verify all inputs with the user and make the user be liable for wrong answers. There is-no AI on the market yet - and not even a hint to have one in The near future, so use us - the humans.
Still believe that the true goal of Sense is to get free sets of data to work with, not to produce a product to notify me - when my pump is running for too long?
@qrnef, you’re right. When I mentioned “too sloppy”, it was only in the context of accurately marking on and off transitions via the form of training that many users raise on this forum. Tagging existing detections with device name and type, negative tagging wrongful detections and tagging time instances where the data should be tagged with a device (as you suggest), are indeed useful for crowd sourcing identification and tuning an already learned model. But model creation based entirely on human training is another story.
@quedak As @kevin1 highlighted, 3 days is definitely not enough time for Sense to detect your 240V appliances. You’re still pretty early into your Sense journey. In an average home, Sense sees about 25-30 detections by the six month mark. Smart plug integrations are a really good option for instant device detection with Sense.
The conversation of a learning mode has come up a lot and there are a few resources (@kevin1 pointed to one earlier) around the community that highlight why this isn’t a solution we’re looking further into.
I think you’ve highlighted a major obstacle right here, which is something that we’re continuously building towards (but won’t happen overnight) due to the amount of variation between different users homes and appliances. I think it’s important to convey is that as a small company, we’re building towards better device detection consistently with calculated, behind-the-scenes work on our end. Device detection has come a long way from where it was a few years ago, and it will continue to improve moving forward.
Although providing users with data is important to our overall mission, Sense is ultimately meant to increase your understanding of what’s using energy in your home. You mentioned you observed your water heater turning on within the Power Meter. You were probably able to see how much electricity your water heater uses, which we’re hoping gives users the ability to change their behaviors around their home based on the costs associated with their energy usage while we continue to add features and integrations to give folks more insights into their energy hogs at home.
@JustinAtSense There is a number of very inexpensive sensors that can plot the overall view of how much energy is used at home, and Sense provides this information as well (the main meter feature), and Sense presents this nicelly. (A side note: I’d like to have a feature of setting up a notification on several basic attributes (fitters) - wattage is over a certain number, number of legs used… this will tell me when my pump is running…, a useful non-“smart” feature) Meantime you can collect the data and massage it looking for models…
I just expected from Sense a bit more. With the approach you’re taking, I feel very pessimistic (I want be wrong) about getting a more detailed picture about energy consumption in my home (at this point, I know when my aquarium heater is on).
Perhaps, something could be changed… An email spam heuristics based algorithm could yield better results - BUT- you have to mark the spam as spam to make it work, over an over again - mark water pump as water pump… I understand that detecting devices is way more complex than detecting email SPAM, and the Sense team is very small, but you could use your customers to evaluate data sets, and yes - it will be very home specific, but it will work for each customer.
Back to my goal - I need to know when my 240V 13 AMPS water pump is running. There is no an integrated (consumer) smart plug for it (well there is one BOSS-1000), so I hope Sense will replace this very expensive tool.
Will enjoy the summary chart of the energy consummation for now.
@quedak Although we’re a relatively small company, everyone at Sense is working towards a better Sense for users. I believe that most community users that have been here for a while would agree that we’re constantly working on new features, many of which are sourced from the community product wishlist subforums.
I can’t say a lot about this right now (we’re still testing), but we do have a couple of projects that we’re working on right now that would address your water pump. My advice, for the time being, is to give Sense some time to recognize the patterns in your home electrical usage and check-in over the coming months for updates or new features that i alluded to .
I noticed that some users reset their Sense to get better results - second, third (even 5th) try worked more effective in recognizing the devices. Can you recommend the resetting the Sense?
The users that you’re seeing reset their Sense units have had them for over a year, closer to two in most cases and had to alter their monitor setup. I would recommend not resetting your Sense, @quedak.
I think the only time Sense support would recommend resetting your data is after you have made physical changes to your monitor setup (i.e. moving the clamps).
I’ve done 3 or 4 resets, but only my first was actually required. The rest were “just for fun” and to see if I could influence the learning/detection rate in a significant way through “artificial on/off cycles”. Most of my major devices are detected within 4 or 5 weeks so it doesn’t take that long to catch up to where I was previously before a reset.
I wouldn’t recommend doing what I do unless you are also keeping good records of when devices are detected, when your monitor receives model updates without a new device detection, etc. Otherwise, you won’t really have a good idea of what was different between this reset and the last one.
Many electrical devices have a dependency on the weather, or even on the current setting of the AC thermostat - when we lowered the temperature in our house my aquarium heater started working more often (my fish want be happy at the temperature they like).
Here is my question - should I start triggering various devices that are otherwise dormant? Examples: boil a glass of water in my microwave (we use our microwave not the often); trigger the the water pump (it only works when it rains); turn on our music amplifier (we use it only when we host a party). Will this help with finding the devices’ models? Should I trigger the devices on a fixed schedule, or randomly?
@quedak, before you spend time doing that, you should know more about your devices. Cycling devices might help speed up detection, but only if they fall into the likely detectable category. Questions to ask:
Does this device have a DC power supply or are there parts of it that are AC operated ? Things with DC power supplies both internally and in the “brick” form, will not be detected because they don’t really have clear “on” and “off” signatures. So most consumer electronics, like stereos, PCs, communications and networking gear, plus TVs won’t be detected. There is a special case with TVs - some smart TVs are picked up by the Sense NDI integration. But don’t bother with your amplifier.
Appliances come in both flavors - Garage door openers, microwaves, refrigerators, washers, etc come in flavors that use both AC and DC inner workings. A garage door opener with battery backup uses DC, and not likely detectable, while a typical AC garage door opener is detectable. Microwaves and refrigerators with the “inverter power” designation are DC, so not likely detectable, but others are. Check the type of microwave you own - I have two old school transformer based microwaves that have both been detected.
Sense is working to solve the “DC/inverter challenge” today, but mostly for devices that use lots of electricity, like EV charging and heat pumps/mini-splits.
Yep, this is the biggest thing I have learned in doing my experiments to speed up detections. Also, to speed up detection by weeks, it seems you have to put days or months worth of cycles into hours which is not practical to do manually.
My wife’s laser printer was detected very early since she prints a lot of shipping labels (at-home business). I rarely print anything but still have a laser printer in my office. For fun I wrote a program that would print out a blank page every 15 minutes. My normally idle printer was detected a day or two later. On another reset, I was able to get it detected using an alternate method. I plugged in the printer to a smart plug and then setup an automation with a motion sensor to turn on the smart plug and then turn it off 15 minutes later. The printer’s startup/warmup routine was enough to get it detected without creating any paper jams
Later I combined the two techniques by writing a program to randomly toggle a smart plug every 10 seconds. I connected a vacuum cleaner to it and let the program run for an hour or two (God bless my wife). It found the vacuum cleaner a few days later. This same technique did not work for a box fan though.
Although this is fun, in all cases, I learned that devices detected in this fashion normally cost me less than $1 a year to use. Also, you are potentially shortening the life of your devices by putting them through more cycles in a period of time than they were normally designed for. All the devices got very hot while doing this and I have jammed at least one relay in one of my early experiments with a bank of LED lights
I would say give it at least a month and if you are not happy with the results, then you at least have some data to know what is detectable in your house within a month of normal/regular use. The longer you go, the more data you will have to use as reference for a reset if you wanted to try it for fun.
I suspected that frequency of cycling could help the process. Sense found my devices in order of frequency they cycling: aquarium heater, garage fridge, dehumidifier and finally the bug fish - the AC.
I think we have a recipe to speed up the detection process, but I think patience is the way.
Patience is definitely the key. My first decide was the basement fridge because it was old and wasn’t sealing properly, so it was on 50% of the day. My AC just took about 2 weeks of constant use to be detected. I’ve had my Sense since January and am only missing a couple larger appliances. It even found a pair of florescent ballasted lights in the home office.
Sense recently found my AC as a 240V device, but when my 240V sump pump runs it also being reported as my AC. My AC running on average for 20 minutes, and the pump runs no longer than 15 seconds, the wattage is similar, but charts look different. It seems that Sense cannot differentiate one 240V device from another.
When the sump is running but sense says it’s the AC, click on report a problem and click “This device is not on”. Also, I take it you have named the AC as the Air Conditioning? My dishwasher shows up as my toaster oven since they use the same wattage heating element. However, you’re look at the detection with hindsight. When the device turns on, you, and the sense, doesn’t know how long it will be on, so it’s not fair to criticize a device based on the future that hasn’t happened yet. I would wager that Sense is able to look at how long a device is on and factors that into its detection of devices, but that’s just a guess. Again, patience is key. Be glad it found your AC so quickly.
.