Disappointed in device discovery

[andy]

I am…and electrically it should make no difference what so ever. A breaker that is “on” is pretty much a piece of wire between the bus bar and the screw. I haven’t tried it on its own breaker, because I’d have to switch the A/C lines to a different box, and I’d be amazed to see any difference.

I have AWFUL detection using a shared breaker (it’s on with my water heater). Sense engineering says it’s because my home is “noisy”; specifically I have a constant pressure water pump that services my domestic use and my geothermal heat pump. The signature is VERY distinctive to the human eye, but apparently not to Sense’s algorithms, and it masks almost all of the other devices in my home.

During the winter (pretty cold up here in NH…days and days below zero lately) the geothermal heat pump runs ½-2/3 the time…depending on outside temps. During the summer, the geothermal reverses and provides A/C, running less than ¼ time on the very hottest summer days. With the heat pump, the deep well, and the associated pumps for radiant, it’s by far the biggest collection of power consumption in my all electric home, but much of it not detectable. Sigh.

[samwooly1]

Did you see where another responded that he had to reset his after
changing where he powered his sense from even though it was the
same legs?
I think the sense is using both clamps AND power wires for measurements
or he shouldn’t have had a problem.
I also have mine on the water heater breaker.
Heat pumps are designed to run almost all the time in the colder weather
even close to 100%. It’s common misconception that they run “backwards”
in the winter. The compressor only turns one direction. It’s a reversing valve
the opens or closes the reverses the flow of Freon. If you have aux heat and
a thermostat with outdoor temperature sensor then you can set it up to turn the
heat pump outside off when temperatures drop to a certain temp. This would
not apply to yours if its geothermal water system.
The geothermal water system shouldn’t reverse, the water runs a single direction.
If its 100 outside the groundwater is easier to cool with the compressor at a constant
58. If its 20 outside its easier to heat at same 58 degrees.

[ben]

At its simplest before any machine learning gets involved

CT clamps are measuring current.
Red/Black wires from Sense to Breaker (shared or not) are
a) powering the Sense (well one of them is)
b) They are measuring voltage on each leg.

By measuring current and knowing the voltage, they can calculate power. Then they can feed all that to the ML to figure out which bumps / spikes = your water heater.

The reason moving things in the panel can screw it all up is that now the math could be wrong.

If we have CT clamps A and B and power leads X and Y.
Part of that first signals check when you first set up the sense is it learning which CT clamp is on which leg which matches with which power lead into the Sense.
It also checks to see which way the CT’s are facing since they are directional. Are they both forwards? Both backwards or split?

So it knows that A and Y go together and B and X are a pair.

So if you go and move something, it is possible that when you put it back, where it was used to a pair of -A/Y and B/X and now its -A/Y and -B/X or A/X and B/Y- it gets confused.

So anyway, thats my really long way of saying that yes, all 4 things are taking measurements. But I don’t personally think the Voltage measurements coming off of a shared breaker vs a dedicated breaker are going to be any different or affect your detection. I think the “noise” of the house has much more to do with it.

[samwooly1]

When reading posts from various users it appears that the ones with the cleanest energy usage, like solar, geothermal and electric vehicles, have the “noisiest” houses. Is it my imagination?
I noticed problems with noise from deep well pumps. I believe there is a way around this using a cyclic stop valve and a traditional pump instead of variable speed. The results would be almost identical but produce less “noise”

[oshawapilot]

FWIW, one of the questions I did pose for the “Ask Sense” thing running right now is a list of what they consider the “noisiest” devices that are known to impede or harm detection.

I think having this info out in the public realm may aid in people either better understanding their issues (at worst), or being able to make changes to potentially remedy the issue (at best).

For example, my own personal detection seems to have hit a wall in the last few months, but If I knew (via a list of known problematic items) that I had something plugged in somewhere in my house that may be the culprit (without needing to file a support ticket and going through that process to perhaps identify it), there’s a good chance I’d either remove or replace said device with one less noisy.

I hope that discussion is enlightening.

[samwooly1]

I agree. I’ve read where users have stated that support told them their house was noisy but it came across as they were basically telling them they were out of luck.
A little more understanding about exactly what their definition of “noise” is. There is the ever present 60hz hum we here but is noise better explained by spikes and fluctuations where sense is concerned?

[samheidie]

I’m starting to think of the noise being sheer wattage being used, rather than device specific.

When you’re drawing 10kW a device turning on using 100 watts is much harder to see vs a quiet house chugging along at a couple hundred watts and the same device turns on.

Other than sheer wattage, I’d say the next biggest deal is how many devices are using the total wattage? 25 random smaller things or a water heater, heat pump, and microwave?

[andy]

What was described to me by Sense engineering was signatures that vary quite a bit during operation…not just a start up followed by steady state. In my particular case, my constant pressure well ramps up gradually, then the instantaneous magnitude depends on the flow required to maintain pressure, with lots of “jitter” (which looks to the algorithms like lots of start-ups) as it runs. For much of its use, a few GPM, the overall magnitude is quite low and the jitter is fairly low, but when the geothermal kicks in it shifts to 15GPM plus whatever domestic use there is. The difference in overall KWH is 8:1 or more, and the instantaneous changes are ongoing and larger than the average usage….e.g. “noisy”.

All of this is pretty visible/recognizable to the human eye, but it doesn’t look much like other device usage. Unfortunately, to Sense it doesn’t look like any one device, and it masks other devices turning on and off while it’s running. During the summer, while the geothermal is kind’a coasting (maintaining 72 degree house temperatures vs 50 degree geothermal temperatures) while providing cooling the heat pump (and associated well) aren’t running all that much. During NH winters, recently at 10-25 below zero (a whopping 50-65 degree difference in the “wrong direction”), it runs much more often, thus masking most of our other devices constantly.

So, our home is “not suitable” for Sense. I agree with those who point out that Sense’s marketing mentions absolutely none of this and some suitability information could save some portion of their customers lots of pain, frustration, and cost.

[kevin1]

Both @samheidie and @andy make good points. Sense looks for signatures (signals) in a combination of features (voltage, current, phase angle) across several different time windows and sampling frequencies. Noise stems from electrical perturbations of those features in the specific time windows/sampling. Perturbations that are close to real signals, but impossible to resolve.

A good analogy is you trying listen to a friend with background sounds:

• 1,000 bats shrieking aren’t going to affect your ability to understand, because they are well outside your “signal range” . The same for whale sounds.
• Within your audible range, the chirping of birds or lions roaring could affect some comprehension, but not as badly as human generated sounds
• You could probably still listen to your friend in a crowded bar, as long as the crowd noise wasn’t overwhelming, because you have some mechanisms for cancelling out other peoples speech, by basically understanding it and putting it aside.
• The most disruptive would probably be a room full of people mumbling, then shouting in Old Norse.

Andy’s well pump is one of those guys muttering and yelling in Old Norse Perfectly understandable if you are listening for Old Norse, but a huge impediment to hearing what else is going on in the room.

[andy]

((- Love the analogy

[RyanAtSense]

Well that’s a pretty fantastic analogy. Don’t be surprised if I borrow it in the future

[Grandpa2390]

I’ve given up on device detection. It was going great, and I had several devices detected. Then sense just lost its mind. started detecting multiple air conditioners. confusing devices with one another that it had been able to tell apart for months.

I attempted to delete all of the devices and let it redetect them with the data, but it didn’t do that too well. I feel like the data has gone bad, and at this point maybe I need to erase everything devices, data, etc. and start over. and repeat that process every few months.

But it’s not worth it. Just need to be happy with the data that I get that is correct and ignore the device detections. Even when a device is detected, it isn’t accurate enough to be worth it. like Sense finally detected kitchen led lights, (8 tubes), but it doesn’t recognize that I turned them on until 5 minutes after. great it is seeing (eventually) that the lights are on. but from the perspective of wondering how much they are costing me, it is not worth anything.

at this point, I just check the app when I want specific information about my overall usage, or to look at the chart and interpret it with my human eyes. I can tell what’s running and what’s not just by looking at the graph. It’s been 12 months, but really you only need like 3 or so before you learn 100x better than sense how to understand what your breaker box is singing.

[samwooly1]

What has advice has tech support had @Grandpa2390?
Was it their idea to delete all your devices or yours?
Have you asked about resetting?
I don’t know anything about reset but I’m thinking that resetting versus deleting all devices would allow to start with a clean slate.
I honestly don’t know enough about it but would consult them before giving up on it.
I do agree that the timeline/graph is very useful and is what I currently use most. At least until detection improves.

[Grandpa2390]

@samwooly1 i’m not reseting all my data. If you reset all your data, you lose it. And eventually I’m going to figure out how to make up for all of Sense’s shortcomings. via tfttt and so forth. Exporting before resetting would be an option if we could export at a decent resolution. But until then.

It just doesn’t work, there’s nothing tech support can do about it.

[Grandpa2390]

@kevin1 the human brain is a lot more powerful and capable of picking out desired sounds from noise than you give it credit for in your analogy. That’s one of the incredible things about our brain versus voice recognition software, face recognition, etc. The only kind of analogy that works is trying to hear your friend at a rock concert. not because it is hard to distinguish his words from the rock music, but because the rock music is so loud, you can’t even hear his words.
That analogy would probably describe why Sense doesn’t detect most of my light bulbs and even my laptop. the small volume of energy/power usage is drowned out completely by louder equipment like the air conditioner.

comparing sense to the human brain is not very fair to sense.

an interesting comparison might be trying to make intelligible speech by looking at a frequency graph of someone speaking with other people speaking. To try doing what the machine is doing. it’s impossible to listen with your eyes. It’s amazing that machines can be taught how to do it.
what sense does is a bit easier than voice recognition though. but only because voice recognition requires the machine to learn a larger vocabulary. listening with our eyes might be just as easy as looking at the Sense chart if english only had a couple of words. Our brain would still be superior, but you can imagine the learning curve it might have just by playing with the sense graph.

unless your house is like mine, with very few devices, and they all pretty much run at different times. my house is easy. but if you have a bigger house with lots of devices…

[kevin1]

All good points, but I think you are simplifying the Sense challenge too much by centering it around what needs to be done in your home.

• First off, my home is a cacophony of devices. I have 61 recognized devices with another 100un recognized. Many of those devices are electronics that eek along on a small near constant power level (mostly AC/DC power supplies) that hide the more variable nature of the digital device on the the side.

• Second, many of the major device patterns in our household are not “words”. Many composite devices offer up entire sentences, like my washing machine below.
  

  

• Third, there are literally tends of thousands of specific devices that Sense has to sort through, all slightly modified by the dialect driven by your home’s unique wiring.

I’ll also point out that we have been working on voice recognition for far longer. When I was graduate student in the 80’s I wrote a survey paper on voice recognition approaches at that time, which mostly relied on pure DSP approaches, but needed computational and contextual help sorting out phonemes when encountering allophones. But good adaptable, real-time, small-footprint voice recognition has really only come into its own in the past 10 years by using a machine learning approach.

What’s fascinating, but not surprising, is that the same types of neural nets used for speech and natural language processing, are also at the center of Sense-type tasks of energy disaggregation.

http://cs231n.stanford.edu/reports/2017/pdfs/804.pdf

[Grandpa2390]

@kevin1 I’m not sure how to interpret your response. Your language says that you disagree and are arguing with what I said, but your arguments are all ones that I have already made, myself, in defense of sense and its developers. So perhaps you just misunderstood me, or I am misunderstanding you? To say the least, the intent of my response, that you replied to was to say that it’s not fair to Sense to compare it to such a powerful “computer” and its abilities. What our brain does so perfectly (sometimes or usually depending on the task) is so much harder for a computer.

But yes, as I’ve said before, scientist have been working on face and voice recognition for decades, and even now it is still spotty. Since this uses the same technology, I don’t think anyone should be buying this device expecting the device recognition feature to work. If that is the primary motivation for buying it, the purchaser will be, at the least, disappointed.

I don’t know what you mean when you say I am simplifying the sense challenge too much by centering it around what needs to be done in my home. What I’m saying is that my home, compared to someone like yours, is a level one challenge. And if it doesn’t work in my home, I’d be surprised if it worked in a home like yours.

Going back to the voice recognition analogy. my home is a soundproof room, and your home is a crowded restaurant, or bar, or whatever. At least in my mind, if the voice recognition doesn’t work in a sound proof room, it is not going to work in a crowded room with lots of noise.

For your first point, It’s detected lots of devices in my house. Just not accurately. For example, a week or more ago, it detected the lights in my kitchen. but it doesn’t detect consistently, accurately, or precisely. It has trouble knowing when the lights turned on. And I find that with other devices as well. Even for the devices that it detects, I can’t trust it because it doesn’t detect the device for the entire duration that the device was consuming power.
I mean they’re leds but its 8 of them, and they are tubes. so they consume a lot of power. I’m amazed that Sense detected them because they don’t exactly turn on in unison. I’m sure you are familiar with that.
I didn’t realize how much power they used until I installed Sense. Now I rarely turn them on, lol. If I can help it, I turn on the led lights in the Dining Room, and that light shines into the Kitchen enough for me to do most of the tasks that I have to do in there at night. During the day, I have sunlight. So don’t think I am ever saying that Sense is a terrible useless product. I just think that one feature is fun but not accurate enough to be useful. Just like the voice recognition on my phone. lol.

Over time, and the Sense Support guy acknowledged this, the device starts to get confused the more data it collects. at one point it detected both my dehumidifier and my air conditioner perfectly. then 3 months later or so, it started confusing the two.

For your second point, We are differing in our vocabulary. I am using the vocabulary of Sense. and to Sense, devices are words. Your washing machine is not a “device”, it is a composite of devices. It is a machine. So yes, I agree with you. it is a sentence made up of individual words. the different components that make it up are words. the different parts of the cycle (spinning, agitating, etc that’s how I labeled them) are words. My dryer is gas, Sense has not detected the tumbler. but it has detected some part of it that I can only guess is the ignition, lol. It has detected the light in my fridge. It detected the blower motor and AC unit separetely. The blower motor and dehumidifying component in my dehumidifier separetely. that’s cool with me. I can understand why it would do it that way, and I prefer it that way.

I can look at my graph and tell that the dryer is running. It will get better with time. Voice recognition still stinks (imho), but it’s def better than it used to be.

Here’s my house over the last 24 hours. Each one of those smaller identical spikes is the refrigerator. around 4 in the morning is the furnace and a space heater as you can see. etc. the cursor fell right on the spike for the fridge. 999 w is not my average power usage.

2 hour shot:

around 10:35 AM is where I turned on the living room lights. (10 cans), and a moment ago the dehumidifier turned on. every other spike is the fridge. The fridge is one device that seems to have accurate detection

As you can see, I don’t need the device recognition. I feel like the market for the device is made up of hobbyist like myself, and professionals (or professionally skilled hobbyist) like yourself. And I can’t imagine anyone buying into this device being less familiar with the technology than I am. I don’t understand when people get on here complaining about device discovery as though that breaks the device and renders it completely useless.

I don’t need the device recognition. as a hobbyist, what I need is a simple interface that allows me to select the data and assign it to devices. hobbyist like myself and professional/hobbyist like yourself, I find, tend to complain more about the lack of data manipulating features (tfttt support for example). Personally, I would rather see the Sense team focus more on releasing these kinds of features that allow us to play with our data more than trying to develop the part of the device that might not work decently for another decade or more.

Maybe that sounds pessimistic… but I’m just looking at the state of similar projects that have much more money invested into them, and the ability to collect data and play with it was the reason I invested in this product. Device recognition was more of a fun bonus if it worked.

anyways, sorry for the book. just trying to say that we’re on the same side.

[kevin1]

Sorry, I think I misconstrued one of your sentences. I interpreted “listening with our eyes might be just as easy as looking at the Sense chart if english only had a couple of words.”, to mean that Sense only had to handle a highly constrained waveform vocabulary.

Understand that we are really coming from the same place, though I might be more enthusiastic that machine learning is coming into it’s own quickly as the right large labeled datasets become available

[Grandpa2390]

I hope you’re right. i look forward to the day when my phone’s voice recognition works properly and Sense’s device recognition feature works flawlessly. until then, I am learning how to program so I can do with my data like some of the other people here.

[don1]

After 3 months with no additional response to a support case opened in October, 2018 it seems time for me to just tap out and accept that this platform is just not going to work for me; and since I’m apparently an ‘edge case’ and therefore not high on the priority list, it never will. I’m not interested in constantly spending more and more money for smart outlets or whatever the next incremental investment will be to make up for the (in my experience) limitations of the platform.

The platform seems to be hamstrung by the obsession with the ‘machine learning’ model and everyone who questions the aggressive reliance on that model despite its clear limitations is just shouted down in this forum. 9+ months after installation my Sense has discovered less than 10 devices (3 of those in the first 2 weeks) that were not assisted by some other investment. That can hardly be considered a success. And when support tells you they can see the devices in question in the logs (washer, dishwasher, TV’s, etc…) but 5+ months pass with no change it’s clearly time for me to give up the fight and just live with it as it is and move on.

It’s certainly not a solution I’d ever recommend anyone invest in and I’ll now start researching alternatives in the marketplace. I had high hopes, but alas the solution does not match the message for me.