At what point do you just give up on Device Detection?

Can understand your pain, but you suggest a couple things that are a bit off-base.

First, I think you might have missed it… The return policy is 120 days now.

Second, you are clearly in the weeds on the technology part if you think it would be easy to have you “mark” devices via patterns in the Power Meter, unless you are talking about manually marking each and every usage. But maybe that’s what you want ? I can see value in being able to measure the on portion of waveforms in the power meter more effectively.

1 Like

Sorry to hear that, @oasis.

As @kevin1 mentioned, Sense has a 120-day return policy. I know you mentioned you received your Sense via your solar installer so a return isn’t an option for you, but please connect with me privately if you had a particular question that I can help with.

1 Like

I am not lost in any weeds. I am a published researcher and computer systems analyst and have developed systems to generate faces starting with nothing but a skull to identify found human remains. So I am quite familiar with AI systems, and their development

It in fact would be a simple matter to allow a user to click on a wattage flag that pops up on the meter display and allow that user to identify that spike. It would be equally easy to segregate that as user supplied data. How that designation is applied in the crowdsourced data stream is up to Sense developers. But such a change in the APP would make this system a lot more useful to people a lot faster.

THe system already does the reverse of this now. Those who are lucky enough to actually get devices detected are sometimes asked to verify that device ID to the system. If you have smart plugs; you identify the device attached to the plug yourself, long before Sense even tries to make an ID.

Many here respond to me as though I am alone in wanting this; and that is far from the truth. IF you look at the reviews from Amazon for this device there are a lot of people who are asking for this same feature. There are many on this forum also asking for this. that fact should say something to the developers.

Not disagreeing with you on the last point - lots of people want it. I do too.

And you could certainly have a UI that lets a user manually label each spike Sense highlights with a power reading (as I suggested). But what good does that do ? I don’t believe users could label on/off signature pairs accurately, except in constrained situations (most other things turned off in the house), by isolating just that on-off cycle. That doesn’t get you anywhere though, since it is a one-off. How would you handle labeling the simple situation in my house below ?

Or are you just assuming once you identify the easy case, Sense would auto-identify the hard cases with noise and other similar spikes popping up and down all over ? Surely you don’t believe that ?

No I do not believe that Sense will identify the hard stuff if I do the simple ones. What I do believe is that this device detection has been hyped WAY beyond the actual state of the art. What I do believe is that for a simpler home like mine; when most stuff is off all day except when I turn it on; that I can accurately identify most devices that matter. To be certain there are some things I can see on the meter display that Ithey can’t identify; but those are small automatic devices and there are not many of them.

I think part of the issue is that I do have a solar installation. During the day the inverter uses power that rises and falls precisely with the amount of power being generated by the solar panels. The sense unit has a separate pair of sensors clamped around the solar circuit but not around the part of the panel that supplies the power to run the inverter. So I do not think it will ever be able to detect the power overhead from running the inverter; even though it is significant and is precious parallel to the generation cycle. Generally running in the 1,200 watt range.

All I am saying is that that this device would be far more useful it it worked like it says it will on “the box”. There Is nothing there about waiting months for devices to be detected. There is a lot on this forum and in the directions about going to the trouble of adding all the devices in your house to the device list, by make and models To assist in device recognition. I did that the first day.

Now I see a response the other day from one of the Sense guys on this forum that the data I entered is not actually being used by my device to identify things. It is being used in the crowdsourced data system they are building that “someday” will improve device detection. No place on the outside of the box does it say that before this thing can work as advertised that I have to join a crowdsourced development program and wait months to see if it can detect my equipment. I am happy for those that have this working ok for them; but it is not what it was advertised to be. All I am doing is suggesting a fix for that.

Now you may not find the ability to identify devices manually to be of use to you; but I would find it to be very usefull. SO apparently would many others. Just yesterday I got a message from Sense that I should set a notification to let me know if my fridge stops working. Problem is, Sense has no clue about identifying my fridge, and I have no way to tell it.

I still find it to be useful; but I also find it to be severely crippled by the rigid insistence that no manual identification system for devices will be added in light of the lack of maturity exhibited by this technology. So we are FORCED to be part of a development program that we were not informed of before we bought this thing

Non-intuitively, these things are not mutually dependent, or mutually exclusive for that matter. A “failing fridge” signature can be detected regardless of whether a working fridge signature has been defined as a device. That would be my assumption anyway.

I disagree. I do not honk it can tell you a device that it has not identified is failing. I suppose I could tell by watching the meter myself, but how would the Sense unit know its failing of it does not know what it is.

Thanks for the explanation… I agree 100% that some of the Sense marketing is over the top on setting the wrong expectation. I’m hold the line on manual identification because so vehemently because the two other devices I looked at before choosing Sense, both had a “manual mode” and both are no longer in this business. There are also a couple of “refugees” from those previous devices roaming around on this forum and share their stories occasionally.

I’m also a solar user and love the fact that Sense brings the solar output together with total usage, plus I’ve seen somewhat better identification than you may have so far (I’ve had 2 units over 3 1/2 years). Not quite sure what you think the issue is with your solar install. I only have a single backfeed from my solar inverter to my main panel, so I never see the energy used by my inverter, since it is upstream from the backfeed (except at night when the inverter pulls a in 2-3W of grid power in night mode).

2 Likes

I wonder if the issue could be that your devices are just not providing Sense with enough on/off cycles to be detected.

I don’t have any details on how the Sense algorithms work, but my understanding is that they are using the same approach that would be used to solve the “cocktail party” problem in machine learning which is picking out a single voice in a crowded room of everyone talking at the same time. The single voice for Sense though is a particular device.

I’m not an expert at machine learning, but I am a software developer and currently working on a project that uses machine learning (it automatically rotates engineering drawings to the correct orientation). The approach I am using in my project is called “supervised learning” where I label this drawing as rotated 90 degrees, that one as 270, this one at 0, etc. After labeling thousands of drawings, I can then have my system interpolate across a number of axes so that it can attempt to classify drawings it has never “seen” before; that is, drawings that were not part of it’s training set. It is working really well.

The cocktail party problem, and I believe Sense, use “unsupervised learning” where rather than labeling training data and then categorizing new data into those labeled buckets, it looks for patterns that emerge by separating the data points into logical groups based on their “distance” to each other. Then, you can label these groups.

One way to solve the cocktail party problem is to use multiple microphones at different locations: You can then analyze the data from the different microphones and look for patterns. There is only one set of mains Sense can monitor so instead of looking for patterns in space, they look for patterns in time. I think that’s where the “give it more time” mantra comes from. It needs to see enough on/off cycles to separate the data into logical groups (i.e. devices) which you can then label as coffee maker, fridge, etc.

Again, I don’t really know how the Sense algorithms work and am not a machine learning expert. This is just my best guess based on my research and current understanding.

3 Likes

Nice explanation @ramon.

And sounds like a fun project.

In Sense sampling frequency terms (important!) and what is sampled (more than just a 1-dimensional amplitude), I’m going to draw a somewhat pointless picture:

In your project, imagine if the drawings were all reduced to 1 pixel. Now everything is the right way around. Problem solved!

Our system here is very simple. @romon commented that perhaps I have not had sufficient on/off cycles of our devices for it to find them. If that is true then I am indeed going to have a long wait. I have read a lot of the materials on this forum and the Sense website explaining why manual intervention will not work. Every excuse they give is based on their desire to perfect the detection System. Basically they have placed the development of the device ahead of the needs of the user community.

While this may fit a technical development model; is is a poor user support model.

I constantly see people falling back on the claim that I could not possibly mark a device in the meter at the millisecond level. I would not plan to even try. This ignores the fact that I do not have to. What I have proposed from the start is to let the Sense equipment do the detecting. Just as it does now with one exception. When it puts up a flag in the meter as it does now; I should be able to click on that flag and tell sense what that device is; if I know; even if I have to do it again when the device turns off.

That identification would in fact be in context; noise and all; and could then be used by Sense to ask me later when it thinks it has seen that device again. I would not be the one identifying when the device comes on an off; Sense would be doing that. All I would be doing is giving it some idea of what that might be. Too be fair; in this scenario I should have already entered the device in the listings for my home as I have.

As to polluting the data base these guys are developing; that would not happen if they recognize and isolate those identifications until they prove out over time. That is far easier than they claim in these forums. Meanwhile, I at least have a chance to reduce this unknown bubble and increase the number of known devices.

I seem to have to keep repeating this. I like the idea of a device like sense. I am enjoying using the one I have. But It is far less useful to me if it is clueless as to what I have in my home. At this juncture it is no better than many other monitors that cost far less and are marketed with the abilities they actually possess.; not those they they say it will have some day.

I should also say that my home is not complex. We have a breaker box that has only one side, so two sided detection is not an issue. There are only two of us here, and things get turned on and off a lot but individually. So the cocktail party analogy is far less applicable to us than it might be to others. Though I recognize the issue, and what I propose might help with that as the Identification would be in context.

I am not suggesting a “teaching” system where users turn on single devices and identify them. I agree that this would be counterproductive. I am suggesting allowing users to identify devices in full context of normal use.

The take away for the Sense folks from all of this should be this. They need to either provide some way for people to manually identify a device that Sense can see going on and off; or they need to explain on the outside of the box that SOMEDAY this device might be able to detect devices on its own; but that the technology is not yet mature.

1 Like

Thanks @oasis,

I understand what you are suggesting. But I have two questions based on your explanation and my picture below.

  • How the heck could someone with my Power Meter picture above know what the +70W device is in the midst of all the other activities ? People have trouble figuring out many of the Sense-detected devices even after Sense has given plenty of additional clues. I guess I could identify a few of the big ones if they weren’t already detected, like the +1206W. And would I need to identify both the on and off signatures ?

  • You suggest that once identified once by you, “it would then ask you and could then be used by Sense to ask me later when it thinks it has seen that device again.” How would this “thinking” mechanism work if isn’t machine learning ?

I think we both want the same thing - better and faster identification. But I have done some proactive things to enhance identification - bought some smart plugs for devices that Sense is unlikely to identify, used one smart plug as a traveller to find troublemaker devices like the one you are searching for below. I agree that the marketing was too aggressive, but I actually have found ways to do what I want and need out of Sense in spite of some of its limitations.

1 Like

The Sense equipment already does half of what I am suggesting. It SEES the device come on and go off and even tells me with a flag how many watts it saw.

For some context here. I am now over a month into this (don’t say it; I am sick of being told to wait). So far Sense detected on its own a coffee maker that has now been replaced. That’s it. I am not talking about some 70 watt light bulb here. I think I will know when I turn on the washing machine and can watch for the detection of that event. I can already see every time the water heater comes on and goes off and so does Sense; it just has no clue what it is. That is the same story with all the devices in my home that are not Hue lights or attached to a smart plug where I could name the device.

I do not have to sort through a complex meter as you have pictured; but I could if I need to. My Sense device clearly shows with a flag when things go on and off. In almost every case I can detect the bounds of the device as things here rarely go on and off in multiples. That said I do not expect this device to detect EVERYTHING. And frankly I don’t care about everything. But I care very much about devices that consume sufficient power to punch my wattage above my generation capacity; or that are critical like the fridge.

I also do not expect to have Sense be 100% correct on subsequent detection of the devices I can identify. That is the part of this that should actually; take some time for Sense to figure out; not the initial offer of a question as to what something it can see might be.

You are correct. Sense does see something and flag it, but you should think of each of those as a “transition of interest”, not a detection. It doesn’t become a detection until two things happen.

  1. It sees enough “transitions of interest” fall into a cluster, based on features of that transition, in a slice of a multi-dimensional space (at least 17 dimensions).

  2. It can pair two opposite “transitions of interest” together, an on and an off. That may actually require some kind of causality.

So you probably could feed Sense by labeling the figged power increases and decreases, but you might need to it enough to provide reliable clustering, maybe a few hundred pairs.

I cant disagree with that because I am not privy to the software of the device. But what I propose is still valid. Even if I only identify the first flag and then allow Sense to do the pairing it could still ask me when it sees that spike again if it is the same device.

But the truth here is simply this. The developers of the device are more interested in the inflexible development model they have then that are responding to the huge volume of user requests for a way to make this device ACTUALLY be what they told us it already was. I am not blaming them. I saw this same sort of inflexibility when the FBI was developing facial recognition systems. Once they got past the single track approach they were using they succeeded. These guys could to if they let go of the idea that users should be force to help them develop the system and set them free to use the device in a bifurcated development path.

1 Like

That’s a very weird solar system, with an inverter drawing power from the mains. Every solar system I’m familiar with self-powers the inverter (before the Sense solar clamps), so what you measure is net gain. My system uses micro-inverters, so each panel does its own thing separately. That costs a bit more than one central inverter, but also isolates individual panels so that a shadow on one or more panels doesn’t drag down the whole array.

1 Like

It is not really drawing power from the mains as it only uses power when it can generate power. So while there is a breaker in the panel that feeds the inverter, in operation the system never uses more than it generates so the power is still coming from the solar system.

The panel on our house is a single stack panel with only one row of breakers. So this limits where the Sense device can clamp on. At present the device is attached at the top where the mains power enters and at the bottom when the solar enters. That is just as the Sense instructions show.

Either way the inverter must get power. If you want an accurate picture of what is happening you need to see that power overhead

I keep thinking about your question based on that very busy meter picture you provided; and also what I keep hearing from the Sense folks about the data the system requires.

It seems to me tht the way the Sense unit is designed it has way too much data rather then not enough. It is possible to flood an AI system with data to the point that it just has way more than it needs.

Perhaps what the system needs is a “tuner” that controls how much data the system actually looks at. FOr example. I see lots of 3, 4, 5 and 10 Wats things going on and off in my system. Almost all of those are LED lights. The system already knows what those are because it can access my Hue lighting data. SO suppose I could adjust the system to ignore fluctuations below a certain value so it could concentrate on just the larger power items.

Moreover, there is really no need for such a high sample rate for the data. We are dealing with a nominal 120/240 volt 60 Hz system. A sample rate of 120hz would be adequate to gather all the data required, but to be certain perhaps over sampling at something like 1Khz would give lots more data than you need. In a 60hz system there is not much going on between cycles.

If you reduce the “noise” data collected that the Sense unit must look through you will have reduced what it has to look at to only the more significant changes. In fact it is pretty clear the the Sense unit is looking at wattage changes at the sub watt level as some of my devices at idle only use about .5 watts and the system can report that. For detection of installed devices that level of data is not really required.

In any case I do not expect this thing to b perfect; I only expect it to do what I was told it would do by the marketing. I am happy to help develop the AI for the system; but in the meantime those who are trying to use this need some relief.

Thanks for considering the screenshot of my home’s electronic pulse …

I think you are right - there probably is more smart “tuning” Sense could do to squeeze out only the most important “transitions of interest”.

As for sampling, at a top level you are right - a 120Hz sample could be used to measure the power if three things were true:

  • Sense could align the sampling directly at the peak of the 120Hz power waveform (see below). But line frequency drifts so sampling requires continuous readjustment (see @dave’s cool posting that shows line frequency drift here)

  • Current was always in-phase with voltage - but it’s not. Motor and transformer displace current in one direction and capacitive loads pull the current waveform in the other, meaning real power (current x voltage) changes with different phase.

  • And Sense only needed power / current / voltage to do identification. But it needs information like phase angles and probably things like estimated relaxation constants for transitions. Those require detailed sampling. You’re just not seeing this data on the transitions because it’s hidden behind a slick UI.

The picture below show the simple case of AC power, when both current and voltage are in-phase. Things get much more complicated when:

  • Current and voltage are not in phase
  • Current is not a nice sine wave - all the time in real life.

image

I’m an electrical engineer, and I think Sense has done a good job hiding all the ugliness behind complex AC measurements that I once had to do in lab. They may have made it so simple that users don’t even realize how hard AC measurements of power really are.

1 Like

I understand what you are saying but I really think that you are adding layers of unnecessary complication with this approach. We face a lot of the same issues with facial recognition. This was especially true in picking out faces in large moving crowds. What we discovered was that we were giving the system too much data. TO see a face the system does not need a lot of specifics that might individualize a face to pick it out of a crowd for identification. It only needs that level of data once is can see a face and is asked to identify a particular person or count how many individuals it can see.

You have the same problem. Sense needs to see and identify that there is a device before it can possibly individualize that device. Only then would it need all this individualizing information and data that you are describing. But in your approach you are asking it to do both step all at once. So naturally it spends weeks/months/ maybe even years gather this data. Because the graininess is so fine; it may never see enough correlations to decide on what it has found. You are asking it to do too much.

If you simplify this to a multi step process it would work much better. First identify the bounds of a paired on/off event. You might even wait to see that same sort of event more than once. But the point is first to identify that there is some kind of device. Then you can step it up a bit and either tell the user that you found something a let them identify it; or go to the next level and start examining the details of that particular device signature.