Detection Challenges Underlying Sense's Approach

Not to go off-tangent here but, I see a lot of discussion on improving device detection with new/smarter algorithms. One must also think of the limits of such approach/technology. Per the CEO’s own words, detecting all devices is akin to being in a room, full of people conversing, and, via speech processing/recognition, identify each unique conversation. The task is truly non-achievable and one must think of the limits of the physics. A similar yet much simpler problem arose 2-3 decades ago when people wanted to improve the time-frequency resolution of wavelet processing … then came the idea that you can have either precise frequency estimation or precise time localization … not both. This is often referred to as Heisenberg’s Uncertainly Principle, at least in this context (for those of you trekkies, few centuries from now they’ll invent Heisenberg Compensators to run the transports/beamers). So, we must ask ourselves that, while for sure algorithms are continuously developing, by either ‘deterministic’ modeling and or AI/ML, realistically speaking, we will never be able to detect ALL devices in anyone’s house because, in the end, when you add all of those devices together, you end up with ‘noise’.

Thoughts/Comments ?

@drjb,
I think there is something else at work - how differentiated / unique are the transitions in your household. If Sense ends up seeing a spectrum across some set of transition features instead of clearly defined clusters, it’s hard to do an identification. My house has a bunch of things that are clearly detected, most often the larger old-school device (single stage AC, vacuum, furnace fans, etc), but a large population of smaller transitions that are not… Here’s my theory:

1 Like

Fully agree @kevin1. But, isn’t that a bit of a limitation in itself? Sense’s algorithms are based on transitions … What if multiple devices fire-up at the same time? Unless one assumes that very rarely this happens (which is very reasonable). This also means I need to adjust the on/off cycles of my 3 grow lights and have them staggered over few minutes, not at exactly the same time. I use those to provide some ‘light’ to the plants that require it, during the cold Illinois winter.

I’m not at all criticizing Sense’s approach, I am instead commenting on how complicated the task truly is, so that we set our expectations at the correct level … and not expect Sense to detect every single device in my or your household.

All in all I’m still fascinated by the technology and how innovative it is. I’m hoping though, in some near future, all the devices we use at home will be able to broadcast their energy consumption and other useful data, and even provide some level of automation/priority. I can load and set my washer to start the cycle not after I press a button, but when electricity is cheapest or whatever other criterion I could come up with. We’re not there yet, but we could see a lot of ‘optimized’ consumption.

@drjb, focusing in on a transition-based approach, and even more specifically, a half-second or so transition-based approach, is definitely a limitation, but really the only way to deliver a working product. without going to a CT per circuit type solution. I would guess that “collisions” between half-second-long transitions are relatively rare unless the transitions are correlated (connected to the same switch, thermostat, etc.). I think the real b beauty of Sense is that they are beginning to expand out in to a broader device-centric hybrid approach.

  • Native detection of half-second type on and off transitions
  • Specialized detection of large, slower ramps - EVs and mini-splits
  • Smartplug integration for plug-in 120V devices and clusters of device that are used together
  • Direct circuit monitoring (DCM) for wired 120V and 240V devices
  • Hue-type hub integration for home automation that is capable of providing accurate power information about some or all of the devices connected to he hub.
  • Other smart device integrations like the Ecobee Historic, which supplies historic ground-truth data to improve models for HVAC, EV charging, etc.
1 Like

@kevin1 agreed. Though I’m not sure the one CT per circuit is a solution either. I have 24 circuits/breakers in my box and there is already a product on the market with up to 16 CTs (look up Amazon). Still, that does not resolve the problem as, what if you have multiple devices in the kitchen circuit? Better yet, how about the family room circuit? In my specific case, 24 devices in there, and a CT by itself would only give me the consumption in the family room, not per device. That is why I like Sense’s approach i.e., including AI/ML and looking at signatures/transitions. Granted, one can think of Sense + 16 CTs, but that quickly moves from a hobby device to a pro-grade. Given that most decent-size houses use a 40-breaker box, and assuming one needs to monitor let’s say 16-20 circuits, that adds-up very quickly to a multi-channel data acquisition board. In my line of work, such devices cost about $1k/channel. Granted, those a pro-grade, but even if we use hobby-grade components, 16 channels is very taxing, need a 4MHz × 16 = 64MHz clock, … let alone the onboard processing power and bandwidth upload requirements. When all is said and done, the Sense device offers a very good introduction to this ‘hobby’ and one needs to appreciate what it’s capable of. Cheers

PS. Not at all lecturing here, love the product and love to ‘understand’ how it works and talk about it. Already showed it to 2 friends and they’re very interested.

All-

I wouldn’t just focus on processing of the data alone. Don’t forget that CTs & PTs play Main role in capturing the waveforms which then being processed by a chip. Alone sampling increase is not going to get us anywhere close.

We often specify Schneider Electric PowerLogic PM8000 series with class 0.2 CTs or POWER ION Monitors for capturing Harmonics and other power quality related issues that present in the PG&E infrastructure at industrial levels. which captures only 264 samples a cycle which is far lower than what you all are talking about although, it does tell us everything we need to know. Yes, it is not going to detect anything but for sure but it does answers majority of the power quality related issues.

Again, sense is built on a different principle in mind which is sensing current waveform and then detect device based on load’s unique signatures. It doesn’t care much about power quality however one can not ignore the power quality when they talk about sensing device because Power quality (harmonics, sags, zero sequence currents, imbalance in the current, Power factor, reactive power, etc.) is the main reason why Sense is having hard time finding each device smoothly. With other noise present (Electronics loads), each waveform shape changes so do the signature of each device which sense rely upon! Which then puts sense at the cross roads on which way to go.

I believe extensive lab testing, data science analyses and utilizing Relay grade power meter with sense and compare both to verify what exactly is useful and how sense process the info vs Relay grade energy meter. Sense is really a great tool IMO but certainly it can take this whole game at different level if done right.

@Manny1314, very interesting thoughts. Sense has commented on the problems of “noise” in the past, though in their books, noise for Sense seems to be more related to “noisy devices” that create large transients within your household and make it hard to spot smaller, less frequent transitions that indicate on and off switches. One device that they have identified in the past as a noise generator is a plasma TV.

I do think that transition waveform variation plays a role in making detection more difficult, but mainly because it helps cause clustering “smear” when Sense tries to cluster transitions in their feature space.

@drjb,
Absolutely agree with you assessment of CT per circuit products - they don’t do what I want i separating out devices. Plus I have over 56 circuit spread over 3 panels - that would cost a pretty penny to instrument on a per circuit basis and I still wouldn’t get what I want. I have purchased a second Sense with Flex sensors so I could isolate a couple 240V devices that weren’t being picked up as reliably or completely as I would have liked. I couldn’t use the Flex sensors on my primary Sense because I already have solar. But with a few workarounds, I was able to get the second Sense operation and using the Flex sensors for two wire 240V devices. It would be nice to be able to repurpose the main CTs on the second Sense for DCM as well, but that doesn’t seem possible yet.

This is the first time i’m seeing the amount of circuits you have for your home, wow.

I was wrong - forgot to count the floor heat subpanel.
14 240V circuits (including solar backfeed, dedicated Sense breaker, & feeder to floor heat subpanel), 49 120V circuits. That’s 63 in all.

2 Likes

I disagree with the point about it being non-achievable. Consider facial recognition technology - unthought of a few years back, now it can pick out an individual out of thousands of faces in an airport. Sure its not 100% but I don’t need 100%, I’m interested in the “big hitters” not every light bulb.

I’m not privy to how it all works but its hard to believe that the user base cannot offer some assistance. I would think that in the background, Sense is building profiles of things that don’t match the other things on the sensors. Then it spends considerable time validating what the thing is that it has characterized. What if Sense exposed those (potential) items and the users could tag them with what it is (or is believed to be).

Yes, it’s a brute force approach thats more like a criminal line up - here are these 8 “criminals” - accuser: which one do your recognize as the assailant? In the Sense case it could be: here is a unique profile that we think might be a dishwasher - would you like to temporarily tag it as such while we validate?

I also understand it is tricky picking a profile out of all the things Sense is seeing. But if I recall a (Sense) video I watched on this, there was a discussion of 240v loads and how a load might be detected on one leg and then it goes and checks the other leg to determine if its a 240v device. So, if Sense can read each leg independently (a big if, I’ll grant) would it not be beneficial to modify the list of manual device entries to indicate which of the 240v legs the device was associated with (the sensors would need to be labeled A & B)? Theoretically, that would cut half the “noise” out of the picture for a given data capture cycle. Using Sense’s own terminology - if you were picking one person speaking out of a room of 100, now you would be picking out of a room of only 50. So I think its all about the AI perspectcive and collecting sufficient data to characterize a particular profile (and its variants). Difficult but not totally unacheivable.

Just an opinion and as such, worth all of 2 cents.

1 Like

@theojt,
I agree that machine learning for energy disaggregation (what Sense does) is on the path to fast advancement , like image recognition was 10-15 years ago, I do thing there are still some fundamental limitations to style of native detection that Sense is doing… Things like:

  1. Always On devices or components of devices, especially one that almost never change usage or barely change. No change = nothing to detect. I have a bunch of these in my house, mostly electronics. For example, the radio / electronics portion of my garage door. For these we’re probably always going to need some kind of smartplug assist.

  2. Electronics with power supplies - for these, the power supply hides the function and characteristic patterns of the underlying device. Be it a DC motor in a garage door or a computer, the power supply homogenizes the usage pattern and spreads it out in time. Once again, we might have to rely on smart plug assist.

  3. Electronics that are usage / media dependent - And even if Sense could get much better at detecting different kinds of power patterns through power supplies, things like OLED TVs and even washing machines have different power usage patterns depending on what’s playing on the TV and how heavy/dirty the laundry is.

I think that the home inventory might be able to help over the long haul, most importantly teaching Sense what kinds of things in the customer base they are missing most frequently, so they can take aim at those devices.

One thing I’ve learned over the years is that of product performance variability, arising from manufacturing variability, material properties, tolerance stackup, etc. After all, things like Monte-Carlo Simulations were investigated and used heavily early on in the electronics industry to assess the spread/variability across supposedly identical electronic circuitry. Two machines coming from the same assembly line will have different sound and other signatures. I bet it’s the same for the electrical signatures Sense is working hard at detecting. If we take 100 dishwashers, fill them up with the same exact dishes, and run the same cycle, we’ll see small variations in their electrical signatures. This to say that, there is a lot of statistics and hypothesis testing and the like in ML/AI. One fundamental fact is that power systems are still 100% ANALOG devices i.e., prone to infinite variation in amplitude (unlike digital signals). The signals/signatures Sense is trying to identify are not fully deterministic. Some parts of such signatures are deterministic/predictable, others are stochastic/random. As such, there is an inherent amount of spread/variability and we’re faced (once again) with a typical problem of SNR (Signal-To-Noise Ratio) and that is where one must appreciate how difficult all of this is, and ultimately accepts that no sophisticated piece of software (today or 20 years from now) will be able to identify ALL conversations in that room with 100 people. Many people have worked on extracting signals from noise (Wiener, Kalman, …) but ultimately there is a floor/minimum that we can’t go beyond.

I recall a debate I had with a colleague many years ago. He was convinced that with very powerful/sophisticated computers, one day we’ll be able to predict the weather for months in advance. This is absolutely ‘rubbish’ as one cannot ignore the inherent randomness built-in the many physical processes in life, electrical signatures included.

My 0.02 :slight_smile: … Very interesting discussion though.

1 Like

I have to agree with this - smartplugs will be the short-term solution, no doubt. But - I hate the whole “wallwart” approach. I’m hoping that Sense will soon certify a manufacturer like TopGreener (or some company like them) that has a whole suite of energy monitoring point solutions, including the most important to me - energy monitoring RECEPTACLES. More expensive than Smartplugs for sure; I would make the investment however. The things that I would want to monitor the most could then be addressed by a combination of receptacles and Flex CT sensors.

@theojt I’m with you on this. However there is another approach. With many of our devices at home becoming smart (e.g. the oven I just bought) I ‘dream’ of seeing the following:

  1. Control/Power Panels (in the basement) that measure/track/broadcast somehow the energy each circuit uses. This might take some time as it’s very difficult to change ‘standards’ that have been in place for many decades, including safety as well.

  2. Smart devices (refrigerator, oven, everything) that measure/broadcast their own power usage.

I’m not too hopeful for #1 as it’s likely difficult for multiple reasons. For #2 however, all that is needed is a small chip to measure current consumption. If TP-Link can make a smart energy/monitoring plug for $20, then Whirlpool/LG/Samsung can incorporate such chip for a fraction of a dollar in their devices.

Of course one also needs to implement ways to ensure manufacturers are ‘honest’ in their reporting. We don’t want them to advertise lower numbers than their devices actually consume. The industry and history are full of such deceptive practices.

For all of this to happen, one also must resolve the issue of cloud/private networks. With more and more people opposed to having their data up in clouds, we’ll see a lot of push-back against smartER devices. I was talking to a friend last week about Sense and he was adamant he did not want another device spying on his habits at home. Complicated, and many ‘opinions’ for sure.

In any event, I hope the recently formed Smart Home group (Google/Amazon/Zigbee/…) comes up with something along these lines… in my lifetime… and then I’d have to buy everything new again :slight_smile:

A little story I remember …and not to digress too much. Back in 95, I lived in a graduate residence at MIT. Then, the school invested in key-activated door locks for the building as a safety measure. This was very convenient as there was no longer a need to carry multiple keys. My student ID doubled as an access key. Many of the residents objected as, per their view, this provided the school with the means to spy on their home in/out habits.

drjb… This is a thought that I have had for sometime. Thanks for airing it… Sure others have come up with this. Put this right into the circuit breaker. Even add a little LED or LCD display to show the current current reading… or whatever else …watts… volts… one might choose… Just like the info the TopGreener is broadcasting…With Sense’s parent company… Someone there is already working on it… Circuit breakers have come a long way with GFCI and AFCI built in… Got to have technical obsolescent to paid for that tooling…Gerry my 2 cents…

1 Like

This actually does exist now. They are expensive and still in early stages, but we are on the way. The trick for Sense may be that the current main product in this space is Leviton, and Sense’s large investment parter is Schneider. But I would imagine part of Schneiders investment in Sense is to be able to create their own smart load center or other products.

Edit to add - I believe there are a few Sense users who have actually installed one of these.

A couple discussions exist in the forums and can be found here

2 Likes

That’s true.
One of the biggest company is Atom. I have been following this company for a long time. Check out their website. If you Google solid state transformer, solid state breaker, etc. You will see many competition out there already. Although some of these product are still in very early stage however it’s certainly doable.

I love the fact that these are so fast, virtually no one will get shock due to its extremely fast over current protection capabilities. We always think about safety when It comes to industrial power system.!

Screenshot_20201219-082105_Chrome|243x500

Glad to see the industry is already thinking along this way (and me proven wrong). When I said ‘difficult’, I was thinking of retro-fit breakers that fit in the current Siemens/SquareD panels. I just replaced my entire box and it is not cheap. My thinking is, to have smart panels (complete new design) installed would only make sense for newly-built homes (assuming the owners are willing to pay the premium). However, if one is to target current homes, one needs a retrofit solution, as no home-owner is going to replace the entire box just for the fun of it.

I recall a couple of products in the tire industry, about 20 years ago. The run-flat tire PACS by Michelin, and the magnetized Intelligent Tire by Continental. Both of them were very ‘ingenious’ but required special tooling to be installed/repaired. Granted, they could very well be installed on new vehicles (PACS was installed on some Honda Odysseys), but doing so meant all tire dealers out there would need to invest in new tooling, and that was not a realistic/affordable option.

So again, and in my opinion, if Smart Panels are the future (for the masses), one needs a retro-fit solution where essentially all that is needed is to swap the breakers. Many of us can do this easily, and change only few at a time, based on need and budget.

1 Like

drjb… Ref the run flat… Remember the Cadillac that could run ‘overheated’ without coolant / water pump…I remember the advertizing back maybe the 80’s maybe… They were saying driving from Detroit to Toledo… No coolant…Sorry for the off topic… Dido on the rest… Love hi tech… Gerry