Disappointed in device discovery

I need to contact support about it. I change the name on each detection by adding three questions marks to the proposed name and flip the “this is a guess” slider to off. So I think there have been a few renamed that have gone missing.
I know for sure there were four I did not rename. They all four were detected at the same time and gone the next day. I didn’t get the notice sense has sometimes given me about a device being removed with these.
Some of these happened during the time of renaming difficulty, which is still happening.

Well, this thread was an eye-opener. I started on the 11th of May and the first ~3 weeks I got several detections. Every other day I would wake and see something new.

But now it’s been two weeks since the last device was detected. I do wish there was a way to see that something else is brewing. Some old posts show under Device Detection area of the Sense Monitor in the app that it lists items currently being learned. I guess this feature was removed, probably caused more support calls than it was worth.

I am still not seeing my Keurig brewing, just when it keeps the water hot for the next brewing. No stove, oven, dishwasher, washing machine seen, and even the dryer seems to be conflated with the hot tub. I had tried to use the TP-Links to figure out some devices, but that only helped me see them, they were still not being Sense’d :slight_smile:

I did watch as my hot tub was detected and after several days of it assuming hot tubs are 2000 watts it finally figured out that ours was 4000 watts and would detect the whole thing. But when I press the cleaning button and the heater AND the pumps come on it doesn’t detect that as the tub at all. But that’s not an everyday thing so I’ll be patient.

I know our Plasma can make detecting things difficult so I have taken to running the dishwasher overnight but still no luck.

I guess I am still confused about the process Sense uses for learning. Elon talks about an AI that is seated in front of a video game with no rules and within a month it beats the game every time, a month later and it can beat any game put in front of it. He expects Teslas to be able to soon cross the line and be better drivers than an average human, but VERY quickly to be many times better than any human. How does that relate to the Sense AI?

Or is it just Sense ML?

Does Sense Machine Learning actually gain those same exponential improvements with exposure to more data? If so I would think in the not too distant future it would be able to see patterns to make device discoveries even thru my noisy Plasma TVs.

It worries me to hear others being told that their environment is just not right for Sense.

Machine-solved problems, like human-solved ones, are limited and defined by their inputs and then the intelligence applied to those inputs.

If you didn’t learn Norwegian as a kid and you hear Norwegian there’s nothing to stop you learning it if you apply your intelligence. But if you are deaf, unless the Norwegian speech hitting your ears is translated into some other form of input, you’ll never learn … or to put it another way, a deaf person will never learn a language by listening because there is no input.

When born-blind people are given eyesight for the first time they usually report not understanding the input: light! Why would they? Until something moves there’s really nothing there to see that will have any kind of corollary in a blind persons well-formed brain.

Sense can only learn based upon its input … which is currently very specific and limited.
In Elon’s example game space the i/o allows for “easy” inference of the rules … the “world” is encompassed by the game so a machine can easily best a human.

As a counter example, imagine a game where you just pull a number out of thin air and say it to the other “players”. I attempt to play this sometimes with friends. The rule is that you respond with another number and then state if you feel like you won. If you don’t feel like you won, that maybe you actually feel like you lose, then you say so. And the game continues. Whether you justify whether you win or lose is up to you and the players. It’s more fun with people who have a sense of humor. I challenge any AI to compete with a human in that game. It’s like the Turing test for machines that can also be believably insane. One response that might start to convince me of a worthy challenge would be along the lines of the machine starting to sing a drunken song as did HAL in 2001.

Ironically, Elon has a significantly more science-fictional view of reality than Kubrick.

[A final thought experiment:
A house with many lamps; all with identical bulbs.
Perhaps, though inference, a machine can work out which of lamps are switched on and off and when – “it’s bed time, that was probably the bed lamp” – but will it ever be able to tell that perhaps some of the light bulbs were swapped around. Would you?]

1 Like

I was really looking for a sense :slight_smile: of how I could expect going forward I would interact with my Sense. Perhaps after a time everyone just waits for the monthly recap. I was motivated and took my panel apart, moving the Sense module to the breaker my router is on so I could turn off the entire house and go breaker by breaker looking for my always on.

So, of course, I wake up and find it has finally identified my Dishwasher and Oven. Not sure if it was at all related to the power cycling, and oddly my Always On value dropped in the process from 420 to -11kW. Guess I need to wait another day for it to have 48 hours of some things always being on.

I would put it this way - different machine learning problems require different machine learning network structures and associated data sets. Machine vision and autonomous driving machine learning have had thousands of researchers working for perhaps 10 years to to build comprehensive datasets (for training and testing) and to devise the most efficient and accurate types of networks for each… The Sense flavor of machine learning (power disaggregation or non-intrusive land monitoring - NILM) hasn’t had comparable resources targeted at it just yet, and most development has been against a diffuse bunch of different datasets. Sense is the one company that seems to have the ability to put together a comprehensive dataset, at least for the North American market.

The following examples might give you an idea of how the breakthroughs in image recognition and autonomous driving happened. It’s been a combination of collaborative datasets, competitions for the best results, plus accumulation of massive datasets. The same is likely to happen one time with Sense, but Elon is a little over-optimistic about how the process happens…

The Imagenet Challenge - competition to create the best neural network for vision.

Imagenet - the real breakthrough - a comprehensive dataset.

The race is on in autonomous driving. There are many public and private datasets as well as many competing models (neural network architectures). The winner is likely to be the guys who have the best datasets and are able to make use of academic competitions to create the best proprietary models. I have done a little work with one of the models, Mask R-CNN. But we’re only about 3 years into the adventure.

One more comment… We’re still in the early going for machine learning in another dimension as well - hardware. Machine learning is still mainly run on traditional Von Neuman processors (Intel, AMD, ARM) as well as traditional GPUs (NVIDIA, AMD), and FPGAs, though it’s not a good fit for any of these for power or performance efficiency. We’re seeing new hardware, especially for machine learning training that substantially speeds up and reduces power usage during training, and inference (detection in Sense-land). We just have to realize that machine learning is very early in it’s life cycle…

Google’s specialized chips

Other specialized machine learning chips

Keep in mind that Sense does not use AI, it’s machine learning and worlds apart in how the two work.
I’ve gone a couple months without detections and others even a year or more. After a couple months of nothing, I got a couple this past week. Not good one with history or anything I’ve been able to locate just yet.
The long spells of zero detections are just part of the experience.
It is puzzling as to why. My Keurig is detected the opposite of yours. I have a detection for when I brew a cup but not for the backup heater.
Hang in there

1 Like

In several years, they have not been able to add something as simple as charging 9 cents per KW for the first 1,000 KWh then 12 cents per KW for additional use. A crappy programmer and a crappy tester could implement that in a day. How do we know if they are even working on improving their ML algorithms and code?

This and may other “this simple thing could be added” comes up all the time.
It is never that simple. Sure, an adjusted calculator like you suggested may be “easy”. But what about the more complicated version. People frequently ask for Time of Use… Now time of use with seasons. Ok, how about time of use, with season, with usage tiers. Those are much more difficult. Not just in the math, but how is the UI managed. What assumptions do they have to make. How many different slight variations of the above exist. How do you build to allow for all of those.

Then they make some of those decisions an implement some form of a new kWh $$ calculator, but what other versions didn’t they take into consideration. Now they are hearing from customers who are un happy because they took care of use case X, but not Y or Z. What is important to you may not be to someone else. Yes, there seem to be a lot of people who have TOU or tiered plans, but there are plenty that don’t. As said by the Sense team many many times, they are a small company. They have to prioritize resources and focus on features that impact the product and matter for the most people and the core product.

THE feature that Sense is marketed for is device detection and device usage. The money calc is just a handy sub feature.

I do not work for Sense, but I do own and run a SaaS platform with a growing client base and go through these development hurdles on a daily basis. Do we add the feature that client X is asking for? It is a simple code change, but what are the repercussions. What will it lead to next.

In its simplest form, its the adult version of “If you give a mouse a cookie”.
It is not about never giving anything because it just grows, but you have to take into consideration what giving that thing means and be prepared for where it leads next.

3 Likes

First, as a new member, I highly recommend you read the Community Guidelines.

Second, in 2018, we increased the number of accurate device detections in homes by around 50%. On top of that, we introduced Community Names, smart plugs, broader EV detection, major refinements to how Sense refines its models and much, much more and we continue to push out model updates and improve our ML infrastructure constantly. I’d say that counts as “working on improving our ML algorithms and code.”

Third — For tiered and ToU support, as has been noted by us and others, it’s far from simple. For ToU, there’s a pretty solid overview of that (including some deep diving by @kevin1) here: Time of use pricing - #4 by MachoDrone and @ben gives another overview directly above . Moreover, even if it were simple, many factors inform product dev decisions beyond just ease of implementation.

2 Likes

At six months and four days, I have 41 active devices. Most of mine were detected in the first few months. Compare the rate and timeframe to what users experienced a couple years ago and it’s clear they are constantly working on ML.

1 Like

Let me just add a simple caveat based on @samwooly1’s reply. Sense will work well in most homes and will detect a lot of devices. However, this is not necessarily true for every single home even if compatibility specs (200A, split-phase, good wi-fi signal, proper temp. range) are met. Sometimes other devices or utility line noise just get in the way and prevent good detection. For that reason, we offer a lengthy return period.

3 Likes

Yes indeed. In my case Sense engineering got involved and narrowed my problem down to a constant pressure (which uses both variable frequency and variable power) deep well pump, which creates “noise”. That not only prevented detecting the well (used for both my geothermal system and my domestic water) from being detected, but prevented reliable detection for most of the rest of my home.

Somethings are difficult, and some things are impossible.

Personally, I’d categorize multi-tier cost reporting as one of the least important features, because I (and many others) bought Sense to help us see where our power was going and that’s all about devices and KWH.

2 Likes

I would expect device detection to increase significantly in a year, even without any effort made to alter the algorithms, due to database population with new data.

I am from Texas where we chose power company on the web site powertochose.org. There are 3, 6, 12, 24, and 36 month contracts. I have always favored 6 & 12 month contracts. Hence, I have been thru quite a few electric companies to pick contracts every 6 months. Common: rate tiered by amount of usage. Rare is time of use plans.

Most people don’t know what a KWh energy unit is (most even confuse it with KW power unit). Thus, I think getting the simple usage tier would 1) be beneficial to FAR more people than the time of use tiers, 2) would benefit FAR more people than adding Alexa, Google Assistant, Hue, and the other things mentioned, 3) I see no reason to deprive the bulk of people from the usage tier simply because you are not addressing the more complex time of use tier. That would be like saying you are not going to implement Hue integration because you are not implementing LIFX!!!

You say people would complain if you implement usage tier but not time of use. Well, consider me complaining that you implemented Hue and not LIFX!!! lol

Bill, I’m not going to argue relative merits of TOU vs tiered vs Hue enhancements. But if you honestly want a quick and easy way to do tiered pricing calculations I can send you a spreadsheet or R code that does exactly that. Interested in being able to do your calculation today ?

One tricky part of the calculation for my utility, PG&E, is that my billing period doesn’t line up with a specific day within a months so I have to input a list of billing periods start dates as part of the calculation. How does it work for you ?

@billsspammail
I would t write off the possibility that this feature will never be available , as @andy stated, it has to do with the priority list.
I’m on flat pricing but thinking of making the switch to TOU so it’s a feature I’d like to see also. When you visit the product wishlist, you’ll see tones of features that we all ask for. There are things I thought should already be there and are not. For instance, I would like to view the timeline without internet. Looks like it would be very simple from my point of view but I’m trying to be patient and enjoy the great features it already has.
In my early weeks I felt as you do and was quite outspoken about it. It took me some time to settle in. Thanks to the community my overall experience with Sense is terrific. There are many things I feel NEED improvement. Voicing your thoughts and opinions here is a great way to eventually get these things implemented.

2 Likes

@RyanAtSense Hi Ryan, in a previous post you mentioned that utility line noise can affect Sense’s performance.

Can you elaborate on what utility line noise actually is?

I read Sense’s explanations regarding devices within the home being too “noisy”. So, I get that part.

What’s this utility line noise though?

1 Like

Perhaps the more proper term is “line quality.”

There are cases where frequent voltage or frequency fluctuations can hamper device detection.

There is all sorts of noise possible on incoming power lines and it’s highly dependent on the provider, distribution topology, neighborhood, age of systems, etc. “Noise” is caused anything from lightning, to equipment arcing, to high tension arcing, to other users of the local service. It varies from time to time.

See http://www.arrl.org/power-line-noise for one good discussion.