This reminds me of my favorite remote sensing story: it was determined that the cheapest, quickest, and most accurate solution to remote greenhouse temperature sensing was a large dial thermometer and a pair of binoculars.

Hi Andreas, great video! This is the developer of the AI-on-the-edge. The reason, why your analog counters are not detected correctly is indeed the inverted color (white on black). In contrast do the digits, this is not part of the training data yet. Have a look on the latest updates, now we can even detect the NaN values with then newest neural networks. Best regards, jomjol

I work as a data scientist and one of the most important questions you always have to ask before using AI is: Can this be done without AI? If the answer is yes it is probably more reliable to do in a traditional way.

Hello, I also have the AI solution for Gas and Watermeter and it works fine. But it took me some hours to do the correct setup. It it important, that the image is horizontal aligned and the scan area for the digital digits are large enough (larger than you wouId expect) , in 3:21 you see the 039 even includes the frames around the digits. This reduces the NaN results a lot and even recognized switching numbers. Also you must setup the initial value and a threshold to only overtake reasonable values (Values to low or big are thereby ignored). I mounted an angle iron on the wall 10 cm above the meter and taped the ESP32 below it. By twisting, you can move the glare point of the internal led to an area, which is not scanned, With this easy solution your energy provider can still read the values manually, if needed. Hope that help's to give it another try.

Going the optical character recognition route, as was initially attempted would require a substantial amount of system training. You provide the images and supply the CORRECT digits in each case. Then after a while it will be able to do it reliably on it's own.

As always great video! The metal piece from the water meter is for inductive sensors like TI LDC0851, I have worked designing this kind of sensors for iot devices

Thanks for video!! Good content, as usual. In my project, I used a coffee cup to mount the camera. I loved it, cheap and recycled.

to eliminate the glare try polarizing film either over the meter face or on an angle below the LED. You can get the film from a damaged LCD screen/ display.

Hahaha, let the MBAs do what already works… priceless! I’m sure your typical audience are mainly engineers so no one will take offence. Great video as always. Thanks!

Great video as always 👍😀 Great tricks Thanks for sharing your experience with all of us 👍😀

Lovely treasure boxes :) I'm affraid all we have similar number of goods at our workshops. Great idea with phototransistor. I've tried AI on the edge and after hours or experiments I have to chose another solution due to big difference between mine water counter and the example one. Great content! Appreciate your work very much.

Great, thanks for sharing. "We want more" is one of the main reasons I watch this channel

"An engineer's work always start when something does not work" Nice adage, thank you.

Thank you for the video. Getting interested in tensorflow now :) You can also use mqtt retain to store the counter value. Works great for me.

A very interesting video. The way to approach the sine wave to process it is brilliant. My water metter is very similar to the one from the project, but has some kind of wireless communication as this is a flat with central hot water (same goes for the heating). I wish I had enough knowledge to work it with the SDR radio to get the numbers in the same way I get the 433 weather sensors.

Excellent and attractive approaches ❤

Boy this sounds like fun, wish I had time to start another project.

Andreas another great instruction video. I enjoyed the water usage analysis. I got thinking, you should add some control to the project and slowly reduce the available hot water after a certain time frame. The idea to minimize water consumption. Results could be water savings but …. 😉😉

Interesting project again!

I made a project like that, reading 8bit segment digits from an aquarium meter. Used ESPcam to take pictures, then a raspberry with OpenCV to locate the digits and tweak the image, then tesseract, trained version for digits. It works nice in the end, i still get some wrong numbers but im able to filter then out.

I just love your sense of humor! I cracked up when you mentioned an engineers work begins....

The needle mechanisms are magnetically coupled to the turbine so you can for sure use hall effect sensor if you remove the lid and gears completely.

Cool project, I have something similar in mind for several months now but my ESP32-CAM boards got lost on their way to me 😁 BTW, using the VoltLink USB to serial adapter, would not require pushing any buttons for programming the ESP devices as it handles that automatically 👍

I guess that metal 'D' plate was designed for a specific optical sensor unit from the OEM with a shaped and focussed beam which probably made the electronic signal processing a bit easier. we've got portable label printers at work that use IR to count and align the label paper by black marks printed on the back of the paper and they do a really good job under a lot of conditions.

Works like a charm!!Tested it on my water meter and worked on the first try! Makes me wonder what I did wrong...🙃

Engineering at it's best Andreas! ;-) I'm using a setup where I'm not really interested in the actual numbers but trying to detect water flow vs no water flow to react on broken pipes, hanging valves etc. My theory is that in 2 hours there should always be 10 minutes with absolute no water usage - at least in my household. If not I use a central valve to close the entire water flow for the house. Works great so far, f.e. detecting hanging toiletts or similar. Had just one false alarm in 5 years. I also have added a ESP32 CAM .. but basically to just allow me to get a "video" or static picture from the counter (that's what the default firmware of the ESP32 CAM does). My water supplier (10 years ago) offered a special device for ~60 euro that is able to put out a open collector contact for every liter water. Imho there is a spinning magnet inside the water meter itself which they plug on. Easy way to get the "real" numbers of water usage. And btw: The "central valve" just a pure mechanical construct which uses an old cordless drill to manually close the central valve ;-) Cheers!

Fascinating! I hadn't thought of using a camera on an analog meter when I went this route. Like you I already had electrical monitoring into Influxdb (thanks Iotawatt!). So I started to wonder about water. I already had some cheap hall sensors on my outdoor water lines to verify that the irrigation system keeps our garden watered. But what about the house itself? Here in Canada (Vancouver) we don't pay for our water, so it's not metered, so I didn't have an analog meter to start with. Digital water meters were pricey, and usually meant for high-volume, so not very precise. Then about a year ago EKM Metering released a version of their pulse water meter that pulses every 0.283 litres. You can hardly turn a tap on without a pulse. Plug the pulse output into a Sonoff running Tasmota, and tada! A count published to MQTT every minute. The meters are cheap enough that I bought two: I still need to put the second one on my hot water line, but soon I'll know water the hot:cold ratio is for my shower. I have yet to convince my partner why this is useful. :-)

I have same counter in Vaud but the water supplier add the remote sensor by default. The M-Bus can be read on 868.625M from your favorite SDR gear and RTL soft. They do that so they don't have to ask you to check the counter. use Izar protocol and NOKEY param.

I've been working an AI agricultural system to read 24k analog meters. When trained with a custom dataset from our meters we've achieved excellent recognition and conversation at the edge. Using a priori information about the meters prior state we can deal with many other edge case issue since we know meters only increase over time.

@2:07 - That trigger button made me think of one of those vintage accordion plate cameras (on a tripod/stand) which has a cord and "button" to trigger the snapshot. Maybe as a project you should make a retro version of that style camera using an esp32-cam and cord. ;)

Straight to the point, as always! State machine rulez 😁👍

Just in time, been on my list for a few months now. Next : eBus from Vaillant 😉

Just a little tip: If your microcontroller doesn't have an ADC pin or you can't spare another ADC for just a simple sensor, you can read it with a digital pin. We can accomplish that with example a PNP transistor on the output of the sensor and with a relative low base and collector resistor, that will make the output clip between gnd and 5v, like a digital signal.

Really love all that amassed "treasure".

Both solutions have their set of problems on which they are useful. Thank you! Edit: I'm having a lot of fun reading the comments, so many clever ideas.

If the project supports training then based on your own images then it really should be possible to get perfect recognition of the main digits at least. Constructing and training on synthetic images of the digit transitions might well be reliable too. I always prefer the simple pulse counting and interval timing approaches though as they're perfect for smaller uC's such as PICs.

I like the more practical approach to the measurement - using pattern recognition is good idea for some events, but reliability is a big problem. So using something like you did is far better, I think.

Thank you for these excellent videos. I love the clear thinking and explanation you bring to them. So, as AI has transitioned from its early mechanistic approaches (e.g., expert systems, etc.) to the current models based on non-parametric statistics, there is also a subtle change in its successful application. The current models are more about being 'less wrong' than 'exactly right in the instance'. This is encouraged by its alignment with large number theory and other assumptions (e.g., Bayesian Analysis, etc.). Given this, we might best think of modern AI technique as most suitable as part of a 'problem-space reduction' tool that gets bolstered and completed by more traditional, deterministic kinds of algorithms, rather than as a central part of some core algorithm. AI's real mission, IMHO, is about abstracting 'the smart' from from everyday good engineering, and this will require a different kind application with different kinds of value produced. I like to think of AI as a kind 'context manager' for whatever problem(s) I am trying to understand and solve. Still, this was a bold and valuable attempt. Bravo to Andreas Spiess, The Guy with Swiss Accent.

Two interesting approaches. i am using an inductive proximity sensor placed above a rotating metal semidisc to pick up a signal

Yup, if water meters where not leaded 😉 we could glue a tiny magnet on the dial arm. Anyway there are also sold water meters with built-in pulse output, but then you had to buy it self and let register and lead it again at the water supplier.

I tried exactly the same with my heating central at the attic and my house power meter. I ended up kicking the pattern recognition in both cases. The fresh installed smartmeter (power) provides a blinkig LED, 5000 ticks per kWh. Easy enough to read using a lightsensor. For the central heating I just tried to find out how often the flame was ignited. I ended up using a lightsensor as well, using the "flame is on" LED of the heating central. I would love to read gas consumption and water consumption as well, but here my problems are more real life related: no power to drive anything around those installations. Seems, the guy that planned installation in this house did not have any digital stuff in mind at all. Still, I need to stick my head into that tensorflow pattern recognition. And many thanks for not calling pattern recognition AI. I hate that buzzword bingo that is happening nowadays.

Hi, thanks for amazing video, very instructive. I have a question: why don't use a Smith trigger to signal conformation to have a 0 and 1 digital signal? Thanks and see you next

You can use induction sensor for reading the water meter, super easy and accurate!

Interesting video Andreas. Creativity for the win, by the time one has trained the AI for all the variables that exist in the application, a lot of time has been spent and functionality is still not guaranteed. The IR is a known workable solution. But, we would not have known all of this if you had not spent the time doing the research, and that is the value of doing these projects. It remains an investment in knowledge, and you kindly share it with us. I appreciate that, and it makes for a nice Sunday viewing. Kind regards, South Africa

7:49 - the dream!

Still nice to experiment with AI. For the water meter, I have created one a year ago, using a hall sensor. A few hints for those who want to create one too: - as my meter has a metal spinning disc, a hall meter does work in my case - I created an MQTT message from which I can update the reading after a reboot. Also, I keep track if an update was done. - I had to use a condensor and a resistor to avoid false reading. They were rare, but completely eliminated with that small circuit added.

Nice!

Great video. A technology which will become mandatory here in Belgium pretty soon ! Btw, I like your vocabulary... a 'ESP32 treasure chest' ! ;-) This reminds me of some years ago when I had a visitor in my workshop at the office. He called it 'the cave of Ali Baba', which was a nice compliment. In my and his opioning, it was full of treasures !

Some meters actually use magnet to trigger reed switch and the reader head either sends a pulse or it's smart enough to keep the count and report the current number

Thanks again for a very interesting and informative video. So many valuable nuggets of valuable information from a seasoned engineer ! If your wife had so many stashes youd probably call her a hoarder :-), so its ok if she is cleaner than you. So the AI solution you found is quite brittle, it breaks when bent to a slightly different problem. I also thought it would be straightforward to take images of my oil tank level and automatically remind me when to reorder ... 3m later Im still going into the basement to read the guage directly. Yes the ESP32-CAM is an amazing device, but requires work to do real-world tasks different from any open source available. Introductory solutions can be found with only 50 lines of code, more featured ones are several thousand lines of rarely commented code.

Instead of changing magical number for the initial value, can also let ESP subscribe to a mqtt topic so you can adjust without needing to recompile/reupload

omg this is so much work.... impressive though!

Thank you for another interesting video, I had this particular ESP32-CAM "AI on the edge" project on my to-do list for some time now, I keep delaying it because I don't have a 3D printer to make the mounting support. At 11:45 I reckon your Y axis labels in Grafana are wrong, possibly by a factor of 1k or even 10k (depending on your showering preferences). I doubt that your wife is using only 4ml of water per minute for the evening shower! :)

I attempted a similar project in the past, reading a water meter on a rotating dial. It was a spoked wheel with many more theeth on the outside. Reading it optically worked super well, to a theoretical accuracy of a few ml per tooth. Unfortunately the water meters were replaced with ones that have only the numbers visible. I was unable to find a rotating magnetic piece, or a detectable piece of metal inside.

My meter has nothing like that and the digits don't align very well for image recognition, but most of these meters need to transfer the motion of some rotor inside the water to the measuring mechanism outside the area where water is, without allowing any water to escape. Almost all meters do this transfer of motion using magnetism, so that in turn means there is some changing magnetic field at the center of that mechanism when the rotor rotates. Sometimes that magnetic field might be sufficiently strong to measure with a Hall sensor, but with my meter - with metal housing - this didn't seem to work. But ... I then used a magnetic field / compass sensor glued to the meter body and from that get a vector that indicates the direction of the magnetic flux. When the meter turns, that vector also turns (in 3d) and by reading it out several times a second and tracking that movement (flattening out some jitter which will always be there with something like this), my esp8266 can measure the number of rotations the meter makes and convert that to actual cubic m of water gone through ... Within one rotation the accuracy is not perfect, but half rotations (so those inverting the vector) are detected fine.

I enjoyed the final solution. More investigation of the camera system would have been interesting, but can't disagree with simple, low power, and reliable solutions.

Sure could have used one of these a few months ago, my water bill went through the roof. The water department told me to check if my water meter's needle was spinning, it sure was, non-stop.

Sali Andreas, ich feiere echt all deine Videos, genial - mach bitte immer weiter so! Zum ESP32-CAM mit AI on the edge watermeter Ansatz kann ich Dir berichten, dass dies bei mir erstaunlich gut funktioniert. Vielleicht wurde mittlerweile der Algorithmus auch weiterentwickelt. Mein Wasserzähler hat ebenfalls weisse rollierende Buchstaben auf schwarzem Grund, und die werden recht zuverlässig erkannt und interpretiert. Ebenso die analogen Zeigerchen. In Home Assistant sehe ich per MQTT präzise den Wasserverbrauch. Das Gefummel bis Kamerabild und Licht stimmten, war aber schon e chli en Murks.. :)

The "MBA" crack was great !! 8-)

Thanks for this video. You didn't talk about training the neural net for the failed image recognition phase of the project. Did you just use a generically trained NN? The training phase of NN development will almost certainly be a really important tool for future makers. Perhaps you should take a closer look at this... Note that training this type of numerical recognition system where the numbers always appear in the same location in a picture is one of the easiest to train because you can get immediate and unequivocal feedback on whether the NN output is correct or not. This is generally not the case, which makes training in the general case much more difficult and explains why it is such a hot current research area.

Andreas nice video! Why do you need the analog input. You can also use a "flip-flop" or? Do you know there are different esp-cam modules are on the market?

To read the metal piece you need to use LC sensor. The Hal of course will not work. And the IR in the field will going to have problems because of the moisture and the water around the meter. Using LC sensor will eliminate that problems and is very robust. Use two LC sensor to find also the rotation direction of the disk. By the way very nice video!!! I'm waiting to see one using LC sensor, there is plenty of documentation out there from ST, IT and meany more.

"An engineer's work always starts when something does not work, the other things can be done by the MBA's". That hits harder than it should hahaha, preach!

Wow, great project and test of "AI" (CNN) vs reflectance counter. I wonder how effective traditional image processing would work, e.g . bitmap matching and arrow angle analysis. Probably you could have retrained the NNs and tweaked the alignment to get it to work, but that's a lot of work. I have a friend who wants to monitor water heater status lights, and seems like an ESP-Cam would work well. I was thinking of using simple image analysis (AI seems harder). You could have written image analysis to detect the spinning disk, but you could have a low power solution (battery) with your reflectance sensor monitored by the ULP peripheral.

what about placing a piezo sensor to the water pipe? when there is a flow it should create some vibration and with some calibration there may be good results :)

Hello Andreas, regarding the photo transistor, how about a Schmidt trigger to take the sinewave and generate a pulse?

Well, if you're storing a single eeprom variable at a time, you can always store it's complementarity, alongside it, to check for writing errors

6:39 🔥🔥🔥

I believe it would be better to send whole image to RPI and use openCV or similar to distinguish the numbers directly on RPI

Very interesting. I like your solution much better than the AI solution. I am curious why you chose to setup the reflective sensor in an emitter follower configuration. I am sure you had to tweak your LED current to come up with such a nice sine wave. Phototransistors are hard to keep in a linear region.

Polarizing filters may help, but the angle of the reflectet should be in the neighborhood of the so called Brewster angle (the closer the better). If the angle of the reflected light is close to zero degrees (vertical incidence) a polarizing filter doesn't help. I such cases two polarizing filters may help: the 1st filter in front of the LED, the 2nd in front of the camera. The two filters must be 'crossed'.

Using a correlation filter would have been much simpler to implement and probably more reliable, considering that the setup has fixed lighting conditions

I think for having better results from the AI, it would be necessary to train it with pictures from the watermeter which is actually used. In German magazine "Make" there is currently a series of articles on how to do this ("AI for ESP32" or "KI für den ESP32" in German, unfortunately behind a paywall). The AI is trained on a PC with python, only the result is transferred to the ESP32. Didn't do it myself, but I would except much better results.

You compared pears with apples. In the traditional approach you were able to choose the right tools and write a program to make them work. The equivalent process with such a focused AI tool is training it with your images and your labels (e.g. black background, rotating reflector, ...) After that and maybe some lighting adjustment, the camera-powered solution can be far more reliable than the traditional one (e.g. being able to recover from outage/disaster and being able to store some pictures for periodic manual verification)

Andreas, could I look at the node red flow please 🙏. I am on the edge to about power.

Hi Andreas.. as usual solution is generally simple..when sought out. AI is needed where normal I wont work ;). Btw.. i have been playing with this cam since quite a while. One way to reduce glare is 3D print diffuser.. other control intensity, lastly as you did rightly.. external LED. But my major challenge is... how do we get rid of that glue.. never managed to remove or losen that focusing lens... any thoughts?

They could also be using a metal detector. i.e. An inductive proximity sensor. Alternatively, a hall sensor and a biasing magnet. Inductive proxies are good for any metal, foil if non-ferrous, while hall and magnet is only good for ferrous targets.

To get rid of the glare, best is to put the camera to read at 45 degrees to the surface, but anything will see improvement. That will take the LED and bounce the light away from the camera. I thought you had to train tensorflow. Would it have better results if ran much longer to train it?

You can use the ESP32 Preferences API which is much simpler, the EEPROM API is marked as deprecated anyway :)

What about capacitively sensing the grey plate? Looking at the design of the grey plate, I am reminded of a tunable capacitor. Is it made of metal, plastic, or since other material? Capacitively sensing the plate might allow you to get more of an analog/positional reading, instead of a single pulse per revolution. It would also be significantly more reliable if the meter is exposed to sunlight (something that I believe the manufacturer may have anticipated).

pretty much all out water and EL meters here, have an IR LED under the clear cover, so that the 'meter man' can simply place a reader on top to take the reading, eliminating human errors. Well that was until recently, now the wast majority of our meters have been replaced by IoT meters and read online anyway.

Hello Andreas, can this be included into Home Assistant, the raspberry pi project? And is the Peters Scargill solution better for home automation?

This is an excellent video! As usual, you do an excellent job of explaining how to properly engineer solutions. Also, thank you for making another ESP32 video! Arduino have done a splendid job of popularizing microcontrollers for students and hobbyists. But generally-speaking-at least for this channel-I suppose using ESP32 for projects is better because it's more powerful than Arduino. AI is a buzzword (well, buzz-phrase to be more precise) that I've always hated. Not only does it confuse more than it clarifies, worse still by anthropomorphising machines the term AI wanders out of the realm of science and into the world of witchcraft, magic, and astrology. Machines cannot be intelligent, nor can they experience emotions, nor can they die; but people can. When machines perform sophisticated pattern recognition they are not being intelligent nor is there anything artificial about the process. Similarly, the term "cryptocurrency" is also a falsehood. An accurate term would be "cryptocollectible." People collect Bitcoin, Ethereum, Dogecoin etc just like people collect stamps, paintings, antiques, etc. Currencies are issued by governments. As you know, charlatans, shills, and, well, many ordinary devious marketers perniciously employ puffery to manipulate those who are unsophisticated. Thank you for taking the time to lampoon popular malevolent falsehoods that are often mistaken for "The Truth." There is nothing inherently wrong with purchasing collectibles nor with employing sophisticated pattern recognition. However, just as the term "artificial intelligence" sounds more impressive than the phrase "sophisticated pattern recognition" the term "cryptocurrency" sounds more impressive than the term "cryptocollectible". In reality, the terms "artificial intelligence" and "cryptocurrency" are lies which have duped many millions of people. As is the case with most people who purchase stocks and bonds, most people who purchase Bitcoin, Ethereum, Dogecoin, etc are not actually investors but rather merely speculators. Similarly, people who buy into the notion that artificial intelligence actually exists-and can actually exist-aren't real engineers; instead they are people who explicitly or implicitly engage in things like witchcraft, magic, and astrology. These days, many people seek to become wealthy as either investors or as engineers. In reality, for the average person it is very difficult to earn a significant return on investment (ROI) as either an investor or an engineer because most of the profits-the so-called "lion's share" or disproportionate portion-go to a relatively small number of people. Nonetheless, the "lure of easy money" (which is a phrase that comes from Glen Frey's song Smuggler's Blues kzfaq.info/get/bejne/grmIftx5speXY5c.html) or a "gold rush mentality" has led millions of people to "check their brains at the door" (that is, act foolishly).

Hi I lived the solution implemented by you and am trying to buld one At 8.03 min I see you have suggested connecing 2 resistors 330 ohm between vcc and emitter and 10K one between Gnd & Diode What power rating should I use for them?

Hi, did you try a capacitive sensor? That semicircle looks metallic

using EEPROM is good but since you have connection via MQTT ( RPI ) and then to Influx DB, why not in startup get the last values from DB instead. Let the DB be your master source of data But awesome project for sure

Is there any mechanism that we can measure AC voltage level, in my flat the voltage fluctuations are common and the AC stops working and gives 88 error which is during high or low voltages. Please make a video if feasible. Thanks

In the past I heard hair spray may help to avoid reflections.

Hi! Can you make a video for water level monitoring through ai camera..it will help a lot..thanks

There are ultrasonic clamp on sensors available who measures water flow by the doppler effect. Didn’t have such yet, but would be my way to solve. Maybe they can be DIY build with transducers?

Noce way to solve the problem! However I'm not sure why you hadn't just used current analog check since it's metal plat, you can print a simple adapter and it would work like a charm. Moreover you can else just glue a tiny magnet and use magnetic field sensor using i2c ( not hall sensor ) which should be a bit more reliable. There's a video somewhere who explain Tesla door, they do something similar but with door. :)

I can only connect my esp cam to the programmer with the jumper on 5v. 3.3v causes connection issues.

90% of what is sold as "AI/ML" is indeed a pattern recognition algorithms. Either detecting the pattern or using detected patterns to detect further anomalies. From computer system monitoring to driver assistance fall in that category.

hello in order to improve IA result, in your cas you could try to nomalize input image in a different way to help neural network in his recognition work, binary color and gaussian could be a start. you can try to treshold you pixel value range, over a long time (find min max, of a pixel zone) Add physically some light reflector to help with changing luminosity. may i suggest you to try tesseract in python, quite effective

What kind of Hall sensor did you use? Linear or digital one? That plate is plastic or some metal?

At around 4:30, the glare problem. Can I suggest three things? 1) introduce a polarizing filter between the camera lens and the meter (this may well solve the problem), 2) Degrade the video signal to monochrome or use a monochrome camera, don't need the colour and 3) Ignore the decimal digit, it doesn't add much information but creates a lot of problems, after all the water company doesn't include this in their billing do they and it wouldn't be evident in your resultant usage graph either?.

In situations where water meter locations usually lack access to power, is there a viable option to adapt this project by implementing RF for low-power signal transmission using batteries? The goal would be to receive these signals on a gateway and subsequently publish the data to MQTT. Thank you for the great videos and the moments of humour.

I wonder if a circular polarizing lens over the camera lens would knock out the glare?

How did you do the aggregating in Grafana at position 0:42? I tried "field(value) last() difference()" and "group by time(1h)". But a) this does not work for months and b) fill(previous) doesn't work, so I always have spikes like in fill(linear)

Instead of using a magic number which overcomplicates things, store the crc of the number or some other calculated value based on the reading to be saved and check for that instead.