We've just launched our map. Add yourself by clicking here!
Our new machine to sort plastics
- This topic has 44 replies, 19 voices, and was last updated 2 years ago by
.
Hi guys,
We are a group of students and we have created a low-cost machine for immediately identifying plastics using infrared rays – see our video here https://www.youtube.com/watch?v=aEFnh86olhg
And we would like to ask for your help! We are interested in how you currently sort plastics and if you have any good hacks for it. Additionally, if you would like to be one of the first people to test our machine, please do let us know!
The best way to get in touch with us is to go to https://matoha.com/pp/.
Be sure to check out our Facebook and Twitter page as well!
Attachments:
Hey thats cool, what did it cost to build?
Do it work with all the recyclable plastic?
Or even with other plastics?
Thank you! The current ‘brain’ inside the machine is a Raspberry Pi – an Arduino won’t be powerful enough to do the analysis of the infrared spectra. We are considering making the device open-source hardware, though it’s not an easy decision – both closed and open-source have their advantages and disadvantages. In any case, it will definitely be closed-source (secret!) until it’s ready for the real world and perfectly functional.
ABS and ABS-PC – I must say we haven’t come across an ABS-PC sample, so I don’t really know. I assume it depends on the composition of the sample and also other factors. You could send us physical samples and we could simply measure them 😉
Thanks for the interest – building it was more than a year of prototyping – starting with big setups on optical benches and gradually making it smaller and functional. And prototyping optics is neither cheap nor easy. If you are asking about the device itself in the current revision, our goal is to make it at least 10-30x times cheaper than the currently existing machines (priced at €20-30k), and most likely even less. We are still quite in the prototyping process, so hard to say.
The machine is able to identify mainly the 5 most common plastics PE/PET/PVC/PS/PP, though it is quite likely other plastics will be possible as well.
@matoha
it is still wornderfull what you are doing
a rasberry-pi is an awsome micro controller so keep up that work
i recycle old computer monitors and i have a batch of abs-pc for you if you are intrested i can send you some (free of charge) just to help you guys
if you are intreseted send me a pm i will send you my contact info
Cheers! Our instrument does best with coloured and clear plastic items. The darker the plastic is, the more our machine struggles; however, industrial automated sorting lines can’t accurately categorise dark coloured nor black plastics either 🙂
Is it of particular importance to you?
Hans
Matoha Ultrascience
@matoha
Your project is awesome. It is already looks cool for prototype.
Sometimes it is really difficult to recognize type of plastic.
There is an example: same manufacturer (Komus), same type of container (food), same color (transparent). But on the left PET, on the right PS.
Attachments:
This is so cool, I’m exited to heard back from you with more progress 🙂
I was doing some reading and found an interesting article from Sandia National Laboratories back from 1993 but it seems they used a high-resolution fourier-transform infrared spectroscopy device plus a well trained neural network. If you are using the same approach I highly doubt this will end up being cheap or open source.
It´s an amazing device, thanks for share it. How can i get it?
i really want to know more!!
Hi @xxxolivierxxx,
Unfortunately, I can’t get hold of a PDF of the article you have linked. But I also have good news – the technology has made a huge progress from 1993 when it was written. There is a wonderful Python library called SciKit Learn, which effectively means the neural net/ machine learning algorithms can run on a £25 Raspberry Pi (as a matter of fact, that’s what we are using).
Also, there are small machines that can do the identifications, priced at something like £30k – that’s obviously horrible and inaccessible. We are four students with very limited budgets, so rather than putting in super-expensive components, we are using our brains to find nice ways how to do it without them. I can’t promise any concrete numbers (still in development), but what I can promise is that it will be much much cheaper than the existing technology 🙂
Martin / Matoha Ultrascience
@chelbig Thank you! We are are working hard to get it fully functional, though it will take some months before it’s ready for beta-testers. We’ll keep you updated – if you are interested you could go to matoha.com/pp to tell us more about what you are doing – e.g. which plastics you normally sort&use and how.
This is impressive. I’m curious what you’re using to detect the plastics’ spectrum. Sure beats £30k as far as accessibility is concerned!
Thanks @copypastestd , those are exactly the plastics we want to be able to sort: ones that look essentially identical to the eye! When we first started prototyping we encountered a lot of waste like that; indistinguishable, and often missing their numbering or were just wrong. The problem will be solved with our machine!
Hans
Matoha Ultrascience
you guys have solved a major issue so far the floating method is what I have been using. I was wondering how long do you need at the sensor for it to read which plastic it is? I have also signed up to be a tester cant wait till its finished. If i can help in anyway just pm me.
Keep up the excellent work!
I wonder if clear plastics don’t pose a major issue. With light there’s less reflection and different refractory properties which would vary with the shape and thickness of the analysed plastics.
Is this a problem with IR sensors ?
Or do you instead use both clear and opaque plastic samples for the training process to capture different spectrum of the same plastic type/color ?
@plastikfantastik Thank you! At the moment, the acquisition of a single spectrum is ~ 100 ms, though because of certain bugs on our PCB (which we’ll fix in the next revision) we have to employ heavy averaging (~10x) to get a decent signal-to-noise ratio. The machine learning classification afterwards is fairly quick, < 10 ms.
The signal intensities we are working with are like 1 nA spectrum peak size, which is obviously very tricky to get working – that means we need ~ 10 pA (10^-11 A) resolution with similarly low noise.
@andyn Our near-infrared source is a special alloy of tungsten, contained in a quartz envelope. Additionally, it is filled with a mixture of iodine and bromine vapours for optimal performance. Also known as a £1 halogen bulb 😃 (usually the simplest solutions are the best 😉).
The reflected light then goes through a monochromator which scans across the wavelengths of interest and the monochromated signal is measured by a photodiode.
For now, we are quite lazy to release the blueprints since it’s not ready yet with many bugs in there. E.g. even if someone built it, it wouldn’t be very useful. (Dave Hakkens does the same thing, releasing stuff only when it’s ready.)
@armbouhali You are absolutely right 😉 With transparent plastics, you get much better spectra using transmittance rather than reflectance. That’s why at the moment we are working on adding a second lamp which will shine through the sample so that both reflectance and transmittance are supported.
In our experience, the reflected and transmitted spectra were not very different (the physical origins of the electromagnetic absorptions are the same). For some samples we observed minor differences, this was for example due to a thin coating layer on the sample.
Near-IR is reasonably immune to changes in colours, though black pigments usually make the samples very non-reflective and pose problems, as we said before.
Martin / Matoha Ultrascience
@basman We have received your samples, thank you! Have a look here what we found out about them using our machine 😉 https://www.youtube.com/watch?v=2DortfmrGgM
@matoha (btw thank you so much for all the work you are doing on the plastic analyzer machine !)
I’m sorry but you got me lost in your last linked youtube video. Your “big machine” as you say so shows there is no presnece of polycarbonate.
I know you are not an all-seeing-&-knowing god, but what happened there?
Was the monitor piece marked (by constructor) ABS PC? Or was just a supposition from @basman ?
I’m sorry to bother asking this but if the error comes from constructor; it might be verry missleading for DIY activists (until you release your baby open source hehe :p)
Anyways, thanks again for all your commitment to this project !
Peace,
Nick
@imuh thank you! So one of the samples was marked ABS by the manufacturer, the other one ABS-PC by basman. So yes, his marking might be inaccurate. Or maybe it was a very clever way to see if we can actually distinguish things – if we said they are from different materials even though they are same that would be quite embarrassing 😀
No worries, we are making plans how to release it open-source 😉 At the moment it’s somewhat a mess, so I don’t think anyone has spare couple of hundreds euros lying around to build something that doesn’t fully work (yet!).
@basman – how did you know which one is ABS-PC?
i did not know it the markings are factory made
some just lied to sell abs as abs-pc i guess
You must be logged in to reply to this topic.
