Central regions | Technology & innovation

Artificial intelligence sorting garbage

27 Oct '21
By 2030, all waste in Russia will be sorted, and half of it will be recycled, follows from the presidential decree. How robots will help this, said Alexander Nevolin, founder of Nevlabs, which develops artificial intelligence-based waste sorting equipment.

To sort waste, we use artificial intelligence, namely convolutional neural networks. Man spied their idea from nature.

The fact is that the vision of humans and animals is also a multi-layered neural network. The first layer recognizes the most basic objects, such as lines and points. Further, the neural network, using information from the previous layer, recognizes more complex objects. For example, among the many lines, she can distinguish some geometric shapes. This happens up to the final layer, which gives the answer - what kind of object is in front of us.

Neural networks, like our brains, learn from examples. We show them a lot of photographs marked, for example, “This is a clear PET bottle” or “This is a green PET bottle with a shrink sleeve.” In the images shown, the neural network finds patterns. This allows her to recognize objects that she has not seen before.

How AI is getting smarter

To train a neural network to recognize garbage, it is necessary to collect a large database of photographs. For this, cameras are installed in factories - as a rule, directly above the conveyors of existing sorting lines. It is advisable to remove garbage in the conditions in which the installation with a neural network will work. This will take into account the flux density, type of lighting and other factors that affect the accuracy of waste recognition.

The next step is to manually mark up the photos. For each waste class, 5 to 10 thousand images are needed. There are about 10 classes in total - in total, you need to mark up 100 thousand pictures.

To do this, we employ home-based workers. They register on a special site and take an introductory course on the types of waste, after which they need to take a test. If we are satisfied with the results, the person is allowed to work. The system also provides for selective control of marking: users check each other's work in order to exclude the possibility of errors.

To do this, we employ home-based workers. They register on a special site and take an introductory course on the types of waste, after which they need to take a test. If we are satisfied with the results, the person is allowed to work. The system also provides for selective control of marking: users check each other's work in order to exclude the possibility of errors.

As a rule, employees mark 100 thousand images per week. Next, we train the neural network by repeatedly showing it photographs. With a high-performance graphics card, training takes two to three weeks. This approach is also used for further training - product manufacturers often change packaging, so it is advisable to update neural networks at least once every six months.

For this, photographs of the waste passing through them are collected from all sorting plants via the built-in Internet modem. First, it allows you to perform manual control of the recognition quality. Second, to support the neural network's knowledge of waste types, such photographs can be selectively marked up and added to training materials. Since data from sorting plants are collected on the server centrally, the neural network is equally successful in recognizing garbage from both Moscow and Vladivostok.

Sorting technologies: robots and compressed air

The heart, or rather the brain of the sorting plant, is an electrical cabinet with a computer on which the main software and a vision subsystem are installed. Receiving data from the camera and a set of sensors, the neural network recognizes the waste and transmits the data to the main software. He, in turn, gives the command for the physical selection of the factions of interest to us.

We produce plants with AI in two versions - the Gurman robot and the Estet pneumatic sorter. “Gourmet” moves the desired fraction from the conveyor to the tank.

Electronics “Estet” gives a command to open one or another air valve. The sorted fraction is shot into a separate bunker, the rest of the garbage falls on the next sorting belt, where another type of waste is selected.

What are the benefits of artificial intelligence

Compared to old technologies for waste recognition, primarily spectral analysis, AI has a number of important advantages:

The cost of equipment and its payback period are reduced. The neural network does not require the use of a hyperspectral camera, the cost of which is more than half of the sorting machine.

The possibility of sorting waste that previously had to be selected manually. The hyperspectral camera sees only the type of material and cannot distinguish, for example, a clear PET soda bottle from a sunflower oil bottle. In reality, however, they need to be processed separately.

Unpretentiousness. Spectral analysis requires regular complex calibration, without which the quality of sorting will inevitably decrease. The neural network is absolutely picky about working conditions. We had a case when half of the luminaires stopped working on one sorting plant, but the recognition quality remained at the same level. Non-working lamps were noticed a month later.

Possibility of training for individual tasks. In particular, a neural network can be trained to recognize construction, medical and other types of highly specialized waste.

Possibility of training for individual tasks. In particular, a neural network can be trained to recognize construction, medical and other types of highly specialized waste.

Low power consumption. Existing sorting plants use powerful halogen lights, and the hyperspectral camera requires an expensive cooling system. Neural networks perfectly recognize debris even with conventional LED lamps that do not require cooling.

Will AI replace manual trash sorting

Manual sorting is a dangerous and dirty job with all the ensuing consequences: high costs, constant staff turnover, unstable quality.

Today there are about 200 waste sorting plants in Russia, where manual labor is mainly used. They plan to increase their number to 800 by 2030, with the widespread introduction of automatic sorters.

We started four years ago, almost simultaneously with companies from the USA and Finland. Several competitors have appeared in Russia together with us. In the development of technologies, we invest the funds earned from the development of outsourced software - the original activity of the company. Also for two years we were in the Bortnik Foundation, and now Nevlabs is a resident of Skolkovo. In the near future, we plan to bring outsourcing to nothing, so that we can only deal with our own developments.

The introduction of AI into municipal waste sorting reduces the cost of equipment - it becomes available to a wider range of users. While our installations are at several waste sorting plants in the Moscow region and Tver, but, I am sure, scaling to other regions is a matter of time.

According to our calculations, with the widespread introduction of automated waste sorting, the use of manual labor will be reduced by 90%. People are needed only at the final stage - quality control. However, the waste here is already much less dirty, and it is much easier to sort it.
Publish in Twitter
Write to Facebook
Google Buzz
Write to LiveJournal
Show in MM
Share MK
COMMENT ON THIS STORY
Find Related Content

Locations:

Tags: AI (4195) /

Latest News: Central regions
2 Dec '21 | Technology & innovation
1 Dec '21 | Technology & innovation
26 Nov '21 | Technology & innovation
24 Nov '21 | Technology & innovation
 

Feature stories

24 Nov '21
Cerebral vascular disorders are responsible for a...
8 Oct '21
Chemists at the St. Petersburg State University...
2 Aug '21
From enhancing the properties of existing materials...
Search (News archive - 22842)
Technology Scouting Solutions For Industry
NATT
Advertising
Marchmont News

Latest News

2 Dec '21
Russian scientists have developed new polymer-based...
1 Dec '21
A Russian electronics maker has launched a new...
25 Nov '21
Scientists in the Russian region of Tatarstan have...

Most read stories from last week

1 Dec '21
The Institute of Human Stem Cells will conduct...
2 Dec '21
Russian scientists have developed new polymer-based...
1 Dec '21
Smart Engines technologies recognize bank card data...