Page 54 from: Our summer issue is out now!

54

loads,’ Wezstein notes. ‘The develop-
ment measurement system will help us
benchmark our future improvements
when it comes to actuator control,
implementation of more complex
image processing and including convo-
lutional neuronal networks.’

NOVEL PCB APPROACH
The recycling of printed circuit boards
(PCBs) could benefit from a deep-
learning assisted X-ray imaging
approach, reports Markus Firsching of
the Fraunhofer Institute. ‘At the
moment, no system is available that
can estimate the monetary value of
used circuit boards. Our team is work-
ing on an intelligent and robust auto-
mated method that sort these based
on features in the DE-XRT images.’
The solution is said to work even in
dirty environments making it a promis-
ing tool for recyclers. ‘We first created
a sample of 104 individual circuit
boards,’ Firsching says. ‘Our approach
relies on the detection of components
that contain valuable materials, like
integrated circuits. Think of ball grids,
pin grid arrays, tantalum capacitators
and connectors.’
Fraunhofer’s research achieved accura-

cy scores of 87.8%, 82.5%, 79.2% and
88.8% in four different classes, namely
integrated circuits, tantalum capacita-
tors, connectors and ball/pin grids.
Their monetary values were very close
to the expected values.
AI predicts particle size
Naturally, today’s sorting industry has
been heavily inspired by state-of-the-
art artificial intelligence-based systems.
Lisa Kandlbauer of Leoben University
in Austria took a deep dive into this
topic to see how the technology can
help with particle size prediction in
solid waste. She used a bi-layered neu-
ral network to set apart five particle
sizes: 10-20mm, 20-40mm, 40-60mm,
60-80mm and fractions larger than
80mm.
‘In the first experiment we achieved a
classification accuracy of 75% and then
we reached 82% for virtually all class-
es,’ she reports. Using another meth-
od, the RUSboosted Trees model, the
classification was about 65% accurate.
The best score was 97% achieved by
the neural network method.
‘Especially for size classes of 20-60mm
of paper and cardboard, the new
AI-powered models can be seen as a
suitable prediction method,’ she

declares, noting that, on average, the
level of falsely classified particles was
far below 20%.
‘To understand and explain abnormali-
ties in the results, relations between
the individual descriptors and images
need to be investigated,’ Kandlbauer
states. ‘Nevertheless, the presented AI
method shows a promising way to pre-
dict and measure particle size in real-
time when treating commercial waste.’

TRANSPARENCY NEEDED
Researcher Steffen Rüger of the
Fraunhofer Institute believes AI can
also be used to enhance aluminium
recovery. He advocates combining dual
X-ray transmission technology and
deep learning to sort a database of
almost 75 000 metal fractions into pure
aluminium, which yielded output frac-
tions with a 91.7% purity and similarly
high recovery levels.
In Rüger’s view, the final deployment
of innovative deep learning systems
depends mostly on whether or not
they are accepted by society and
industry. ‘That’s an easier said than
done scenario,’ the researcher tells the
audience. ‘With new technology, peo-
ple don’t understand how it works – if
they can believe the hype. That’s why
transparency is key.
‘We employ local explainable AI solu-
tions to hopefully open these black
box models. To provide insights and
foster trust amongst the people in the
value chain. That’s step one.’
We have plenty to review in two years
when the next gathering in Aachen
takes place. .

Labels can make sorting tasks much harder.

Louise Hagesjö is working on a laser-
induced breakdown spectroscopy R&D
project, with promising results for black
plastics sorting.

48-49-50-52-53-54_sensorbasedsorting-3.indd 54 02-05-2022 14:42