The importance of Machine Learning for RFID tunnels
RFID is playing more and more an important role in the speed and efficiency of logistics centres. Using RFID tunnels, it is possible to automatically receive 600 to 1200 boxes per hour, and know exactly the contents of those boxes — up to the last item, down to size and color. This allows brands, wholesalers and other logistic parties exactly what they received — compared to just assuming that the supplier was right, or doing intermittent sample checks.
If you are going to employ an RFID tunnel, it should be highly accurate — otherwise you’ll have a drop in efficiency (manually having to verify a lot of boxes after the tunnel), or wrongfully blame suppliers for not shipping the right quantities. This means in practise that the RFID reader should read everything inside the box, but nothing of the box before and after — and any boxes that might linger around the tunnel. Especially at high speed, low seperation and larger box quantities, this becomes a significant challenge.
Reading everything you need to read…
The easy part is reading everything you want to read. You can use as much RFID power as you can, but should also consider efficient configuration of the way the RFID reader talks with the RFID tags. This should be optimised for speed, but there is danger in suffering interference from other readers. Mitigation techniques should be applied.
…except what you don’t want to read
More difficult is to not read anything of the previous and next boxes in line. A first step is to select the best RFID antennas for the job: ones that read very carefully close by, but not too much farther away. Also the placement of the antennas is of critical importance. However, this will not solve all stray tag reads.
Some solution providers use mechanical metallic doors or roller shutters in front and after the tunnel to shield the box you want to read from boxes before and afterwards. However, this solution is most likely to slow down the throughput, is more expensive, and places an additional burden on maintenance due to the mechanical components involved.
The Nedap solution is to use Machine Learning. In the past few years, Machine Learning has come a long way — most recent examples include the ChatGPT chatbot, that offers an almost human-like conversation.
Machine learning in RFID tunnels: TunnelML
With TunnelML, our model creates an estimate for every RFID tag read by the reader on how likely it is that it was included in the box that we want to read, or whether it was a ‘stray’ read (belonging to a box before or after the one you want to read — or a box just lying around the tunnel).
The accuracy of the algorithm needs to be extremely high: if you have 50 items in a box, and you want to estimate the contents of that box correctly, the algorithm needs to be right 50 times. So, having a 99.9% correct algorithm, will still yield one in thousands estimates to be wrong, and thus a one in twenty-boxes false rejects. The algorithm thus needs to be very accurate.
Therefore we have carefully selected more than thirty parameters that our algorithm takes into account when deciding on an RFID tag. The algorithm was then trained by reading thousands of known boxes of different vendors and brands, in different reading set-ups. This has made our algorithm robust enough to run at high accuracy at high speed — in some circumstances only with a simple curtain at the entrance of the tunnel, or no shielding at all.
Nedap Harmony
With Nedap’s TunnelML algorithm, you can design and use RFID tunnels that require less hardware. Our algorithm can accurately read and identify the contents of boxes without the need for mechanical metallic doors or roller shutters that can slow down the process and increase maintenance costs. As a result, your initial investment will be lower, and your TCO will be reduced because there is less hardware to maintain and that can fail. All up to 1200 boxes per hour.
TunnelML is part of Nedap Harmony, a RFID platform-as-a-service build for logistic applications — available for system integrators and end-users. We have a tunnel-model that can be licensed, or you can bring your own hardware. Of course, TunnelML can also be trained on your products and your situation to ensure highest accuracy possible.
