"With AI we can find the needle in the haystack" 

DR. GRÖNZIN, what does your role as CTO at Leuze entail?

I am responsible for all technical issues, which is commonly associated with development. But it’s more than that: Traditional product management is also under my purview. I am also responsible for getting the developed devices into series production. And lastly, technical service for the devices after purchase by the customer.

How come Leuze as a component manufacturer is working with AI?

For us, it has always been a matter of course to create change ourselves. That is why we are always looking into new technologies and their relevance for our main industries. For example, we were the first sensor manufacturer to develop sensors capable of communicating with the cloud directly via OPC UA. We believe that artificial intelligence will be relevant for our sensors and their use in our customers’ applications. That is why we are looking into it and why we are communicating with our customers on the issue.

What specifically is Leuze working on at this time?

Right now, we are in a trial phase with an automobile manufacturer. We are using AI in connection with our barcode readers. A barcode reader captures labels and delivers relevant IDs to a higher-level system. This sensor self-monitors. That means the sensor knows its own status and is able to communicate this to the system. This has been so for many decades. The only problem is: The sensor’s view of the status is very limited, meaning only from its own perspective. For example, the sensor can signal "I’m currently reading," "Excellent reading," or "Very poor reading." But a single device is incapable of determining the reason for this. Our AI project aims to solve this problem: Is the poor reading quality due to the device, the barcode label, or are there interfering factors in the environment?

How does AI come into play?

First, it is important to know that we have countless components in any given system. In logistics, we often have up to 1,000 barcode readers at different installation locations, and thousands of labels on pallets or boxes that pass by this barcode reader. AI allows us to handle and analyze the data volume. Picture this: In the course of a process, a label passes many barcode readers. It is also read at different installation locations. Overall, this offers an equation with many unknowns: the countless barcode readers, even higher quantities of labels, and the different installation locations. At every station and for every label I get different results. Take the reading quality, for example: sometimes the quality is 90%, the next time 80%, etc. And at some threshold value the label is no longer read. Now I can use AI to resolve this complex equation by retracing and finding the cause. Does the poor reading rate always occur on one barcode reader, only one label, with a certain type of label, or at a certain installation location?

What learning method is used for this? A neural network?

Correct. We use so-called recommendation algorithms. This is commonly used by streaming services, which evaluate user behavior and use this information to recommend films or series. To use this analogy, imagine the barcodes being the films and the barcode readers being the users of the streaming service. So then a barcode is more or less "attractive" for different barcode readers.

What are the benefits for the customers?

There are two types of situations to consider: commissioning and ongoing operations. During commissioning, the system builder is under maximum stress—everything must get done quickly. Picture a large warehouse where materials have to travel over large distances of up to 40 kilometers. If only one of about 1,000 barcode readers installed is poorly aligned and you don’t know which one it is, you have to walk the entire route. You are looking for a needle in a haystack, and the clock is ticking. But if I can get straight to the faulty barcode reader, then this adds enormous value. In addition: Sometimes it’s not just a single barcode reader, it’s several. Other difficult borderline cases: the barcode reader is somewhat aligned and reads most of the time, but labels are still frequently lost, such as because the reader is slightly inclined, reads in the border area, something in the system is vibrating, a viewing window is fogged up, or a label is too damaged. AI enables us to filter and find the causes quickly.

...and what are the benefits for system operators during operation?
 
For a system operator, there is nothing worse than an unexpected shutdown. It costs time and money. A planned shutdown, on the other hand, is much more palatable: For example, the operating company can produce in advance, outsource in advance, or keep the delivery performance for customers high and make up for it with subsequent production. Our approach also enables predictive maintenance. Sometimes we use data from several years of operation, on the basis of which early detection works optimally—and the system keeps learning constantly.

There is only one question left: When will the solution be available?

That will still take some time. There is always a bit of a lag for new underlying technology to become established. Take OPC UA, for example. We introduced the first sensor with OPC UA in 2016, and the first larger installations based on this technology are only now starting to operate. With new technologies, there are also always other issues that have an indirect effect on the application. For example, right now we are looking into IT security issues, in terms of what can be sent where. In particular, this is an issue that concerns larger systems in industrial environments, when data has to travel through the cloud. Alternatively, it is also often possible to use edge devices, which keeps the data on site with the customer.

www.leuze.com/en/GB/