Exhibitors & Products
Events & Speakers

They established their company, known as Erium, in 2017, secured business grants and landed their first customer projects, which they often had to work on in their spare time as both had a job surveying stars at the Max Planck Institute. "A scientific career was too unpredictable for me," says Steininger. But surely self-employment is even more precarious? "We had so much interest from industry. We weren’t worried about the future."

Together with eight colleagues, they now occupy premises right next to Munich’s Technical University. So what do astrophysicists have in common with industrial processes?

Satellite data, for example, are very costly, and the measurements involve a great deal of work. For this reason scientists try to construct powerful theoretical models of reality in order to add to their knowledge with relatively little data to go on. "It’s just like measuring the light from a star." The researchers construct an image of how things should be, and then fill in the gaps in their knowledge with statistics. "And that’s how our software works in the car industry, too."

In collaboration with the client, Erium works out how a process functions, putting together the unknowns with the available expert and domain knowledge: small data rather than big data. "The problem is that satellites are super-calibrated. In industry, the data situation is more uneven. We find unstructured databases, as well as bad or corrupt data, caused by such things as a defective sensor cable." These sources of error need to be identified. So data preparation in industry is always a problematic issue for the team. "If you set out to include all the data from the outset, you’ll never get beyond the preparation work." In short, it’s better to start with fewer data.

A fascinating use case for their software can be found in the car industry. "We help the assembly line workers with the correct positioning of attachments such as doors, tailgates and boot lids." This is a real problem for them as the hinges are bolted on the bodyshell in its bare metal state – before it is painted, glazed and fitted out on the inside. Why? Because the part that attaches to the other leaf of the hinge has to be painted separately first. "The process sequence means that the mounting location of the hinge has to be determined at a time when the bodyshell is basically just bare sheet metal. The car then goes through the paint shop, and gets fitted out with all its special equipment – all of which obviously adds weight and alters the geometry of the bodyshell. But it is important for the appearance of the car – and also for effective noise suppression – that the fit of the doors in their frames, for example, is accurate to within a tenth of a millimetre.

So the question is: How can you hang a door so that its final position, allowing for any deformation and slump, is accurate to within 0.1 millimetre?" says Steininger, posing the same question that his clients ask. At present, the problem is solved by manual adjustment. "But obviously our clients want to have as little extra work as possible. So there is a permanent iteration loop between the finishing stage and body assembly, so that the setpoints for the assembly robots are continually revised. The problem is the delay involved." So what’s the solution? "We send setpoint instructions to the robot based on inline measurements. This means that if the inline measurements after painting show that the process has altered from the previous vehicle, we can react instantly and give the robot a new setpoint. And that reduces the stress for the assembly line workers, who otherwise would always be under pressure to react to change as fast as possible."

But there is no black box behind all this, programmed for machine learning. "That wouldn’t work, because we would need many thousands of data points for that. Instead, working with the engineers, we have constructed a statistical model that contains as much expert knowledge as possible. Specifically, we exploit the fact that the measurement points on the vehicle are spatially corre-lated." Which means in practice: "If the door, for example, is twisted out of line, that means that all the measurement points will show a degree of rotation. Correlations of this sort tell us a great deal about the structure hidden within these data. And by means of this model, we are able, with relatively few data points, to make predictions and recommendations for action. In other words, we can give the assembly robots the correct setpoints."

And it’s not just the car industry that is interested in rapid small data analysis. In various use cases, this technology is helping to increase efficiency in the machine tool and additive manufacturing industry, as well as in the e-commerce sector.