Background discussion on the current research focus of our division »Mathematics for Vehicle Engineering« and its involvement in the European Consortium for Mathematics in Industry
Machine Learning and Big Data for Cable Simulation
At the beginning of December, the web seminar »Math for Industry 4.0 – Models, Methods and Big Data« of the European Consortium for Mathematics in Industry – ECMI for short – took place. Our department »Mathematics for the Digital Factory» is intensively involved in ECMI, was co-organizer of the web seminar and participated in the event with a presentation. This was an opportunity to ask the department what its activities at ECMI are and where its current research focus lies.
Swenja Broschart (Team Communication):
Joachim, you head the department »Mathematics for the Digital Factory« in the divison »Mathematics for Vehicle Engineering« and are one of the three German representatives on the ECMI Council. What does the involvement in ECMI look like?
The connections of the Fraunhofer ITWM with ECMI go back a very long time. After all, our institute founder and former director Helmut Neunzert was co-founder and even ECMI president in 1988. In the beginning, almost the entire institute was active at ECMI – although we were, of course, considerably fewer employees at that time than we are today.
Our department has been increasingly involved with ECMI again since 2013/2014 – we organize and contribute to seminars and mini-symposia. At that time, the Special Interest Group (SIG) »Mathematics for the Digital Factory« was founded, which corresponds exactly to the department's focus of work – after all, that is also the name of the department »Mathematics for the Digital Factory«.
The recent ECMI seminar was then a major activity of the Special Interest Group, where we were represented with the topics of cable simulation and digital human modeling – i.e., the two main research areas of our department in the field of the digital factory.
Kaiserslautern is even one of the ECMI Nodes. What does that mean?
In addition to our division »Mathematics for Vehicle Engineering«, our colleagues in the department »Transport Processes« have been active at ECMI for a very long time. The AG Technomathe of the TU Kaiserslautern is also still strongly involved in the consortium. Thus, Kaiserslautern is an important location for ECMI. Our Gothenburg subsidiary institute Fraunhofer Chalmers Centrum, which develops the IPS Cable Simulation software together with us, has also been involved in ECMI since 2000.
The software package IPS Cable Simulation, which you developed, enables real-time simulation of cables and hoses. Fabio, you are involved in the development. What is your current research status and what are you currently working on?
The simulation tool IPS Cable Simulation is already widely used, especially in automotive development. An important pillar in simulations are realistic model parameters. In our case, these are the mechanical properties of the cables. In order to be able to meaningfully simulate the behavior of cables and cable bundles in cars already in the purely virtual product development phase, we need mechanical properties that are as accurate as possible. The surest way to obtain realistic model parameters is to measure concrete performance samples.
The number of cables in the vehicle is still manageable, but the number of possible cable bundles – i.e. bundled cables – quickly becomes confusing. Measuring them all would be very time-consuming or even impossible, because the necessary samples of cable bundles are often not yet available in early, virtual development phases. In this case, the only option is to estimate the mechanical properties. In principle, there are two ways to do this: One can work model-based or data-based.
Fabio and Lilli, can you briefly explain how different the two options are and what your approach is?
Model-based means that detailed models of the cable bundles are created, these are usually parameterized on the basis of physics and the unknown variables are simulated – in our case, the effective mechanical properties of the bundle. However, basing the estimates purely on models is not practical for bundles because the variety of these is too large. There would be a huge amount of different models. Nevertheless, we know quite a bit about the bundles, which we can also exploit for a data-based approach.
In a data-based approach, you work with existing (measurement) data, which are often – as in our project – considered as input and output pairs. Machine learning essentially means using the computer to derive patterns or models from the given data sets and then using them to make predictions. The transfer models are trained using the existing data, the patterns are learned and can then be applied to new input data
A crucial point is of course the pre-selection of suitable models as well as the pre-processing of the characteristic input parameters, here especially system understanding and knowledge about the expected transmission behavior play a decisive role.
You co-organized the recent ECMI web seminar »Math for Industry 4.0 – Models, Methods and Big Data« and also gave a presentation on this very topic. Vanessa and Lilli, you gave the lecture. What was the presentation about?
As Lilli just explained, we need a lot of training data for machine learning. We are in the fortunate situation that we have our self-developed cable measuring machine MeSOMICS to determine training data. So we can generate ourselves the Big Data we need to train the estimation algorithm.
The first part of our talk focused on our experimental activities to measure the cable stiffness parameters using the MeSOMICS machine. We have now built up a large database and measured a great many cable samples and cable bundle samples that we have fabricated ourselves. So our Big Data is this data on cables and cable bundles. Now an estimation algorithm is to be developed or an estimation is to be carried out on the basis of these data sets in order to estimate the cable bundle stiffnesses from the cable bundle parameters, which are required for the simulation in IPS Cable Simulation.
To estimate the bundle stiffnesses, we use Gaussian processes for regression, an established method from the field of machine learning. Therefore, in the second part of the talk we first presented the main properties of Gauss processes as well as their application in our specific project.
Finally, we presented our results. We were able to show that with Gaussian processes we can derive cable bundle stiffnesses based on cable bundle parameters and parameters of the individual cables combined in the bundle with sufficient accuracy.
How far along are you in this project?
As we were able to show in our presentation, it is possible to estimate cable bundle stiffnesses with Gaussian processes and our good database, which represents a great added value for many applications in industry: Existing data sets from performed measurement series can be sustainably utilized and new tests and measurements can be saved.
For many applications, the torsional stiffness of the cable bundles is also required. The measurements in MeSOMICS for this have been largely completed, and we are now in the process of training Gaussian process models for the estimation of torsional stiffnesses in order to evaluate the initial estimates.
Another focus of current work is in the methodological area: we are investigating approaches to optimize the creation and enrichment of the training database.
Good luck with this project! And are there already further activities in the pipeline at ECMI?
This year, ECMI 2021 will take place in mid-April as a virtual conference in Wuppertal, Germany, as a replacement for last year's conference planned in Limerick – which is the ECMI asset in Ireland – that unfortunately had to be cancelled due to the pandemic. I am active there again as co-organizer of the mini-symposium of the SIG »Mathematics for the Digital Factory«, in which we will participate with four contributions. There will be one more contribution from MF, to the minisymposium »Data-driven Optimization«, organized by Simone Göttlich and Claudia Totzeck – both are alumnae of the AG Technomathe. I guess there will be some more from other departments and also from the AG Technomathe, and the ECMI node Kaiserslautern will be represented at the conference again with a good dozen contributions in total.
So we are trying to keep up our networking in the Technomaths community even in pandemic times. Of course, we all hope that we can meet again at a »real« ECMI conference in 2022. Personally, I'm hoping for the Limerick conference to be made up there soon. I would like to intensify my contacts made there in March 2018 at our SIG workshop on site.