ZEUS – Data-driven process monitoring methods

In the joint project "LuFo-ZEUS" "Zero Emission aircraft with sustainable fuselage concept and technology", Leibniz-Institut für Verbundwerkstoffe (IVW) is working on the development of integral construction methods, simulation methods and manufacturing processes for thermoplastic fuselage components of the future, such as the thermoplastic door environment and thermoplastic profiles. A central focus is also on innovative, data-driven methods for process monitoring. Data collected in a manufacturing process is to be used to train an artificial intelligence (AI) that will then enable a robust prediction of the component quality achieved in situ during the utilization phase. The process selected for this is intermittent hot pressing. Its incremental mode of operation is ideal for characterizing and evaluating individual production stages and thus generating many data sets.

There are three main requirements for implementation. First, it must be ensured that the collected data provides sufficient information content in terms of both scope and quality to enable the training of a powerful, reliable and meaningful AI. Only if the collected data is sufficiently related to production quality, a reliable statement can be made. In order to ensure that sufficient data is available for the interval hot-pressing process under consideration, close coordination is taking place with the project partner XELIS, a manufacturer of thermoplastic profiles with carbon and glass fibers and an experienced user of interval hot-pressing for decades and a global market leader in this segment. The unique processes and systems enable the production of endless FRP profiles in almost any configuration. In the project, their long experience with the production process forms the basis for the selection of exactly these decisive data.

One focus of the collaboration in the project is on the preparation of the data. For example, it must be evaluated whether data points influence each other and whether the process is subject to relevant temporal fluctuations. Powerful, relatively simple models such as K-Means clustering or support vector machines can be used for independent data. If, on the other hand, there are correlations in the sequence of the data, more complex AI approaches such as neural networks (NN) with closed recurring units (GRU) or long short-term memory units (LSTM) and vector autoregressive models must be used. These more complex machine learning methods require a larger data set than the methods mentioned above in order to represent a model with a comparably high scope of influences under consideration. Due to the limited amount of data collected in "LuFo-ZEUS", this must then be supplemented by existing knowledge about the process. This can be introduced in the form of a "Physics Informed Neural Network" (PINN) or by further results from other IHP processes by means of transfer learning. The use of individual sub-models for carefully selected correlations also represents an opportunity to use the knowledge of the process from the wealth of experience of the project partners to investigate more complex correlations with a limited data set. The selection of the methods to be finally used can only be made taking into account the available process data, which is why their preparation is a crucial step.

If this is the case, some popular AI models such as K-means clustering or support vector machines can only be used to a very limited extent. Even neural networks (NN) with closed recurring units (GRU) or "long short-term memory" units (LSTM), as well as vector autoregressive models, can only be used for single problems of low complexity.

The second major challenge is the efficient and targeted use of the data obtained to train a wide variety of AI models while simultaneously considering additional options such as the use of federated or transfer learning. This requires many years of expertise in the versatile, production-related use of a wide range of AI solutions. Here, the strategic cooperation with the Chair of Machine Tools and Controls (WSKL) at RPTU Kaiserslautern-Landau and its unique experience represents an important component on the path to successful AI application.

The last crucial task is the reliable assessment of the process results based on reliable, classical and established measurements on the produced components. This guarantees the quality of the training data sets and allows a genuine transition from a purely academic view of the problem to solving the challenges faced by an industrial end user. This task is performed by the project partner Airbus, Europe's largest aerospace company, which has unique expertise in the development and evaluation of FRP processes and components in aviation.

Together, the aim is to establish the correlation and train a powerful AI that will deliver significant added value in the future: more robust processes can save money, not only through more efficient use of materials, but also by reducing the cost of supporting the process. Once the leap to advanced AI methods has been made, these can be developed further instead of being built from scratch. As a result, the application becomes simpler each time and requires less and less human intervention. Once the application is fully established, the number of measurement methods involved can also be reduced. This is highly relevant for the expensive in situ measurements and even more so for the time-consuming ex situ tests. Together with its project partners, Leibniz-Institut für Verbundwerkstoffe is thus taking a major step towards the production of the future.

The project "LuFo-ZEUS" Zero Emission aircraft with sustainable fuselage concept and technology is funded by the Federal Ministry of Economics and Climate Protection (BMWK) on the basis of a decision by the German Bundestag, funding reference: 20W2106F.