CVC Leitprojekt – Nutzung von KI in der Nutzfahrzeugproduktion nimmt Fahrt auf
Exploiting the potential of machine learning in commercial vehicle production
Commercial vehicle production is characterized by industry-specific features that make it difficult to directly transfer solutions from related areas, such as automotive engineering. The most striking characteristic is a high number of variants with frequently small quantities of the individual variants. The resulting increase in complexity in production often leads to increased susceptibility to errors or high costs.
Machine learning (ML) methods, a branch of artificial intelligence, are increasingly being used to master such complex production systems. The aim of these methods is to derive conclusions and forecasts for future processes from large and heterogeneous amounts of data. Due to the advancing digitalization in the context of Industry 4.0, near-real-time sensor and control data also make up a significant proportion of the available data. This increasingly large amount of available data favors the use of ML. While the data for individual variants accumulates in mass production (e.g. in automotive engineering) and thus comparatively large amounts of data are available for learning models, the data volumes in commercial vehicle production are distributed across the entire spectrum. Broken down to the number of different variants, fewer data records are available per component. The resulting increased heterogeneity of the database poses new challenges for the applicability of ML, which generally requires larger homogeneous data volumes to derive meaningful models. While highly specialized and very accurate models are created in mass production due to the large database, the diversity of variants in the commercial vehicle industry requires the development of models with a higher degree of generalizability, which are easier to transfer to other applications. In addition, there is still a high proportion of non-digitized manual activities, which further reduces the amount of available data per variant.
The aim of the lead project is to determine the potentials, obstacles and challenges in the implementation of machine learning solutions, taking into account the specific characteristics of commercial vehicle production, and to create a guideline for manufacturing companies based on this. The ML methods are examined with regard to their ability to handle heterogeneous data, as well as methods for homogenizing or artificially enriching the database (data augmentation).
To achieve this objective, three characteristic use cases were identified in the lead project, each of which raises further research questions in the field of ML.
One of the application problems concerns the optical detection of missing parts; the challenge here is the small database per variant. To counteract this, data augmentation methods are being investigated using heuristic procedures. Another application is the optical quality control of spot welds, which is made more difficult by the high number of variants. Classic feature-based methods expect a variant-specific adaptation, which is associated with a very high manual effort. The use of the Neural Radiance Field (NeRF) offers the possibility of generating any additional perspectives using ML in order to provide the necessary database for differentiating between variants.
The current focus of research is on the potential of the Neural Radiance Field (NeRF) for the generation of virtual models in comparison to approaches from photogrammetry, which extracts a point cloud from a series of images. A neural network learns to reconstruct a 3-dimensional scene, including depth information, from a series of images. The universal use of this technology, combined with simplified data processing, offers great potential for image-based quality control. This is of particular interest for variant-rich production to increase and homogenize the database.
In addition to these research questions, the project also addresses the introduction of machine learning methods in everyday operations. Increasingly better access to computing power and user-friendly software is opening up new and broader fields of application for ML technologies. In addition, the commercial vehicle sector is strongly characterized by small and medium-sized companies, for which independent empowerment without concrete instructions often represents a high risk. This is where the funded project comes in directly through a close exchange with associated partners from the industry and develops practical guidelines for action.
Contact
M.Sc. Patrick Rüdiger
E-mail: patrick.ruediger(at)mv.uni-kl.de
Phone: 0631 205 - 4282