Project description

A key challenge for the use of data-based methods for the analysis and optimization of processes in electronics production is the imbalance of the data. Despite the large amounts of data generated in high-volume production, the information content of the data is limited, as the vast majority of inspections do not detect any errors. In order to avoid error slippage, the automated inspection steps, which are usually based on conventional image processing methods, are set so sensitively that there is usually a relatively high rate of pseudo errors. These components actually meet the specifications, but are incorrectly classified as faulty, resulting in a high manual inspection effort.

In order to reduce this effort, components identified as faulty can be rechecked using AI methods so that some of the pseudo faults are recognized, which results in a reduction of the testing effort. The quality of the AI methods used crucially depends on the amount and quality of the available data. In order to have a large database available despite the unbalanced data situation in electronics production, the research project ColEP is investigating the advantages and information technology solutions for the collaborative utilization of process, quality and metadata from electronics manufacturers and system manufacturers. Selected AI use cases in electronics production will be used to analyze potential improvements for which collaborative data sets appear particularly suitable. In this regard, the data is automatically anonymized and processed to ensure basic security. Furthermore, a framework will be developed using the GAIA-X concepts to ensure the sovereignty of the data for the owner.

In this context, both frameworks for federated learning and the integration into existing sovereign data infrastructures are examined. In contrast to centralized learning, where data from different sources is collected on a central server, federated learning keeps training data local to the participating companies at all times, ensuring data security and sovereignty. Model training takes place locally, after which the trained models are sent to a central server, which combines them into a global model.

The goal of the project is to investigate the applicability of federated learning in an industrial context using use cases in electronics production. It aims towards comparing the model performance of federated learning with centralized and local learning. In addition, data protection-specific advantages of federated learning will be investigated and security-relevant challenges such as data manipulation and reconstruction of training data will be critically examined and prevented using suitable methods.