Partnership for Applied AI Research

RTE and the Fraunhofer Institute for Digital Media Technology IDMT in Ilmenau have entered into a three-year research collaboration. The focus is on developing self-supervised learning (SSL) methods for acoustic part inspection.

The Problem: Training Data Is Expensive

Classical supervised ML methods require large amounts of labeled training data – typically several hundred to thousands of OK and NOK samples. For new components or defect patterns, this means weeks of data collection under production conditions. This significantly delays the ramp-up of new test stations.

The Approach: Self-Supervised Learning

SSL methods leverage the structure of unlabeled data to learn meaningful feature representations. The model recognizes patterns and anomalies without requiring each sample to be manually classified. This is particularly promising for acoustic testing, since OK parts are typically available in abundance.

Project Goals

  • Reduction of required labeled training samples by at least 50 %
  • Comparable or better selectivity versus current supervised methods
  • Transferability of models between similar component groups
  • Integration of results into the SonicTC platform as a "quick-start" training mode

Why Fraunhofer IDMT?

IDMT has decades of expertise in acoustic signal processing and machine learning. The "Industrial Media Technology" department has already conducted numerous projects in acoustic production monitoring. The combination of Fraunhofer's fundamental research and RTE's practical experience from over 200 industrial projects forms an ideal foundation for application-oriented results.