Recycled endless fiber-reinforced thermoplastic composites offer a valuable source of high-quality short fibers for new reinforcing materials. However, the final product performance is significantly influenced by length and distribution of these recycled fibers.

State-of-the-art methods often struggle with crossed or incomplete (cut-off) fibers, thus they are limited to analyzing non-crossed, singular fibers, and need numerous measurements with continuous stirring of the fiber-water mixture, which can degrade fibers and requires high dilution rates. Obviously, determining fiber length distribution from images with a high fiber density would be beneficial: The appearance of fibers in such optical measurements is highly variable, influenced by the fiber lengths, the local concentration, the diluting medium (typically H2O), the number of crossed or cut-off fibers and the measurement methods used.

Our solution implements a three-step machine learning approach that systematically addresses the challenges of fiber analysis: First, a regression model determines optimal image processing settings based on fiber characteristics and image properties, automatically adjusting parameters such as contrast thresholds, noise reduction levels, and edge detection sensitivity to accommodate varying image qualities. Second, a binary CNN classifier differentiates between simple, non-crossed fibers and complex fiber clusters, enabling efficient processing paths for each fiber type. Finally, for crossed-fiber structures, a UNET++ architecture with skip connections segments individual fibers by identifying junction points and potential fiber paths, effectively resolving even complex overlapping scenarios. The network analyzes local features including fiber angles, orientation continuity, and gradient transitions to accurately reconstruct original fibers from overlapping segments, utilizing a custom loss function that penalizes discontinuities and implausible fiber geometries. This integration of complementary ML techniques enables robust performance across diverse imaging conditions without manual parameter tuning.

This optimized approach enables reliable, rapid, and highly accurate measurement of prepared fiber solutions, even in the presence of complex fiber crossings and convoluted images. Our approach significantly outperforms conventional image processing algorithms in terms of reproducibility and the total number of fibers measured, as validated across diverse fiber measurement systems and image types.

Keywords: fibre length measurement, crossed-fibre analysis, machine learning, CNN classification, UNET++ segmentation, recycling, reinforced glass fibres
Preferred Lecture Type: On-Site Lecture in Bochum