The tensile failure of composites is governed by the development of fibre breaks, which increase stresses in intact neighbouring fibres. Critical failure is reached once a configuration of fibre breaks (a “cluster”) increases the stresses in its intact neighbours to an extent that no fibre can bare the additional load, hence the cluster propagates unstoppable throughout the composite. Current state-of-the-art strength prediction models are able to predict the development of fibre breaks, but more experimental results are needed for their validation. Observed differences indicate a lack of understanding of the influence of the local micro-structure on the development of breaks. By providing access to fibre trajectories during in-situ tensile testing, the performed study offers new possibilities for a more sound implementation of the micro-level composite behaviour into strength prediction models by simulating real composite micro-structures. The observed fibre break development can be used to setup fibre strength distributions as crucial input parameter. This will help to improve existing strength prediction models and in the long-term provide reliable virtual testing capabilities for the design of novel high performance composites. Image analysis pipeline: The acquired projections were first reconstructed with a filtered back projection algorithm provided by ESRF (European Synchrotron Radiation Facility). The high fidelity of the reconstructed images reveals many fibre breaks, which can be manually annotated in the consecutive volumes. The rigid body movement of the volumes can be compensated due to the observed movement of pairs of fibre breaks. By adjusting the histogram threshold, dark objects can be filtered from the volume, which would typically consist of voids and different types of cracks within the volume. Potential fibre breaks can be filtered based on their small volume and penny shape and by comparison with the annotated coordinates. For the fibre segmentation the InSegt tool (provided by Emerson et al. from DTU) is trained using a representative sample image. The dictionary built up from the annotated fibre centres and matrix areas is used to identify the fibre paths in the first acquired volume (specimen just under a small preload). The algorithm processes a predefined number of the transverse slices (usually 10-20% of the total) and returns probability density values for the two annotated categories for each voxel. These values are filtered using a threshold probability and analysed with connected component analysis. The centroids of the identified objects are grouped into fibres, providing a fibre trajectory by connecting them with an interpolation function. Using the mean fibre diameter from laser micro-metre measurements, the original 3D volume of the fibres is computed. Comparison of the fibre and fibre break object volumes reveals their overlap and allows to assign fibre break objects to their corresponding fibres. Repeating this step for all acquired volumes allows to track the fibre break development from volume to volume–hence offering time/strain dependent analysis of the effect of fibre re-alignment and the local state of the influence zone around fibre breaks.