Autor: M. Schiele

Mask R-CNN has recently achieved great success in the field of instance segmentation. However, weaknesses of the algorithm have been repeatedly pointed out as well, especially in the segmentation of long, sparse objects whose orientation is not exclusively horizontal or vertical. We present here an approach that significantly improves the performance of the algorithm by first pre-segmenting the images with a PSPNet algorithm. To further improve its prediction, we have developed our own cost functions and heuristics in the form of training strategies, which can prevent so-called (early) overfitting and achieve a more targeted convergence. Furthermore, due to the high variance of the images, especially for PSPNet, we aimed to develop strategies for a high robustness and generalization, which are also presented here.

In this work, a dissimilar copper/aluminum lap joint was generated by force-controlled hybrid friction diffusion bonding setup (HFDB). During the welding process, the appearing torque, the welding force as well as the plunge depth are recorded over time. Due to the force-controlled process, tool wear and the use of different materials, the resulting dataseries varies significantly, which makes quality assurance according to classical methods very difficult. Therefore, a Convolutional Neural Network was developed which allows the evaluation of the recorded process data. In thisstudy, data from sound welds as well as data from samples with weld defects were considered. In addition to the different welding qualities, deviations from the ideal conditions due to tool wear and the use of different alloys were also considered. The validity of the developed approach is determined by cross validation during the training process and different amounts of training data. With an accuracy of 88.5%, the approach of using Convolutional Neural Network has proven to be a suitable tool for monitoring the processes.

The growth of electric propulsion systems motivates the automotive industry to transfer the focus from exhaust to non-exhaust emissions, with special attention to brake-related emissions. The literature lacks well-established approaches that describe the particulate emissions through reliable analytical correlations. Moreover, the mechanisms of brake particulate formation entail highly stochastic phenomena, which cannot be captured by means of traditional deterministic modelling tools. Machine learning algorithms have been recently used as an alternative method to seek for a branched correlation between tribological properties (i.e. friction coefficient and wear rate), pad composition, environmental and operating conditions. In this regard, the presented work focuses on the study and identification of sophisticated stochastic meta-models for the prediction of the number of emitted brake particulate and associated uncertainty. Specifically, artificial neural networks are developed and validated against brake emission data collected in real driving conditions at Technische Universität Ilmenau. The developed algorithms are intended for multiple use: (i) in the course of real driving emissions (RDE) testing, to support the experimental data; (ii) while driving, to inform the driver about the brake-related emission levels; (iii) as an on-board optimisation tool that identifies the brake actuation rules to minimise the release of particulate emissions.