Towards Online-Prediction of Quality Features in Laser Fusion Cutting Using Neural Networks

The fine-scaled striation structure as a relevant quality feature in laser fusion cutting of sheet metals cannot be predicted from online process signals, today. High-speed recordings are used to extract a fast melt-wave signal as temporally resolved input signal and a surrogate surface profile as output. The two signals are aligned with a sliding-window algorithm and prepared for a one-step ahead prediction with neural networks. As network architecture a convolutional neural network approach is chosen and qualitatively checked for its suitability to predict the general striation structure. Test and inference of the trained model reproduce the peak count of the surface signal and prove the general applicability of the proposed method. Future research should focus on enhancements of the neural network design and on transfer of this methodology to other signal sources, that are easier accessible during laser cutting of sheet metals.