Image-based modelling and visualisation of the relationship between laser-cut edge and process parameters. (September 2021)

• Record Type: Journal Article
• Authors: Tatzel, Leonie; Tamimi, Omar Al; Haueise, Tobias; Puente León, Fernando
• Is Part Of: Optics & laser technology. Volume 141(2021)
• Publication Date: 2021-09
• Subjects: Laser cutting -- Machine status -- Process parameter regression -- Convolutional neural network (CNN) -- Layer-wise relevance propagation (LRP); Optics -- Periodicals; Lasers -- Periodicals; Electronic journals; 621.366
• DOI: 10.1016/j.optlastec.2021.107028
• Languages: English
• ISSNs: 0030-3992

Abstract

Highlights: Artificial neural network deduces process parameters from image of laser-cut edge. Network estimates the focus position with a mean error of 0.2 mm. Broad training database containing 3336 stainless-steel cut edges was generated. Layer-wise relevance propagation provides insights into neural network. Approach enables automated condition monitoring of laser cutting machines. Abstract: This article presents a new way of evaluating the laser cutting process. We show that it is possible to deduce the underlying process parameters directly from the laser-cut edge using a convolutional neural network (CNN). For this purpose, we developed a suitable CNN architecture and generated a broad database of 3336 stainless steel (1.4301) edges that were cut with different combinations of four process parameters. RGB images and 3D point clouds of the edges were used as input to the network, and the process parameters were the regression targets (output). We found that the CNN estimates the process parameters well and performs better on the RGB images. The mean error is 1.1m/min, or 7% of the range, for the feed rate and 0.2mm, or 4% of the range, for the focus position. The proposed method could be used to monitor the condition of a laser cutting machine by evaluating an image of a cut edge. Because defective machine components can cause the actual process parameters (in the sheet metal) to differ from the set values, they can be identified quickly by comparing the CNN output with the chosen settings. As our approach offers a new perspective on the process, the visualisation of the CNN might offer a better understanding of the process. We applied layer-wise relevance propagation to visualise the relationship between each individual pixel of the input image and the output of the CNN. We show the potential of this technique with some examples.