Full penetration detection in fiber laser remote cutting of electrical steel sheets from <i>in situ</i> monitored images using convolutional neural networks

In laser remote cutting, a cut through is usually realized by repeating a scanning process for several ablation cycles. The material removal rate depends primarily on the laser power, feed rate, and number of cycles. Too few cycles will not achieve a full penetration, whereas too many cycles may cause the workpiece to reweld on the sheet. In either case, the workpiece cannot be detached from the sheet. In this contribution, we present an in situ monitoring system using convolutional neural networks (CNNs), from recordings of the process zone during remote cutting, to detect full penetration in thin electrical sheets. The monitoring system, based on a high-speed camera, is attached to a conventional laser scanner, whereby the optical detection path is aligned coaxially to the incident laser. Images with a size of 256 × 256 px2 are recorded with a sample rate of 20 × 103 fps and an exposure time of 20 μs. Without external illumination, the thermal radiation of the process zone is captured in the VIS spectrum. A dataset of images is labeled for cycles before a cut through, during a full penetration and for too many cycles, to train the networks. As a result, a full penetration is recognized by the trained CNN models, and the detection accuracies of AlexNet, VGGNet, and ResNet are compared.