Advanced Deep Learning Techniques for Real-Time Defect Detection and Quality Control in Laser Cutting Processes
Parveen Kumar and Sumitkumar Rathor
Laser cutting is a high-precision manufacturing technique that relies on the stability of a focused laser beam and an assist gas to produce intricate geometries with minimal mechanical stress. However, fluctuations in beam parameters and gas flow can destabilize the keyhole, leading to defects such as kerf-width deviation, dross attachment, heat-affected zone irregularities, and micro-cracks. Conventional rule-based inspection methods image thresholding integrate multi‐modal in-process sensing (optical, thermal, acoustic) with post-process machine vision, and employ state-of-the-art convolutional neural networks (U-Net, Mask R-CNN) and hybrid CNN–Transformer architectures for pixel-level segmentation and classification of defects. Self- and unsupervised learning strategies reduce annotation overhead by modelling the manifold of “good” cuts and flagging deviations in real time. Our edge-AI implementation achieves inference times under 30 ms, enabling closed-loop feedback to CNC controllers that adjust process parameters on the fly. Benchmarking on publicly available metal‐surface datasets demonstrates detection accuracies > 95%, Intersection over Union scores > 0.80, and F1-scores > 0.90. We further analyse the hardware and software requirements for industrial deployment, quantify the cost–benefit, and address integration challenges with MES/SCADA systems. Finally, we outline future directions miniaturized edge-AI modules, Deep Learning based AI diagnostics, and digital-twin coupling to realize zero-defect, self-optimizing laser-cutting cells in smart factories.
Keywords: Deep learning, defect detection, laser cutting, quality control, realtime monitoring