Real-time defect monitoring in high-power laser-MAG hybrid welding with an improved multi attention mechanisms convolution transformer network

High-power laser-MAG hybrid welding (HPLMHW) offers good penetration and bridging ability for medium-thick plates. However, the instability of intense interaction between the dual heat sources and the weldment causes defects like root humping, with inadequately understood mechanisms and insufficient monitoring of defect types. Additionally, in HPLMHW monitoring, similarities among defects and intense periodic interference make it difficult to monitor accurately and adjust subsequent welding parameters. To address these issues, this research studies and defines the formations and types of root humping in HPLMHW, and proposes an improved transformer-based network for real-time monitoring of five weld types including diverse root humping defects. Firstly, the formation mechanism of diverse root humping in HPLMHW is revealed and diverse categories of root humping are defined. Next, using the established top-view monitoring platform, a dataset of five weld types including distinct root humping is constructed, differing from other welding monitoring studies. Then, the research proposed an enhanced convolutional transformer network for real-time defect monitoring in the mentioned HPLMHW process environment with intense noise interference, periodic objects overlap, and similar top-view formation processes of multi defects, named Sparse and Multi Attention Convolution Transformer Network (SMACTnet). SMACTnet combines sparse and multiple attention mechanisms within Transformer and CNN frameworks. It leverages the advantages of Transformers for global feature extraction and CNNs for local feature extraction, enhances defect class features, minimizes noise and periodic overlap interference, and improves monitoring accuracy. Comparison experiments on a test set of new welds demonstrate that the SMACTnet outperforms other models in overall performance and effectively monitors welds, including incomplete penetration, two types of root humping, surface collapse and well-formed.