News on China's scientific and technological development.
- Thread starter Quickie
- Start date
China's first commercial 2x6MV (12-megavolt) tandem accelerator recently completed assembly and "cold commissioning" at the Accelerator R&D Center of the Ruichang Institute of Applied Nuclear Physics in Jiangxi Province. With its key performance indicators meeting design specifications, the unit has officially rolled off the production line, marking a significant milestone in the development of domestically produced, high-end scientific research equipment.
Shanghai Institute of Optics and Fine Mechanics has made new progress in laser welding process monitoring.
Recently, the team led by Professor Yang Shanglu of the Laser Intelligent Manufacturing Technology R&D Center of the High-end Optoelectronic Equipment Department of the Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, has made progress in laser welding process monitoring. The relevant results, entitled " An explainable multi-modality fusion framework for laser weld penetration classification of Al–Si coated PHS 22MnB5" , were published in Optics & Laser Technology .
As high-end manufacturing demands increasingly higher welding quality and reliability, monitoring the weld penetration during laser welding has become a crucial aspect of ensuring structural safety and service performance. However, traditional methods relying on single sensor signals and manual feature engineering are insufficient to fully characterize the complex physical behavior of the welding process, thus limiting the improvement of online precision identification capabilities.
To address these challenges, the research team constructed a multi-sensor fusion monitoring platform that simultaneously acquires acoustic signals, plasma spectral signals, and molten pool image signals during the welding process, establishing a multimodal dataset covering four types of weld depth states. Based on this, they innovatively proposed a hybrid model framework (Fusion-XGBNet) combining a neural network ( NN ) and extreme gradient boosting ( XGBoost ) . This model utilizes neural networks to achieve automatic feature extraction and deep fusion of raw multi-source data, and introduces XGBoost for efficient discrimination of fused features. Finally, an adaptive strategy enables collaborative decision-making between the two model outputs, overcoming the reliance on manual feature engineering in traditional methods.
A multimodal machine learning framework for fusion state classification (Fusion-XGBNet)
Experimental results show that the fusion model achieved an average recognition accuracy of 95.29% on the test set. Furthermore, using Grad-CAM++ visualization analysis, it was clearly identified that the molten pool contour region and the characteristic spectra near 367 nm , 437 nm , 580 nm , and 622 nm make crucial contributions to the model's decision-making, significantly enhancing the model's interpretability and reliability for engineering applications.
supercat
Colonel
Huawei's net profit grew 8.6% in 2025.
