Scholarlink

<p>Proactive forecasting of the dynamic molten pool state in laser metal deposition (LMD) is critical for autonomous feature extraction, early warning, and process analysis. In this study, an encoder–decoder architecture based on Convolutional Long Short-Term Memory network (ConvLSTM-ED) was developed to achieve end-to-end prediction of future molten pool morphology from historical image sequences. Temporal dependencies were primarily captured by the ConvLSTM architecture. Furthermore, a temporal attention mechanism was incorporated, which leveraged global average pooling over spatial and feature dimensions to compute attention scores for each time step. This allowed the model to adaptively focus on the most critical historical frames, thereby enhancing the hierarchical spatiotemporal feature extraction and aiming to improve the recognition of abrupt state transitions. To train the ConvLSTM-ED model, Ti-6Al-4&#xa0;V single-bead LMD experiments were conducted, during which a high-speed thermal imager captured 2300 frames of molten pool videos. Based on this data, convolutional layers hierarchically extracted spatial features, enabling predictive analysis of molten pool behavior under three typical scenarios: stable deposition, laser defocusing, and localized bulging structures on pre-deposited surfaces. A hybrid loss function combining Structural Similarity (SSIM) and L1 loss was employed to jointly constrain perceptual quality and pixel-level accuracy. The proposed ConvLSTM-ED framework was extensively evaluated, and the results showed that the model successfully predicted five subsequent frames from five input frames, achieving a 160 ms advance prediction of molten pool evolution.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic evolution analysis on molten pool morphology in laser metal deposition based on temporal attention augmented

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Jiayu Wang,
  • Jiali Gao,
  • Qin Dong,
  • Yingjie Zhang,
  • Guangyi Ma,
  • Kai Zhao

摘要

Proactive forecasting of the dynamic molten pool state in laser metal deposition (LMD) is critical for autonomous feature extraction, early warning, and process analysis. In this study, an encoder–decoder architecture based on Convolutional Long Short-Term Memory network (ConvLSTM-ED) was developed to achieve end-to-end prediction of future molten pool morphology from historical image sequences. Temporal dependencies were primarily captured by the ConvLSTM architecture. Furthermore, a temporal attention mechanism was incorporated, which leveraged global average pooling over spatial and feature dimensions to compute attention scores for each time step. This allowed the model to adaptively focus on the most critical historical frames, thereby enhancing the hierarchical spatiotemporal feature extraction and aiming to improve the recognition of abrupt state transitions. To train the ConvLSTM-ED model, Ti-6Al-4 V single-bead LMD experiments were conducted, during which a high-speed thermal imager captured 2300 frames of molten pool videos. Based on this data, convolutional layers hierarchically extracted spatial features, enabling predictive analysis of molten pool behavior under three typical scenarios: stable deposition, laser defocusing, and localized bulging structures on pre-deposited surfaces. A hybrid loss function combining Structural Similarity (SSIM) and L1 loss was employed to jointly constrain perceptual quality and pixel-level accuracy. The proposed ConvLSTM-ED framework was extensively evaluated, and the results showed that the model successfully predicted five subsequent frames from five input frames, achieving a 160 ms advance prediction of molten pool evolution.