Scholarlink

<p>During dynamic roll forming process, downward pressure parameters nonlinearly affect the final radius of curvature, making accurate prediction a significant challenge in precision molding. To address this issue, this paper proposes a hybrid neural network model, ARIMA-TCM-XLSTM-Informer, designed to efficiently model the complex temporal dependencies in the roll forming process. First, the ARIMA module captures the short-term linear trends in the material deformation sequence by focusing on the elastic response and transient stress changes occurring during the early stages of forming. Next, the TCM module identifies the dynamic evolution of local buckling and residual stress by using a multi-scale convolutional structure to extract local nonlinear features and short-term dependencies, thus enhancing the model’s sensitivity to local changes. Furthermore, the XLSTM module improves the ability to characterize long-term dependencies by modeling the roll-over process using an enhanced long-and-short-term memory structure. This enables accurate representation of the stress-strain history and cumulative deformation features during roll-over. Finally, the Informer module optimizes long-sequence modeling by employing a sparse self-attention mechanism to capture remote correlations and trend changes in the forming sequence, which improves the model’s predictive capability for critical forming steps. The experimental results show that the ARIMA-TCM-XLSTM-Informer model can effectively capture the linear and nonlinear dependencies of the molding sequence data. The comparison with other advanced models on the physical experiment dataset demonstrates the proposed model’s accurate prediction ability of the curvature radius.</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

Prediction of curvature radius in roll-bending forming based on the ARIMA-TCM-XLSTM-Informer neural network model

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Menghang Wan,
  • Guoyan Huang,
  • Tong Wu,
  • Yafeng Zhang,
  • Peng Shi

摘要

During dynamic roll forming process, downward pressure parameters nonlinearly affect the final radius of curvature, making accurate prediction a significant challenge in precision molding. To address this issue, this paper proposes a hybrid neural network model, ARIMA-TCM-XLSTM-Informer, designed to efficiently model the complex temporal dependencies in the roll forming process. First, the ARIMA module captures the short-term linear trends in the material deformation sequence by focusing on the elastic response and transient stress changes occurring during the early stages of forming. Next, the TCM module identifies the dynamic evolution of local buckling and residual stress by using a multi-scale convolutional structure to extract local nonlinear features and short-term dependencies, thus enhancing the model’s sensitivity to local changes. Furthermore, the XLSTM module improves the ability to characterize long-term dependencies by modeling the roll-over process using an enhanced long-and-short-term memory structure. This enables accurate representation of the stress-strain history and cumulative deformation features during roll-over. Finally, the Informer module optimizes long-sequence modeling by employing a sparse self-attention mechanism to capture remote correlations and trend changes in the forming sequence, which improves the model’s predictive capability for critical forming steps. The experimental results show that the ARIMA-TCM-XLSTM-Informer model can effectively capture the linear and nonlinear dependencies of the molding sequence data. The comparison with other advanced models on the physical experiment dataset demonstrates the proposed model’s accurate prediction ability of the curvature radius.