Scholarlink

<p>The production of high-speed wire products (e.g., cord steel, hard wire) requires strict quality control for intra-batch uniformity, inter-batch consistency, and real-time monitoring. However, existing time-series prediction models fail to effectively extract temporal features and can only model single steel grades. To address these limitations, LSTM-SAT-ResNet-MTL is proposed, a neural network combining Long Short-Term Memory (LSTM), Self-Attention, ResNet and Multi-Task Learning (MTL) for real-time mechanical property prediction. The model processes production time-series data (e.g., temperature, rolling parameters) using LSTM to capture temporal dependencies, while self-attention dynamically weights critical process features. MTL leverages correlations between multiple steel grades to enhance prediction robustness. Factory trials on SWRH72A-DY and SWRH77A wires achieved 91.96% and 96.33% prediction accuracy rates (within ± 3% error tolerance), respectively, significantly reducing reliance on manual testing. This multi-grade approach enables real-time prediction, substantially improving efficiency while meeting stringent industrial quality standards.</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

Mechanical property prediction of high-speed wire based on multi-modal deep network with multi-task learning

  • 期刊论文

  • 发表日期：

    2026年1月31日

查看全文
  • Fei He,
  • Yimin Ding,
  • Baole Lou,
  • Yungui Zhang

摘要

The production of high-speed wire products (e.g., cord steel, hard wire) requires strict quality control for intra-batch uniformity, inter-batch consistency, and real-time monitoring. However, existing time-series prediction models fail to effectively extract temporal features and can only model single steel grades. To address these limitations, LSTM-SAT-ResNet-MTL is proposed, a neural network combining Long Short-Term Memory (LSTM), Self-Attention, ResNet and Multi-Task Learning (MTL) for real-time mechanical property prediction. The model processes production time-series data (e.g., temperature, rolling parameters) using LSTM to capture temporal dependencies, while self-attention dynamically weights critical process features. MTL leverages correlations between multiple steel grades to enhance prediction robustness. Factory trials on SWRH72A-DY and SWRH77A wires achieved 91.96% and 96.33% prediction accuracy rates (within ± 3% error tolerance), respectively, significantly reducing reliance on manual testing. This multi-grade approach enables real-time prediction, substantially improving efficiency while meeting stringent industrial quality standards.