Scholarlink

<p>To address the challenges in quality prediction within cigarette processing production engineering, this study proposes a quality prediction method based on the xLSTM-Transformer model. First, an xLSTM-based feature correlation layer is constructed, utilizing extended memory units and a dynamic projection mechanism to model local temporal dynamics and capture short-term fluctuations as well as inter-stage feature correlations. Next, a Transformer-based feature extraction layer is designed, employing multi-head self-attention with a multi-scale attention mechanism to analyze global interactions among process parameters and unearth deep spatio-temporal features. Subsequently, a gated fusion module adaptively merges local and global features, while a lightweight decoder that combines temporal convolution with sparse attention ensures efficient prediction. Experimental results demonstrate that the proposed model achieves RMSE and MAE values of 0.0165 and 0.0126 respectively on the test set, showing significant improvement in prediction accuracy compared to time-series forecasting models such as iTransformer and RobustTS, while maintaining a stable R² coefficient above 0.95. Through its synergistic mechanism combining local temporal feature enhancement with global interaction modeling, this method innovatively realizes dynamic dimension matching of process parameters and multi-scale feature fusion, thereby providing a novel technical approach for online quality monitoring in batch production processes.</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

XLSTM transformer based quality prediction for tobacco cut rag in intermittent processing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Peipei Li,
  • Peitong Sun,
  • Mingrui Zhu,
  • Jinsong Bao,
  • Xia Chen,
  • Yi Jin

摘要

To address the challenges in quality prediction within cigarette processing production engineering, this study proposes a quality prediction method based on the xLSTM-Transformer model. First, an xLSTM-based feature correlation layer is constructed, utilizing extended memory units and a dynamic projection mechanism to model local temporal dynamics and capture short-term fluctuations as well as inter-stage feature correlations. Next, a Transformer-based feature extraction layer is designed, employing multi-head self-attention with a multi-scale attention mechanism to analyze global interactions among process parameters and unearth deep spatio-temporal features. Subsequently, a gated fusion module adaptively merges local and global features, while a lightweight decoder that combines temporal convolution with sparse attention ensures efficient prediction. Experimental results demonstrate that the proposed model achieves RMSE and MAE values of 0.0165 and 0.0126 respectively on the test set, showing significant improvement in prediction accuracy compared to time-series forecasting models such as iTransformer and RobustTS, while maintaining a stable R² coefficient above 0.95. Through its synergistic mechanism combining local temporal feature enhancement with global interaction modeling, this method innovatively realizes dynamic dimension matching of process parameters and multi-scale feature fusion, thereby providing a novel technical approach for online quality monitoring in batch production processes.