Scholarlink

<p>In this study, a prediction method based on a data augmentation strategy using GANs Generative Adversarial Networks (GANs) and DBN (Deep Belief Network) algorithms is proposed for predicting D-type inclusions in the electroslag remelting (ESR) process. The study focuses on the ESR process, with data collected from the production line, comprising over 2000 samples. A GANs-based data augmentation strategy is employed to enhance the model’s accuracy and generalization ability, increasing the dataset size to 10,000 samples. Comprehensive data preprocessing is conducted on the expanded dataset. Using this augmented dataset, five predictive models for D-type inclusions are constructed with DBN and other algorithms. The performance of these five models is systematically compared through evaluation metrics such as confusion matrices, ROC curves, and prediction accuracy. To analyze the impact of process parameters on inclusion prediction results, the SHapley Additive exPlanations (SHAP) interpretability framework is introduced for model explanation. This analysis reveals the key factors influencing D-type inclusions and their mechanisms, providing new perspectives and approaches for predicting inclusions in the ESR process.</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

Interpretable Prediction Model Based on GANs–DBN Data Enhancement Strategy for Electroslag Remelting Inclusions

  • 期刊论文

  • 发表日期：

    2025年4月3日

查看全文
  • Yuxiao Liu,
  • Yanwu Dong,
  • Zhouhua Jiang,
  • Xi Chen

摘要

In this study, a prediction method based on a data augmentation strategy using GANs Generative Adversarial Networks (GANs) and DBN (Deep Belief Network) algorithms is proposed for predicting D-type inclusions in the electroslag remelting (ESR) process. The study focuses on the ESR process, with data collected from the production line, comprising over 2000 samples. A GANs-based data augmentation strategy is employed to enhance the model’s accuracy and generalization ability, increasing the dataset size to 10,000 samples. Comprehensive data preprocessing is conducted on the expanded dataset. Using this augmented dataset, five predictive models for D-type inclusions are constructed with DBN and other algorithms. The performance of these five models is systematically compared through evaluation metrics such as confusion matrices, ROC curves, and prediction accuracy. To analyze the impact of process parameters on inclusion prediction results, the SHapley Additive exPlanations (SHAP) interpretability framework is introduced for model explanation. This analysis reveals the key factors influencing D-type inclusions and their mechanisms, providing new perspectives and approaches for predicting inclusions in the ESR process.