Scholarlink

<p>Data-driven classification techniques exhibit notable advantages in the quality inspection of plastic products but face challenges such as data imbalance and unclear category boundaries. To address these issues, this paper proposed a semi-supervised method that uses a dual-model structure, leveraging abundant qualified samples to reduce reliance on defective samples. Specifically, qualified samples were divided into two categories based on size distribution, and sub-models based on variational autoencoder and Wasserstein generative adversarial network were trained separately. The outputs were then integrated using weighted aggregation for final classification. Experimental results demonstrated that the dual-model structure improved classification accuracy on the test set by 4.5 % and 6 % compared with single-model structures and by more than 5.3 % compared with the supervised learning approach.</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

Anomaly detection in injection-molded products under data imbalance and fuzzy boundaries

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Xiaoqiang Deng,
  • Wei Xiang,
  • Wenwen Lin,
  • Renhao Yu,
  • Yu Chen,
  • Mengzhong Ji,
  • Lei Wu

摘要

Data-driven classification techniques exhibit notable advantages in the quality inspection of plastic products but face challenges such as data imbalance and unclear category boundaries. To address these issues, this paper proposed a semi-supervised method that uses a dual-model structure, leveraging abundant qualified samples to reduce reliance on defective samples. Specifically, qualified samples were divided into two categories based on size distribution, and sub-models based on variational autoencoder and Wasserstein generative adversarial network were trained separately. The outputs were then integrated using weighted aggregation for final classification. Experimental results demonstrated that the dual-model structure improved classification accuracy on the test set by 4.5 % and 6 % compared with single-model structures and by more than 5.3 % compared with the supervised learning approach.