Scholarlink

<p>The plastic injection molding industry represents a major manufacturing domain requiring both high-volume production and precise quality control. Recently, global production strategies have been adopted, enabling geographically distributed factories to produce identical products in parallel to meet local demand efficiently. However, ownership boundaries, legal restrictions, and competitive interests often limit data exchange among factories, making centralized learning (CL) impractical. To address this challenge, this study applied federated learning (FL) to injection molding processes to verify whether quality prediction models (QPMs) can be developed under data privacy constraints. Data collected from a single factory were used to simulate multi-factory conditions by varying temperature, humidity, and process parameters. Three data partitioning scenarios were designed, ranging from independent and identically distributed (IID) to highly Non-IID conditions. The study analyzed FL performance with respect to four factors: comparison with CL under different training data sizes, variation in the number of clients, changes in local data volume per client, and aggregation methods including Federated Averaging (FedAvg), Federated Matched Averaging (FedMA), and Federated Batch Normalization (FedBN). Results showed that FL effectively mitigated the limitations of CL in data-scarce or heterogeneous manufacturing environments. Increasing the number of clients and enlarging local datasets improved model performance and stability. FedAvg and FedBN achieved stable performance under Non-IID conditions. These findings demonstrate that FL can serve as a practical alternative for building QPMs in the global injection molding industry by overcoming restrictions on data sharing and supporting rapid adaptation in newly established factories.</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

Federated Learning for Privacy-Preserving Quality Prediction in the Globalized Plastic Injection Molding Industry

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Joon-Young Kim,
  • Songho Lee,
  • Jecheon Yu,
  • Junhyeong Lee,
  • Seunghwa Ryu

摘要

The plastic injection molding industry represents a major manufacturing domain requiring both high-volume production and precise quality control. Recently, global production strategies have been adopted, enabling geographically distributed factories to produce identical products in parallel to meet local demand efficiently. However, ownership boundaries, legal restrictions, and competitive interests often limit data exchange among factories, making centralized learning (CL) impractical. To address this challenge, this study applied federated learning (FL) to injection molding processes to verify whether quality prediction models (QPMs) can be developed under data privacy constraints. Data collected from a single factory were used to simulate multi-factory conditions by varying temperature, humidity, and process parameters. Three data partitioning scenarios were designed, ranging from independent and identically distributed (IID) to highly Non-IID conditions. The study analyzed FL performance with respect to four factors: comparison with CL under different training data sizes, variation in the number of clients, changes in local data volume per client, and aggregation methods including Federated Averaging (FedAvg), Federated Matched Averaging (FedMA), and Federated Batch Normalization (FedBN). Results showed that FL effectively mitigated the limitations of CL in data-scarce or heterogeneous manufacturing environments. Increasing the number of clients and enlarging local datasets improved model performance and stability. FedAvg and FedBN achieved stable performance under Non-IID conditions. These findings demonstrate that FL can serve as a practical alternative for building QPMs in the global injection molding industry by overcoming restrictions on data sharing and supporting rapid adaptation in newly established factories.