Scholarlink

<p>In smart manufacturing monitoring platforms, real-time sensor data must be labeled either manually or automatically before they can be used for training and inference with Artificial Intelligence (AI) prediction models. However, such data often suffer from noisy labels arising from sensor inaccuracies, human annotation bias, or noise in automated labeling systems, which significantly degrades the performance of AI models. To address this issue, this paper proposes a label noise suppression method based on ensemble learning, with the aim of improving the quality of real-time sensor data and the effectiveness of AI prediction models in smart manufacturing monitoring systems. The proposed approach employs multiple machine learning classification algorithms for training and recognition using the same dataset, and then applies confidence scores and majority voting strategies to identify and correct potential noisy labels. The refined labels are subsequently used to retrain the AI prediction model, thereby mitigating the adverse impact of noise on the overall recognition performance. To validate the effectiveness of the proposed method, an empirical evaluation is conducted on a smart manufacturing monitoring platform deployed in a rubber and plastic production line. Moreover, previous studies have primarily focused on injection molding machines, which typically operate with longer production cycle times. The label anomalies considered in this paper occur more frequently, on machines with much shorter cycles, meaning that a significantly larger amount of vibration data can be accumulated within the same time period. Since research involving vibration analysis of linear stretch blow molding machines is relatively scarce, real-world experiments and analysis are conducted in this study, on a linear stretch blow molding machine deployed on an actual production line. Experimental results demonstrate that the proposed method outperforms traditional single or multiple machine learning classifiers by approximately 1% to 5% in terms of recognition accuracy, thereby indicating its robustness in handling noisy label data.</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

Application of ensemble learning for label noise mitigation in intelligent manufacturing monitoring platforms

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Ping Chen,
  • Song-Yen Hsu,
  • Ming-Fong Tsai,
  • Lien-Wu Chen

摘要

In smart manufacturing monitoring platforms, real-time sensor data must be labeled either manually or automatically before they can be used for training and inference with Artificial Intelligence (AI) prediction models. However, such data often suffer from noisy labels arising from sensor inaccuracies, human annotation bias, or noise in automated labeling systems, which significantly degrades the performance of AI models. To address this issue, this paper proposes a label noise suppression method based on ensemble learning, with the aim of improving the quality of real-time sensor data and the effectiveness of AI prediction models in smart manufacturing monitoring systems. The proposed approach employs multiple machine learning classification algorithms for training and recognition using the same dataset, and then applies confidence scores and majority voting strategies to identify and correct potential noisy labels. The refined labels are subsequently used to retrain the AI prediction model, thereby mitigating the adverse impact of noise on the overall recognition performance. To validate the effectiveness of the proposed method, an empirical evaluation is conducted on a smart manufacturing monitoring platform deployed in a rubber and plastic production line. Moreover, previous studies have primarily focused on injection molding machines, which typically operate with longer production cycle times. The label anomalies considered in this paper occur more frequently, on machines with much shorter cycles, meaning that a significantly larger amount of vibration data can be accumulated within the same time period. Since research involving vibration analysis of linear stretch blow molding machines is relatively scarce, real-world experiments and analysis are conducted in this study, on a linear stretch blow molding machine deployed on an actual production line. Experimental results demonstrate that the proposed method outperforms traditional single or multiple machine learning classifiers by approximately 1% to 5% in terms of recognition accuracy, thereby indicating its robustness in handling noisy label data.