Scholarlink

<p>This study proposes an acoustic-based approach for diagnosing faults in centrifugal pumps by leveraging optimized ensemble learning models. Acoustic signals recorded under natural operating conditions were categorized into three classes: healthy, slightly faulty, and faulty. A set of key features was extracted from the time, frequency, and time-frequency domains to construct the dataset. Five ensemble classifiers—gradient boosting (GBM), AdaBoost, stacking, bagging, and voting—were optimized using the most suitable hyperparameters identified through grid search, and these classifiers were subsequently trained and evaluated using 10-fold cross-validation. All models, particularly GBM and AdaBoost, achieved high classification accuracy exceeding 99 %, demonstrating their effectiveness. Although the “slightly faulty” class posed challenges due to overlapping feature distributions, results confirm the robustness and reliability of ensemble models under real industrial conditions. Compared with vibration-based methods, the proposed approach offers a more cost-effective and practical alternative, enhancing predictive maintenance by enabling early and accurate fault detection in pump components.</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

Acoustic-based diagnosis of impeller and bearing faults in centrifugal pumps using optimized ensemble learning models

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Savaş Koç,
  • Idris Saçaklıdır

摘要

This study proposes an acoustic-based approach for diagnosing faults in centrifugal pumps by leveraging optimized ensemble learning models. Acoustic signals recorded under natural operating conditions were categorized into three classes: healthy, slightly faulty, and faulty. A set of key features was extracted from the time, frequency, and time-frequency domains to construct the dataset. Five ensemble classifiers—gradient boosting (GBM), AdaBoost, stacking, bagging, and voting—were optimized using the most suitable hyperparameters identified through grid search, and these classifiers were subsequently trained and evaluated using 10-fold cross-validation. All models, particularly GBM and AdaBoost, achieved high classification accuracy exceeding 99 %, demonstrating their effectiveness. Although the “slightly faulty” class posed challenges due to overlapping feature distributions, results confirm the robustness and reliability of ensemble models under real industrial conditions. Compared with vibration-based methods, the proposed approach offers a more cost-effective and practical alternative, enhancing predictive maintenance by enabling early and accurate fault detection in pump components.