Scholarlink

<p>The continuous operation of single-phase pumps is essential in industrial environments, yet their reliability is often compromised by mechanical and electrical faults that lead to unplanned downtime, higher maintenance costs, and production losses. This study proposes a multivariate Space Vector Machine with Error-Correcting Output Codes (SVM-ECOC) that integrates vibration, current, and temperature signals to detect and diagnose failures in pumps used for level control in storage tanks. The method addresses common fault conditions, including misalignment, voltage drop, cavitation, level-sensor bias, and proportional-valve bias. The acquired signals are processed through feature extraction in the time, frequency, and time–frequency domains, followed by normalization and dimensionality reduction. Classifiers are trained using a 70–30 partition and evaluated through confusion matrices, accuracy, macro-averaged F1 score, and computational performance metrics on an ESP32 microcontroller. To assess the effectiveness of the proposed approach, its performance was compared with Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) under three sensing configurations: single-variable (vibration), two-variable (vibration and current), and three-variable (vibration, current, and temperature). The results indicate that incorporating all sensing variables improves the performance of both SVM and ANN models; however, the proposed SVM-ECOC method demonstrates superior stability and generalization across all scenarios, achieving up to 73.33% accuracy in the three-variable configuration and consistently outperforming the ANN, which is more sensitive to dataset variability. The proposed diagnostic framework is based on a multiclass SVM that uses an Error Correction Code (ECC) scheme. Unlike the standard SVM formulation, commonly applied to binary classification problems, the ECC strategy decomposes the multiclass task into multiple coded binary problems, improving robustness against class overlap and error propagation. This formulation constitutes a key methodological contribution of the proposed system.</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

Failure diagnosis of a single-phase pump used in level control using a multi-sensor SVM with vibration, current, and temperature signals

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Cielo Espinoza Diaz,
  • Javier Aranda Lertora,
  • Carlos H. Inga Espinoza

摘要

The continuous operation of single-phase pumps is essential in industrial environments, yet their reliability is often compromised by mechanical and electrical faults that lead to unplanned downtime, higher maintenance costs, and production losses. This study proposes a multivariate Space Vector Machine with Error-Correcting Output Codes (SVM-ECOC) that integrates vibration, current, and temperature signals to detect and diagnose failures in pumps used for level control in storage tanks. The method addresses common fault conditions, including misalignment, voltage drop, cavitation, level-sensor bias, and proportional-valve bias. The acquired signals are processed through feature extraction in the time, frequency, and time–frequency domains, followed by normalization and dimensionality reduction. Classifiers are trained using a 70–30 partition and evaluated through confusion matrices, accuracy, macro-averaged F1 score, and computational performance metrics on an ESP32 microcontroller. To assess the effectiveness of the proposed approach, its performance was compared with Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) under three sensing configurations: single-variable (vibration), two-variable (vibration and current), and three-variable (vibration, current, and temperature). The results indicate that incorporating all sensing variables improves the performance of both SVM and ANN models; however, the proposed SVM-ECOC method demonstrates superior stability and generalization across all scenarios, achieving up to 73.33% accuracy in the three-variable configuration and consistently outperforming the ANN, which is more sensitive to dataset variability. The proposed diagnostic framework is based on a multiclass SVM that uses an Error Correction Code (ECC) scheme. Unlike the standard SVM formulation, commonly applied to binary classification problems, the ECC strategy decomposes the multiclass task into multiple coded binary problems, improving robustness against class overlap and error propagation. This formulation constitutes a key methodological contribution of the proposed system.