Scholarlink

<p>Reliable condition monitoring of rolling element bearings is essential for improving machinery availability and reducing unplanned downtime. Machine learning (ML) techniques have been widely applied for vibration-based bearing fault diagnosis; however, reported performance is often difficult to interpret due to variations in datasets, operating conditions, and validation protocols. This study presents a controlled experimental benchmark evaluating three classical ML algorithms—Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Artificial Neural Network (ANN)—for vibration-based fault classification of cylindrical roller bearings. Experiments were conducted on SKF N204 bearings using a machinery fault simulator under three bearing conditions (healthy, outer race defect, and roller defect) at four rotational speeds (100–400 RPM). Artificial defects were introduced using electrical discharge machining (EDM), and vibration signals were acquired using a National Instruments data acquisition system. Root mean square (RMS), kurtosis, and crest factor were extracted as interpretable statistical features, resulting in a dataset of 119 feature samples. Model performance was evaluated using stratified train–validation–test splits, with k-fold cross-validation applied uniformly across all classifiers, including the ANN, to ensure fair comparison. Under the present laboratory conditions, the ANN exhibited the most consistent performance across validation and testing, while SVM and k-NN showed slightly lower validation accuracy but comparable testing behavior. The results demonstrate that high classification accuracy can be achieved using simple statistical features and classical ML models in a controlled setting. However, the findings should be interpreted as upper-bound diagnostic performance under laboratory conditions, rather than direct indicators of industrial-scale generalization. The study provides a transparent benchmarking reference and highlights the trade-off between interpretability, data efficiency, and classification performance in bearing health monitoring applications.</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

Machine Learning for Bearing Health Monitoring: A Controlled Experimental Benchmark Using Vibration-Based Statistical Features

  • 期刊论文

  • 发表日期：

    2026年4月30日

查看全文
  • Prasanta Kumar Samal,
  • Srinidhi R,
  • Imran M. Jamadar,
  • Razikha Amreen M. I.

摘要

Reliable condition monitoring of rolling element bearings is essential for improving machinery availability and reducing unplanned downtime. Machine learning (ML) techniques have been widely applied for vibration-based bearing fault diagnosis; however, reported performance is often difficult to interpret due to variations in datasets, operating conditions, and validation protocols. This study presents a controlled experimental benchmark evaluating three classical ML algorithms—Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Artificial Neural Network (ANN)—for vibration-based fault classification of cylindrical roller bearings. Experiments were conducted on SKF N204 bearings using a machinery fault simulator under three bearing conditions (healthy, outer race defect, and roller defect) at four rotational speeds (100–400 RPM). Artificial defects were introduced using electrical discharge machining (EDM), and vibration signals were acquired using a National Instruments data acquisition system. Root mean square (RMS), kurtosis, and crest factor were extracted as interpretable statistical features, resulting in a dataset of 119 feature samples. Model performance was evaluated using stratified train–validation–test splits, with k-fold cross-validation applied uniformly across all classifiers, including the ANN, to ensure fair comparison. Under the present laboratory conditions, the ANN exhibited the most consistent performance across validation and testing, while SVM and k-NN showed slightly lower validation accuracy but comparable testing behavior. The results demonstrate that high classification accuracy can be achieved using simple statistical features and classical ML models in a controlled setting. However, the findings should be interpreted as upper-bound diagnostic performance under laboratory conditions, rather than direct indicators of industrial-scale generalization. The study provides a transparent benchmarking reference and highlights the trade-off between interpretability, data efficiency, and classification performance in bearing health monitoring applications.