Scholarlink

<p>This study presents a multilayer perceptron classifier for fault diagnosis in rotating machinery using statistical features derived from vibration signals. Eighteen time-domain parameters were initially extracted from the underhang bearing radial acceleration signal. Correlation analysis, IG, F-scores, p-values, and SHAP values were applied for feature selection. Four key features were retained for training. The model was trained and evaluated on non-overlapping 1&#xa0;s signal segments, which provided a balance between temporal resolution and dataset size. Additional tests using 2.5 s segments slightly improved the macro-average F1-score from 0.97 to 0.98 but reduced the sample count, confirming that 1&#xa0;s segmentation offers a robust and scalable baseline, while longer segments enhance separability of overlapping fault classes. The optimized model was validated on the CWRU bearing dataset, achieving a macro-average F1-score of 0.98 and demonstrating strong generalization across fault types and operating conditions. Comparative experiments with baseline models showed that the proposed model achieved the best balance between accuracy and computational efficiency. Computational performance analysis revealed that it delivered the fastest inference time with moderate memory usage. Hyperparameter sensitivity analysis identified learning rate and optimizer type as the most influential factors. SHAP analysis highlighted RMS and Skewness as the most significant diagnostic features, while Entropy and Kurtosis refined understanding of signal complexity and defect localization. Requiring only a single accelerometer and a minimal feature set, the model provides a compact, interpretable, and efficient solution for real-time predictive maintenance in industrial 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

Efficient Fault Diagnosis in Rotating Machinery via Feature-Optimized Multilayer Perceptron

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • A. Khanfir,
  • A. Bouguecha,
  • M. Haddar

摘要

This study presents a multilayer perceptron classifier for fault diagnosis in rotating machinery using statistical features derived from vibration signals. Eighteen time-domain parameters were initially extracted from the underhang bearing radial acceleration signal. Correlation analysis, IG, F-scores, p-values, and SHAP values were applied for feature selection. Four key features were retained for training. The model was trained and evaluated on non-overlapping 1 s signal segments, which provided a balance between temporal resolution and dataset size. Additional tests using 2.5 s segments slightly improved the macro-average F1-score from 0.97 to 0.98 but reduced the sample count, confirming that 1 s segmentation offers a robust and scalable baseline, while longer segments enhance separability of overlapping fault classes. The optimized model was validated on the CWRU bearing dataset, achieving a macro-average F1-score of 0.98 and demonstrating strong generalization across fault types and operating conditions. Comparative experiments with baseline models showed that the proposed model achieved the best balance between accuracy and computational efficiency. Computational performance analysis revealed that it delivered the fastest inference time with moderate memory usage. Hyperparameter sensitivity analysis identified learning rate and optimizer type as the most influential factors. SHAP analysis highlighted RMS and Skewness as the most significant diagnostic features, while Entropy and Kurtosis refined understanding of signal complexity and defect localization. Requiring only a single accelerometer and a minimal feature set, the model provides a compact, interpretable, and efficient solution for real-time predictive maintenance in industrial applications.