Scholarlink

<p>Rolling bearing fault diagnosis must jointly identify fault types and quantify severity to enable predictive maintenance. Conventional feature-based pipelines often rely on 15–25 statistical descriptors [29, 30], which increases computational cost and introduces redundancy that is problematic for embedded/edge deployment. This paper proposes MIF8, a compact, fault-mechanism-informed time-domain representation of eight descriptors capturing impulsiveness (envelope- and peak-related measures, clearance factor, zero-crossing rate), periodic impact repetition (autocorrelation), energy/variability (RMS and standard deviation), and signal regularity (predictivity ratio). MIF8 is benchmarked against TFstat18, a widely used 18-dimensional time–frequency statistical set [29], within an SDI–CDA analysis framework that quantifies feature discriminability, constructs a canonical discriminant space, and derives a severity indicator as the distance to the healthy centroid; nonlinear structure is further examined using t-SNE. Experiments on two complementary datasets—URMA (5 classes: healthy, outer-ring, ball, inner-ring, combined defect) and CWRU (4 classes: healthy, ball, inner-ring, outer-ring)—with linear SVM (ECOC), AdaBoostM2, and a FeatCNN-1D under 5-fold cross-validation show that MIF8 achieves accuracy comparable to, and in several cases higher than, the 18-feature baseline while substantially reducing computational requirements (≈ 57.7% fewer operations). SDI, CDA, and t-SNE analyses consistently indicate a more compact and interpretable class structure, and Pareto trade-off evaluation highlights MIF8 as providing the most favorable accuracy–complexity compromise.</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

Fault-mechanism-informed compact features for rolling bearing diagnosis: discriminability, severity, and efficiency via SDI–CDA

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Brahim Oudjani,
  • Djalel Drici,
  • Issam Atoui,
  • Khouloud Bedoud,
  • Hichem Merabet

摘要

Rolling bearing fault diagnosis must jointly identify fault types and quantify severity to enable predictive maintenance. Conventional feature-based pipelines often rely on 15–25 statistical descriptors [29, 30], which increases computational cost and introduces redundancy that is problematic for embedded/edge deployment. This paper proposes MIF8, a compact, fault-mechanism-informed time-domain representation of eight descriptors capturing impulsiveness (envelope- and peak-related measures, clearance factor, zero-crossing rate), periodic impact repetition (autocorrelation), energy/variability (RMS and standard deviation), and signal regularity (predictivity ratio). MIF8 is benchmarked against TFstat18, a widely used 18-dimensional time–frequency statistical set [29], within an SDI–CDA analysis framework that quantifies feature discriminability, constructs a canonical discriminant space, and derives a severity indicator as the distance to the healthy centroid; nonlinear structure is further examined using t-SNE. Experiments on two complementary datasets—URMA (5 classes: healthy, outer-ring, ball, inner-ring, combined defect) and CWRU (4 classes: healthy, ball, inner-ring, outer-ring)—with linear SVM (ECOC), AdaBoostM2, and a FeatCNN-1D under 5-fold cross-validation show that MIF8 achieves accuracy comparable to, and in several cases higher than, the 18-feature baseline while substantially reducing computational requirements (≈ 57.7% fewer operations). SDI, CDA, and t-SNE analyses consistently indicate a more compact and interpretable class structure, and Pareto trade-off evaluation highlights MIF8 as providing the most favorable accuracy–complexity compromise.