Misalignment Severity Classification in Rotor-Bearings using the mRMR Feature Selection Method
پذیرفته شده برای ارائه شفاهی
کد مقاله : 1051-ISAV2025
نویسندگان
1دانشگاه صنعتی امیرکبیر، دانشکده مهندسی مکانیک، آزمایشگاه تحقیقاتی آکوستیک
2دانشگاه صنعتی امیرکبیر، دانشکده مهندسی پزشکی
چکیده
The accurate detection and diagnosis of mechanical faults are essential for predictive maintenance and operational safety. A crucial step in this process involves extracting informative features and selecting the most effective subset for classification. This study presents a comprehensive evaluation of nine entropy-based measures and nine time-domain statistical features for fault classification using an optimized Support Vector Machine (SVM), where the optimization is performed via a Genetic Algorithm (GA). The analysis is conducted on an acoustic dataset containing normal conditions and four distinct fault severities. Results show that while individual entropy measures, such as Bubble Entropy, achieve high fault detection accuracy (95.75%), their performance in multi-class fault diagnosis is limited (47.70%). In contrast, a combined set of time-domain features provides a stronger baseline, achieving 95.03% detection accuracy and 85.55% diagnosis accuracy. To address the limitations of individual features, a hybrid framework integrating Minimum Redundancy Maximum Relevance (mRMR) for feature selection is proposed. The resulting mRMR-GA-SVM model demonstrates superior performance, achieving near-perfect fault detection (99.55%) and excellent fault diagnosis accuracy (95.21%). Overall, the findings confirm that a strategically selected hybrid feature set using the mRMR approach significantly outperforms any single feature type, establishing a robust and reliable methodology for complex fault diagnosis tasks.
کلیدواژه ها
Title
Misalignment Severity Classification in Rotor-Bearings using the mRMR Feature Selection Method
Authors
Abdolreza Ohadi, Farshad Alamsganj
Abstract
The accurate detection and diagnosis of mechanical faults are essential for predictive maintenance and operational safety. A crucial step in this process involves extracting informative features and selecting the most effective subset for classification. This study presents a comprehensive evaluation of nine entropy-based measures and nine time-domain statistical features for fault classification using an optimized Support Vector Machine (SVM), where the optimization is performed via a Genetic Algorithm (GA). The analysis is conducted on an acoustic dataset containing normal conditions and four distinct fault severities. Results show that while individual entropy measures, such as Bubble Entropy, achieve high fault detection accuracy (95.75%), their performance in multi-class fault diagnosis is limited (47.70%). In contrast, a combined set of time-domain features provides a stronger baseline, achieving 95.03% detection accuracy and 85.55% diagnosis accuracy. To address the limitations of individual features, a hybrid framework integrating Minimum Redundancy Maximum Relevance (mRMR) for feature selection is proposed. The resulting mRMR-GA-SVM model demonstrates superior performance, achieving near-perfect fault detection (99.55%) and excellent fault diagnosis accuracy (95.21%). Overall, the findings confirm that a strategically selected hybrid feature set using the mRMR approach significantly outperforms any single feature type, establishing a robust and reliable methodology for complex fault diagnosis tasks.
Keywords
Fault diagnosis, Entropy Measures, feature selection, mRMR Feature Selection