An analytical benchmark of feature selection techniques for industrial fault classification leveraging time-domain features

Abstract

The growing size and complexity of industrial datasets have intensified the need for efficient fault diagnostics tools. This study addresses the challenge of handling large-scale data by developing a data-driven architecture for fault classification in industrial systems. To extract meaningful insights, 15 time-domain features were combined with 5 Feature Selection Methods to optimize model performance by eliminating redundant features. The sensor data and selected features were analyzed using the Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) algorithms to enable accurate fault detection and prediction. The proposed framework was validated using publicly available datasets, namely the Case Western Reserve University (CWRU) bearing dataset and the National Aeronautics and Space Administration Ames Prognostics Center of Excellence (NASA PCoE) lithium-ion battery dataset. The results demonstrate the framework's adaptability and high efficacy across diverse scenarios, achieving an average F1-score exceeding 98.40% using only 10 selected features. This finding highlights the effectiveness of embedded Feature Selection Methods in improving classification performance while reducing computational complexity. The study underscores the potential of the proposed framework as a foundational tool in intelligent manufacturing, offering a versatile solution to enhance fault diagnostics in diverse industrial applications.