Hybrid Diagnostic Framework for Interpretable Bearing Fault Classification Using CNN and Dual-Stage Feature Selection

Timely and accurate fault diagnosis in rotary machinery is essential for ensuring system reliability and minimizing unplanned downtime. While deep learning approaches, particularly Convolutional Neural Networks (CNNs), have demonstrated strong performance in vibration-based fault classification, their limited interpretability poses challenges for adoption in safety-critical environments. To address this, the present study introduces a hybrid diagnostic framework that integrates CNN-based transfer learning with interpretable supervised classification, aiming to enhance both predictive accuracy and model transparency. A key innovation of this work lies in the dual-stage feature selection process, combining Analysis of Variance (ANOVA) and Permutation Feature Importance (PFI) to refine deep features extracted from a pre-trained VGG19 network. This strategy improves both dimensionality reduction and classification performance in a statistically grounded, model-agnostic manner. Furthermore, SHapley Additive exPlanations (SHAP) are employed to interpret the predictions, offering insight into the most influential features driving the classification decisions. Experimental evaluation on the Case Western Reserve University (CWRU) bearing dataset confirms the effectiveness of the proposed approach, achieving 100% classification accuracy using ten-fold cross-validation. By uniting high performance with transparent decision-making, the framework demonstrates strong potential for explainable and reliable fault diagnosis in industrial settings.