Towards Fault Diagnosis Interpretability: Gradient Boosting Framework for Vibration-Based Detection of Experimental Gear Failures

Abstract

Accurate and interpretable fault diagnosis in industrial gear systems is essential for ensuring safety, reliability, and predictive maintenance. This study presents an intelligent diagnostic framework utilizing Gradient Boosting (GB) for fault detection in gear systems, applied to the Aalto Gear Fault Dataset, which features a wide range of synthetic and realistic gear failure modes under varied operating conditions. The dataset was preprocessed and analyzed using an ensemble GB classifier, yielding high performance across multiple metrics: accuracy of 96.77%, precision of 95.44%, recall of 97.11%, and an F1-score of 96.22%. To enhance trust in model predictions, the study integrates an Explainable AI (XAI) framework using SHAP (SHapley Additive exPlanations) to visualize feature contributions and support diagnostic transparency. A flowchart-based architecture is proposed to guide real-world deployment of interpretable fault detection pipelines. The results demonstrate the feasibility of combining predictive performance with interpretability, offering a robust approach for condition monitoring in safety-critical systems.