<p>Prognostics and health management (PHM) has become essential in modern industry, particularly for systems such as harmonic drives (HDs), also known as harmonic gears or gearboxes, which are specialized mechanical gearing systems used in industrial robots. The HD reducer, susceptible to various faults owing to continuous operation, requires timely PHM to maintain smooth and steady operations. Traditional methods for diagnosing HD reducer faults, such as demography analysis, calibration, and acoustic emission analysis, encounter various challenges. This study introduces a novel approach utilizing wireless data collection via an acceleration sensor for fault diagnosis, uniquely focusing on load and velocities, an area not explored in existing studies. By applying machine learning (ML) algorithms to the collected data, prominent features are identified from acceleration signals. Subsequently, the dataset is trained using various ML classification algorithms. The original features with kinetic energy-based signals perform better than other feature selection techniques. Specifically, the random forest classifiers achieve the highest accuracy of approximately 99.81%, significantly outperforming other classification algorithms such as k-nearest neighbors, decision tree, and XGBoost. Additionally, this study employed explainability tools, including Shapley Additive Explanations (SHAP) and local interpretable model-agnostic explanations (LIME), to enhance the interpretability of the diagnostic process.</p>