Fault Diagnosis of Harmonic Reducers Used in Industrial Collaborative Robots via DWT, MB-FFT, EMD and Machine Learning Classifiers (RF, XGBoost, SVM and KNN)

Abstract

Harmonic reducers are critical components in industrial collaborative robot joints but are prone to faults because they operate under cyclic motion and fluctuating load conditions. This work focuses on three representative failure modes in harmonic reducers: gear tooth of the flexspline breakage, flexible bearing outer race defects, and wear at the flexspline–circular spline interface. To enhance weak fault signatures, three signal preprocessing schemes are evaluated: Discrete Wavelet Transform (DWT), Multiband Fast Fourier Transform (MB-FFT), and Empirical Mode Decomposition (EMD), followed by the extraction of 11 time-frequency domain features from each processed signal set. The resulting features are used to train four classical Machine Learning (ML) classifiers, Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM) and K-Nearest Neighbors (K-NN).  Experimental results show that the proposed MB-FFT technique provides the lowest computational cost while delivering superior classification performance, achieving 100% accuracy when combined with random forest and XGBoost for both vibration and current signals. Compared with deep learning models, the results demonstrate that signal enhancement can significantly improve classification performance despite weak fault characteristics, and that current signals can serve as effective indicators for harmonic drive fault diagnosis in cobots.