The Performance Prediction of HD Gear Reducer in Industrial Robots using Machine Learning Approach

Abstract

Harmonic drive reducers are essential components of industrial robot arms, primarily due to their high reduction ratios, concentric shafts, compact size, and low backlash, making them particularly suitable for small to medium load industrial robots. However, their intricate mechanical structure, coupled with low vibration characteristics, poses challenges in determining performance attributes. One common failure mode in harmonic drives is flex spline fatigue fracture, which requires specialized testing platforms and removal of harmonic drives for measuring torsional stiffness. To address these challenges, this study introduces a machine learning-based vibration analysis approach for the online prediction of performance attributes in harmonic drive reducers. An accelerated testbed is employed to simulate loading and operational conditions of harmonic drives in real robots. This testbed collects vibration data during operation and measures the torsional stiffness and transmission error of the harmonic drive. The collected vibration data is then processed using discrete wavelet transformation, followed by feature extraction from the transformed signals. These extracted features are used to train an artificial neural network designed to simultaneously predict the torsional stiffness coefficient, average transmission error, and the maximum transmission error. The mean squared error and mean absolute error of the proposed method are significantly lower, demonstrating the superiority of the proposed machine learning approach in addressing the challenges associated with evaluating and predicting the performance of harmonic drive reducers in industrial robots.