Harmonic reducer in-situ fault diagnosis for industrial robots based on deep learning

The harmonic reducer is an essential kinetic transmission component in the industrial robots. It is easy to be fatigued and resulted in physical malfunction after a long period of operation. Therefore, an accurate in-situ fault diagnosis for the harmonic reducers in an industrial robot is especially important. This paper proposes a fault diagnosis method based on deep learning for the harmonic reducer of industrial robots via consecutive time-domain vibration signals. Considering the sampling signals from industrial robots are long, narrow, and channel-independent, this method combined a 1-dimensional convolutional neural network with matrix kernels (1-D MCNN) adaptive model. By adjusting the size of the convolution kernels, it can concentrate on the contextual feature extraction of consecutive time-domain data while retaining the ability to process the multi-channel fusion data. The proposed method is examined on a physical industrial robot platform, which has achieved a prediction accuracy of 99%. Its performance is appeared to be superior in comparison to the traditional 2-dimensional CNN, deep sparse automatic encoding network (DSAE), multilayer perceptual network (MLP), and support vector machine (SVM).