Abstract:Due to the cyclic motion, different working beats and transient speed, it is difficult to effectively describe the running state and assess the health state of industrial robot harmonic reducer. In this study, a method based on integerperiod data and convolutional neural network is proposed to achieve health state assessment of harmonic reducer. Firstly, the phase difference spectrum correctioncross correlation method is used to adaptively segment the original vibration signal and construct the integerperiod data samples to accurately describe the running state information of the harmonic reducer. Secondly, continuous wavelet transform is applied to decompose the integerperiod data sample to obtain the timefrequency map to fully show the transient characteristics of harmonic reducer in the operation cycle. Finally, convolution neural network is utilized to translate and scale the input signals in time and space with high invariability to fully learn the transient characteristics of the harmonic reducer in each operating cycle. In this way, the health state of the harmonic reducer can be evaluated. Experimental results show that the identification accuracy of the proposed method is over 90%. The effectiveness of the proposed method is verified, which has good generalization ability and robustness.