At present, the health state recognition of industrial robot harmonic reducer is mainly based on vibration signals, which requires additional test system, increases the difficulty and cost of data acquisition, and its accuracy and effectiveness are affected by the installation location of sensors. Based on this, the health state recognition method of harmonic reducer based on depth feature learning of voltage signal is proposed. The industrial robot motor voltage signal is used to characterize the health state of harmonic reducer, and the continuous wavelet transform is used to transform the voltage signal into time-frequency diagram to obtain the time-frequency information of voltage signal under different health state of harmonic reducer, and the data sample set is constructed. The convolutional neural network is used to self-learn the time-frequency information of the voltage signal, and the network parameters are supervised to adjust. In this way, the health state of harmonic reducer can be recognized while the depth characteristics of voltage signal under different health state of harmonic reducer are obtained. Experiment results show that the recognition accuracy of the proposed method reaches 90% above, which proves that the proposed method can effectively recognize the health state of harmonic reducer, and has good generalization ability and robustness.