In order to solve the problem of fault diagnosis of transmission error over limit in the use process of industrial robot harmonic reducer, the electromechanical simulation model of servo motor and harmonic reducer was established, and the influence of the increasing of reducer clearance on the current of driving motor was simulated and analyzed. The experimental platform of the electromechanical system of industrial robot harmonic reducer was constructed, and the current signals of drive motor corresponding to harmonic reducer with different transmission errors were collected. Based on the motor current signal, two fault diagnosis methods were proposed, namely Mel frequency cepstral coefficients (MFCC) combined with support vector machine(SVM), and MFCC combined with probabilistic neural network(PNN). Based on a variety of indicators, the above two fault diagnosis methods were comprehensively compared and analyzed. The results show that the comprehensive performance of the MFCC-SVM is better than that of MFCC-PNN.