Rotor Fault Diagnosis Using Domain-Adversarial Neural Network with Time-Frequency Analysis

Abstract

Intelligent fault diagnosis of rotors always requires a large amount of labeled samples, but insufficient vibration signals can be obtained in operational rotor systems for detecting the fault modes. To solve this problem, a domain-adaptive transfer learning model based on a small number of samples is proposed. Time-domain vibration signals are collected by overlapping sampling and converted into time-frequency diagrams by using short-time Fourier transform (STFT) and characteristics in the time domain and frequency domain of vibration signals are reserved. The features of source domain and target domain are projected into the same feature space through a domain-adversarial neural network (DANN). This method is verified by a simulated gas generator rotor and experimental rig of rotor. Both the transfer in the identical machine (TIM) and transfer across different machines (TDM) are realized. The results show that this method has high diagnosis accuracy and good robustness for different types of faults. By training a large number of simulation samples and a small number of experimental samples in TDM, high fault diagnosis accuracy is achieved, avoiding collecting a large amount of experimental data as the source domain to train the fault diagnosis model. Then, the problem of insufficient rotor fault samples can be solved.