Intelligent fault diagnosis of planetary gearboxes under time-varying condition based on bilateral adversarial encoder

Characteristic frequencies of planetary gearboxes are time-varying when they are running under time-varying condition, and conventional statistics are unsuitable to explore time variability and potential properties of nonstationary signals. These make fault diagnosis of a planetary gearbox manually difficult. In order to address these issues, a bilateral adversarial encoder model was proposed. Firstly, a time-frequency representation image of a sample was obtained to reveal the time-varying frequencies. Then, an encoder and a decoder were established. The input of the encoder and the output of the decoder were utilized to train discriminator 1 to ensure that the reconstructed signal follows the distribution of original signal, and the features of an image are effectively extracted. Besides, a Gaussian mixture distribution was established and samples were collected from the Gaussian mixture distribution according to the label information. Discriminator 2 is utilized to make the extracted features follow the distribution of the samples in order to enhance the discriminability of features among different classes by controlling the mixture distribution. Finally, a Softmax classifier was trained by the enhanced features and the testing features were identified. This method was validated via planetary gearbox data set. The results indicate that the reconstructed signal can be made by an adversarial game to follow the distribution of the original signal and the features can be controlled by a Gaussian mixture distribution to improve their clustering performance and diagnose the gear faults accurately. In comparison with other methods, it works better to some degree. © 2021, Editorial Office of Journal of Vibration and Shock. All right reserved.