Deep residual learning with demodulated time-frequency features for fault diagnosis of planetary gearbox under nonstationary running conditions

Due to the tough and time-varying working conditions, fault diagnosis technique is of critical significance for drive-chain system in rotating machines. In recent years, many statistical and spectral feature extraction methods have been developed and applied, but unfortunately, they are incapable of dealing with mechanical behaviors under varying running conditions. Besides, the lack of specific dynamical knowledge also becomes an obstacle for effective diagnosis through direct spectral analysis. Accordingly, a data-driven fault diagnosis method based on time-frequency analysis and deep residual network is proposed in this research. Firstly, a deep residual network is pre-trained on spectral features extracted under fixed rotating speeds. For the transient signals, an accurate phase function is constructed via probabilistic instantaneous angular speed (IAS) estimation algorithm based on time-frequency representations. Then the generalized demodulation operator is utilized to remove rotating speed fluctuation. Afterwards, several groups of instantaneous features demodulated from time-frequency representations are input to the deep residual network to test the performance of proposed method under nonstationary running conditions. The diagnosis results of a planetary gearbox test rig are compared with other traditional methods; the comparisons show that the proposed data-driven fault diagnosis method achieved significant improvement on incipient fault detection accuracy under varying rotating speed. (C) 2019 Elsevier Ltd. All rights reserved.