Interpretable Intelligent Diagnosis Method for Aero-engines Based on Deep Signal Separation

To address the problem that existing harmonic extraction methods such as time-frequency analysis and signal decomposition cannot adaptively extract the rotational speed harmonic components with clear physical meaning when performing tacho-less order tracking(TLOT) applications for rotating machineries, an interpretable rotational speed harmonic signal adaptive separation method based on a time-frequency spatial deep binary mask model is proposed. Firstly, a series of simulated template signals with clear physical meaning of the instantaneous frequency(IF) change of the tacho harmonics are constructed as training samples for the deep binary mask model, and then the time-frequency representation(TFR) of the tacho fundamental harmonics and higher-order harmonics of the constructed simulated template signals are used to generate the training target of the binary mask model in the time-frequency(TF) space of the fundamental harmonics. The non-linear mapping relationship between the TF space of the simulated template signal and the TF image of the tacho fundamental harmonics is established to achieve the accurate separation and extraction of the tacho fundamental harmonic components and the credible masking suppression of the tacho higher-order harmonic components, ensuring that the harmonic components obtained by the deep binary mask model have a clear physical meaning. Then, the instantaneous phase information related to the deep nonlinear mapping of the rotational speed signal is obtained by the Hilbert transform of the fundamental harmonics obtained by deep separation, and applied to the original signal with equal angle resampling to accomplish accurate order tracking. In summary, the proposed method overcomes the problem that the traditional TLOT method relies heavily on expert experience, which can provide instantaneous phase support with clear physical meaning for TLOT of the transmission system. © 2024 Chinese Mechanical Engineering Society. All rights reserved.