Data-driven adaptive chirp mode decomposition with application to machine fault diagnosis under non-stationary conditions

Abstract

Rotating machineries play a significant role in industrial application and fault diagnosis is an important technology to ensure their safe operation. However, the complicated operating environment makes the condition monitoring signals usually display nonlinear and non-stationary characteristics, which brings severe challenges to fault diagnosis. Although adaptive chirp mode decomposition (ACMD) shows good adaptability and high time-frequency resolution for non-stationary signals, it depends on an instantaneous frequency (IF) initialization based on Hilbert transform, which limits its practical applications. In this paper, a fully data-driven adaptive chirp mode decomposition (DD-ACMD) is proposed to address the issue. Firstly, the high-frequency modes of the signal are enhanced by derivative operation, and then the IF of the highest-frequency mode is preliminarily estimated based on a normalization operator. Next, an iterative time-varying filtering method based on a demodulation technique is proposed to reduce the influence of noise and thus obtain good estimates of initial IFs for the ACMD. In addition, a time-varying low-pass filter is introduced into the recursive framework of mode extraction to further improve the noise robustness of the whole algorithm. The DD-ACMD has both high adaptability and good noise robustness, and can even separate non-stationary signals with very close modes. The effectiveness of the DD-ACMD is validated by both simulations and real-life applications to machine fault diagnosis.