Application of a coarse-to-fine minimum entropy deconvolution method for rotating machines fault detection

Abstract

Due to the severe working condition and long-term service, the key rotating parts including the bearing and gearbox, are susceptible to damage. Blind deconvolution which can eliminate the influence of the transfer path and enhance the fault-related feature is widely used for machinery fault diagnosis. Not only the objective function but the initial filter can influence the convergence properties of the deconvolution algorithm. For example, in practical application, the classical MED trends to converge to fault-unrelated outliers if the initial filter coefficients are not appropriate. Motivated by this, the effect of initialization on the convergence of MED is firstly investigated. Subsequently, a coarse-to-fine MED (CFMED) is designed to highlight the fault-induced repetitive impulses. Specifically, the procedure of pre-iterating is used to search for the convergence direction and coarsely update the initial filter. Then, the dropout step is introduced to drop out the redundant candidates and reserve the expected solution by quantitatively measuring the sparsity of all possible solutions in the frequency domain. Meanwhile, CFMED is more robust to filter size than MED. Finally, the simulation and experimental data with bearing and planetary gearbox fault verify CFMED is more suitable for the fault diagnosis of rotating machines compared with original MED.