This paper proposes the order parameter optimization technique of autoregressive minimum entropy
deconvolution(ARMED). Traditionally, the best order is selected by Akaike information criterion(AIC)
which is same as expected squared prediction error (ESPE). However, in practically, there are many
cases that AIC value does not have minimum value and decreases monotonically. This is because
residual signal could be disturbed by the other vibration source or noise even appropriate
preprocessing. Therefore, AIC’s basic assumption could be violated and it does not work well. This
paper proposes new order selection techniques based on health feature related with envelop frequency
using Hilbert transform. The simulation and experimental result show that the selected order using
these techniques make better performance in signal filtered by ARMED for differentiating normal and
fault state than the order obtained by AIC.