A JOINT TIME–FREQUENCY EMPIRICAL MODE DECOMPOSITION FOR NONSTATIONARY SIGNAL SEPARATION N. Stevenson, M. Mesbah, B. Boashash Perinatal Research Centre University of Queensland Herston, QLD, 4029, Australia email: n.stevenson@uq.edu.au H.J. Whitehouse Linear Measurements, Inc. 4174 Sorrento Valley Blvd, San Diego, CA 92121 ABSTRACT This paper outlines the application of the empirical mode de- composition (EMD) to a frequency domain representation of a signal. The application to frequency domain representa- tions, as opposed to time domain representations, may pro- vide more useful decomposition in the case where signal com- ponents overlap in frequency. The combined use of the EMD for both time and frequency domain representations of a sig- nal is capable of generating more desirable signal decompo- sitions for nonstationary signals. The results of the time– frequency EMD on two example signals shows improvement in nonstationary signal separation with little loss of desirable decomposition performance measures such as IMF orthogo- nality. 1. INTRODUCTION The empirical mode decomposition (EMD) has been proposed by Huang et al., in [1], as a tool for analysing nonlinear and nonstationary time series. The EMD decomposes a signal into a series of intrinsic mode functions (IMFs). These IMFs are optimised for instantaneous frequency (IF) estimation using the analytic associate of a signal. To this end, the EMD uses an iterative estimate of the local or short–time mean, based on the signal envelope, to generate IMFs that have an equal number of extrema and zero–crossings. The use of the local mean has the implication that the EMD is an adaptive decomposition technique, as it decom- poses a signal based on a set of signal dependent basis func- tions. This is contrary to well known signal decomposition methods, such as Fourier, wavelet or Wigner–Ville. It is the behaviour of the EMD when decomposing signals of a non- stationary nature that is of interest to the time–frequency (TF) signal processor. Of particular interest is the representation of the signal using the EMD as s(t)= Q i=1 IMF i (t)+ r(t), (1) where, Q is the number of IMFs and r(t) is the decomposition residue. This representation is identical to that given in [2] to define signal components in a TF signal processing context, s(t)= Q i=1 a i (t) cos(θ i (t)) + r(t), (2) where a i (t) cos(θ i (t)) = IMF i (t), a i (t) is the envelope and θ i (t) is the phase of the i th component or IMF. According to this decomposition, an IMF is equivalent to a signal compo- nent. In the context of the EMD, this representation suggests that the EMD is capable of separating signals with different frequencies at a single time instant, but it cannot separate sig- nals with identical frequency content at different times. The representation in (2) defines each component as ex- isting for the entire signal duration. In practice, however, a component can additionally be defined with limits in the time domain. This is a more complete definition of nonstationary signal components and can be outlined as, s(t)= L k=1 a k (t)rect t τ k b k cos(θ k (t)) + r(t), (3) where rect((t τ k )/b k ) is the rectangular function of width b k , centre τ k and L Q. The advantage of such a definition is that a signal component can be localised in both the time (us- ing cos(θ k (t))) and frequency (using rect((t τ k )/b k )) axes of the TF plane. This representation also results in monotonic components with respect to time and frequency. This has applications in kernel design for reduced interference time– frequency distributions (TFDs), [3, pp. 168]. The aim of this paper is to adjust the EMD so that it pro- vides IMFs that fit the definition of (3) rather than (2). That is, force the EMD to separate signals with identical frequency content at different times, in addition to separating signals with different frequencies at a single time instant. The pro- posed solution involves using the discrete cosine transform (DCT), [4], as an additional processing step in the EMD. 1-4244-0779-6/07/$20.00 ©2007 IEEE