Based on the decomposition of the space-time eigenstructure of signals

a novel time-frequency signal subspace fitting (TF-SSF) method is proposed in this paper to estimate the DOA of signals.Through the bi-linear Cohen class time-frequency distribution

the time-frequency correlation matrix C

x

is constructed to replace the traditional correlation matrix R

x

. Accordingly the signal subspace and noise subspace are separated with the eigen-decomposition of C

x

. Because the observed data are processed in spatial domain and 2-D time-frequency domain simultaneously

the method can separate the signals that have different time-frequency signatures and is suitable for both stationary and nonstationary signals

while the traditional subspace methods have to assume that the signal is stationar

y.It is also proven herein that the traditional subspace methods is the special case of the TF-SSF method.Furthermore

with the time-varying filtering available

the method has the signal-selectivity and is capable of suppressing interference and noise that are difficult to handle in time or frequency domain only.The simulation results show the effectiveness of the method.