Time-domain methods for the identification of linear structural dynami
c systems are studied. The stochastic autoregressive and moving averag
e (ARMAX) model is used to process the measured excitation and respons
e records contaminated by noises. The study focuses on the sequential
prediction-error method incorporating several techniques for improving
the parameter estimation. They are the exponential data weighting, th
e global data weighting, and the square-root estimation techniques. Ef
ficient procedures of the square-root estimation are developed for the
multi-input and multioutput (MIMO) case as well as the multi-input an
d single-output (MISO) case. Verifications of the present methods are
carried out using the simulated time histories for the input excitatio
n and output response, as well as using the experimental data on a bui
lding model. The results indicate that the square-root estimation tech
nique is particularly effective for improving the convergence and accu
racy of the sequential estimation, even with crude initial guesses.