An improved time domain polyreference method for global model identification is presented in this paper. A time domain preprocessing technique is developed to reduce the effects of random noise contamination on measured data. Compared with the polyreference complex exponential technique, the size of the estimation problem is considerably reduced and the judgement of the required computational order is made easier and more reliable in the low signal-to-noise ratio cases. A total least squares algorithm with singular value decomposition for parameter estimation is adopted to minimise the bias error. An improved procedure for residue calculation is proposed, which takes residual terms into account in the time domain.