In this paper, the problem of identification of discretetime system with a timedelay is addressed. This problem involves both the estimation of the timedelay and the estimation of the dynamic parameters. An Improved Stochastic Gradient Descent Method based on auxiliary model theory and lowpass filtering technique with a variable forgetting factor is proposed to simultaneously estimate the timedelay and dynamic parameters on line. An auxiliary model is established to estimate the noisefree output of system. A variable forgetting factor based on prediction error is introduced to enhance the convergence rate and identification accuracy. Moreover, a lowpass filtering technique is used to widen the convergence region and the sharp peaks of cost function are ‘spread out’, which makes the global minimum easier to reach. Simulation results are included to show the effectiveness of the proposed method.