The weakly damped low frequency oscillation is one of the main factors that influence the stabile operation of interconnected power grids. At present based on the measured data， the alarm signals can only be generated after the oscillation. It has been observed from the wide area measured data that small amplitude fluctuations which are similar to ambient signals exist continuously in power grids because of the random changes of loads. In order to realize the early warning under normal operation， the autoregressive moving average（ARMA） method is used to estimate the low frequency oscillation modes based on ambient signals. As a key step of the ARMA method， the selection of model order is studied in detail. By comparing different model order selection criterions， Bayesian information criterion（BIC） is chosen to determine the model order， and for the online application， ARMA（2n，2n -1） modeling procedure is adopted to improve the calculation efficiency. Finally， the approach is employed to process the simulation data from the 36-node benchmark system and practical ambient signals measured in Southern China power grid， and the results validate the feasibility of this approach.