For structure dynamic signals with low signal to noise ratio (SNR), Multi-scale EMD is proposed to extract useful information from signals contaminated by noise. This method is an adaptive way, and achieves better performance than traditional methods. A complex signal composed of signal and noise, which is first decompose into a series of different characteristic IMFs. Then sum of the first few IMFs including useful information and noise are further decomposed by EMD, and iterate this process for several times in order to extract useful information. The final step is deleting some IMFs obtained from the last decomposing which is considered as noise, and rebuilt the signal using the rest IMFs from different scale. This method can smooth signal effectively. The effectiveness of the proposed method is verified with the numerical simulating signal, and useful information in signals is retained. It also has some superior performances such as concise computer process, robust application, etc.