Abstract: In the recognition problem of time series signals such as seismic,acoustic,and EEG,ECG signal,it is necessary to extract features from time series.In this paper,a new method for feature extraction of time series is proposed,which decomposes time series into two parts and extracts features respectively.Its implementation is by AR model and Wiener filter,AR coefficients and Wiener filter coefficientsare combined into feature vector.As an example of application,this method is usedto extract features from seismic signal,in order to recognize sedimentary sequenceand then help the inference of sedimentary faces.The computer experimental result shave shown that the method is quite effective.