The performance of an electroencephalography (EEG) automatic detection and classification system mainly depends on the feature extraction of EEG signal. This paper analyses the advantages and disadvantages of the current EEG feature extraction methods, and then presents a new EEG feature extraction method based on echo state networks (ESN). The new method is a nonlinear method, and can extract the EEG features reversibly. Therefore, the information lost in the process of feature extraction is much less than that of the traditional EEG. Additionally, the realization of this method just needs to compute the pseudo inverse of a matrix, which keeps it efficient. Experimental results have showed that the new method could well accomplish the task of automatic detection and classification of EEG signals.