Focused on the inability of traditional multiscale time series reconstruction method to extract EEG signal characteristics completely, a new multiscale reconstruction method of time series, moving variances was proposed. EEG signals are reconstituted by new method of moving variances, then two eigenvalues, sample entropy and variance entropy, are extracted at multiple scales. Finally, the extracted features are tested of p-value by method of KS test. Experiments show extracting EEG signal features from time series at multiple scales which are reconstituted by the method of moving variances can effectively distinguish EEG signals of patients with epilepsy during period and interval of onset and thus provide basis to diagnose mental disorders by EEG signals.