Targeting at improving the accuracy of multi-classification problems of epileptic EEG signals, one algorithm based on the combination of a signal to difference module and convolutional module is proposed. The signal to difference module performs multi-order differential operations on raw EEG signals to obtain its incremental representation which depicts the fluctuation features of EEG signals. Then, this representation is converted to images by convolutional module using dynamic learning parameters rather than static transformation. And pre-trained convolutional networks are applied to extract features and classify them automatically. The classification results show that this method improves the classification performance by up to 8.1% when compared to recent researches. This method achieved 99.8% accuracy in two-class classification problems, 92.8% accuracy in three-class classification problems and 86.7% accuracy in five-class classification problems, which indicates that signal to difference module has an important effect on EEG classification problem.