Preprocessing based solution for the vanishing gradient problem in recurrent neural networks