Some methods of pre-processing input data for neural networks

Abstract

Two techniques of data pre-processing for neural networks are considered in this paper: (i) data compression with the application of the principal component analysis method, and (ii) various forms of data scaling. The novelty of this paper is associated with compressed input data scaling by the rotation (by the "stretching") in neural network. This approach can be treated as the new proposition for data pre-processing techniques. The influence of various types of input data pre-processing on the accuracy of neural network results is discussed by using numerical examples for the cases of natural frequency predictions of horizontal vibrations of load-bearing walls. It is concluded that a significant reduction in the neural network prediction errors is possible by conducting the appropriate input data transformation.