Application of Time-series Smoothed Excitation CNN Model

Abstract

The deep learning network simulates the human neural system and its nonlinear hierarchical characteristics, extracts the nonlinear features of the information layer by layer and processes them comprehensively. This is suitable for various evaluation models. The excitation level time-frequency spectrum is used to establish the convolution neural network (CNN) evaluation model. In this paper, the excitation is smoothed in time-domain by using filter first, then the mapping relationship between the global subjective evaluation result and the time sequence smooth excitation is constructed by CNN. The overall comprehensive CNN evaluation model was established based on the time sequence smooth excitation. The time series smoothing excitation CNN model has better performance in the evaluation than the ordinary CNN model and improves the prediction accuracy (the mean error is reduced by 8.64 %), stability (the error variance is reduced by 31.97 %) and consistency (the Pearson correlation coefficient is increased by 2.48 %).