Smoothed deep neural networks for marine sensor data prediction

Abstract

Deep learning has become a mainstream method in marine data processing field. However, the raw marine data, characterized by fluctuations, outliers and noise, is a serious obstacle to its performance improvement. To address this problem, a hybrid deep computing model is developed for marine sensor data prediction, combining smoothing and deep belief echo state network (DBEN). The proposed structure adopts a two-stage data processing mode to deal with the complex characteristics of marine time series. In the preprocessing stage, four smoothing methods are considered for fluctuation reduction and outlier handling, so that purer data conducive to achieving a higher prediction accuracy can be gained. In the prediction stage, DBEN equipped with efficient feature learning serves as a nonlinear approximator. The effectiveness of the constructed hybrid model is tested on real-world marine time series of different sources and traits, and its superiority in prediction accuracy is demonstrated by readout comparisons and statistical measures.