Parameter Estimation of Ocean Impulsive Noise Using Hybrid Deep Neural Networks

Abstract

Impulsive noise exists widely in ocean environment and statistical parameters of impulsive noise are necessary for ocean signal processing and communication system. However, statistical parameters are generally assumed known or acquired through complicated calculations. In this letter, a hybrid deep neural network based parameter estimation is proposed for ocean impulsive noise. The proposed method formulates the parameter estimation as a non-linear mapping problem to be solved by deep learning network. The network incorporates one dimensional convolutional neural network to extend noise signals into new feature space without destroying large-amplitude characteristics. Then the refined features are input to stacked long-short term memory modules for temporal feature exploration considering temporal correlations inherently in sequential noise signals and the parameters are output through fully connected layers. The proposed network is verified and analysed through impulsive noise datasets from acknowledged ocean impulsive models and real-test ocean noise. Experimental results prove the advantages of proposed method in parameter estimation accuracy.