Modulation Recognition of Underwater Acoustic Communication Signals Based on Denoting & Deep Sparse Autoencoder

Abstract. It is difficult to identify the modulation types of underwater acoustic communication signals via traditional modulation recognition methods since the received communication signals always have a low signal to noise ratio due to the oceanic ambient noise and signal distortion. Deep denoising autoencoder is a special type of deep neural network which can learn essential representations of input that are robust to stochastic corruption in input, besides, sparse coding can make the representations learning more efficient efficiently. We propose a denoising and modulation recognition model for underwater acoustic communication signals which is based on deep denoising autoencoder. We use a denoising autoencoder to reconstruct the signals as a preprocessing in order to enhance the signals and use greedy layer-wise training to train a stacked denoising autoencoder, then we use the reconstructed signals as a set of input to train a deep neural network which is initialized by the stacked denoising autoencoder we have trained above. In this process, sparseness constraints are used to improve the performance of the model. Comparative experiments clearly show the advantage of using denoising autoencoder in modulation recognition of underwater acoustic communication signals.Keywords: underwater acoustics, communication signal, modulation recognition, deep denoising autoencoder