Underwater Acoustic Preamble Detection via End-to-End Complex-Valued Synchrosqueezed Wavelet Neural Network

Abstract

Preamble detection is critical in underwater acoustic systems due to its impact on reliability and operational coexistence. Traditional methods are limited due to the types of interference found in underwater environments, which can easily falsely trigger the system. In this study, we propose an end-to-end neural network for preamble detection, using a single deep learning model without preprocessing. Our approach employs a simple convolutional neural network architecture with a minimal number and size of layers. We integrate neural network with time–frequency analysis knowledge via the complex-valued wavelet synchrosqueezing layer to extract crucial time–frequency features, which is essential for distinguishing the preamble from underwater acoustic interferences. In addition, we adapt the network to handle complex values, capturing both magnitude and phase information in preamble signals. Experimental results demonstrate that, even with similar preamble interferences, our proposed network, leveraging the Morlet mother wavelet under the LeNet1d framework, exhibits superior detection performance compared to conventional networks. Notably, the performance is very robust even with a small training data set and small computational complexity, highlighting the effectiveness of the network's knowledge-based design.