Ship Noise Target Classification Based on Deep Convolutional Networks

Due to the presence of not only a large amount of natural and anthropogenic noise in the ocean, but also the undulating and variable seabed and inhomogeneous seawater layer, which makes it more difficult to recognize and classify targets in the ocean acoustic field. In recent years, most of the research results in related fields are based on private datasets, but the private datasets are not publicly available and the data volume is small, which hinders the development of the field of underwater acoustic target recognition. The use of deep learning for hydroacoustic signal processing is a current trend, so this paper proposes a method for underwater noise classification using a deep convolutional network model (AlexNet), which will extract three features, namely, the short-time spectrum, the FBank spectrum, and the Mel cepstrum coefficients, of underwater ship noise. The data is obtained using a large open-source hydroacoustic dataset, DeepShip, which consists of 47 hours and 4 minutes of real-world underwater recordings from 265 different vessels in four categories. The classification results of different acoustic features are firstly compared and analyzed based on deep convolutional network type, and the classification results under different features are evaluated by parameters such as accuracy, precision, recall and F1 score. Then the FBank spectra with the best classification results were analyzed, and the effects of different model hyperparameters on classification accuracy were investigated.