Underwater acoustic target multi-attribute correlation perception method based on deep learning

The underwater acoustic target recognition via passive sonar can extract the features of target attribute from the ship radiated noise to complete the recognition task. The underwater acoustic target is characterized by a variety of attributes and the correlation among them, and the attributes include ship types, ship size, propeller type, etc. The correlation contains the information of the joint distribution of multiple attributes, which is an important complement to target description. Existing underwater acoustic target recognition methods based on deep learning usually classify one attribute of the target without considering the correlation among attributes. In this paper, an underwater acoustic target multi-attribute correlation perception method based on deep learning is proposed. Firstly, deep time–frequency representations are extracted from time-domain ship radiated noise. Then, a group of neurons with learnable wights is designed to extract correlation deep features, which could model the correlation among multiple attributes. Finally, the correlation deep features are utilized to realize multi-attribute correlation perception. The results of visualization experiment show that the learned correlation is similar with the true correlation among the multiple attributes. And the results of prediction experiment show, by adding information of correlation among multi attributes, the proposed method achieves better recognition performance (accuracy is 82.1%). And in ship monitoring experiment, the proposed method achieves the furthest recognition distance and the most stable correct recognition.