Underwater acoustic target recognition based on knowledge distillation under working conditions mismatching

In marine economics and military activities, the timely recognition of underwater acoustic targets in complex environments is particularly important. The conditions of ship navigation can greatly affect the radiation noise. However, due to the difficulties of marine experiments, the collected data only contains a limited part of the working conditions, which makes underwater acoustic target recognition difficult. To solve the problem of working conditions mismatching, we proposed a deep neural network based on knowledge distillation. In the proposed method, the teacher network uses sufficient straight working condition data to obtain prior knowledge. Then, while training with a small amount of target working data, the student network patiently learns the classification knowledge from the feature-based knowledge and the response-based knowledge of the teacher network by using the knowledge distillation method to extract incremental knowledge. Knowledge distillation is used to improve the accuracy of the classification of target working condition data by student network. The effectiveness of the method was verified on multi-working conditions ship-radiated noise datasets. The experimental results show that the proposed method can improve the performance of underwater acoustic target recognition under working conditions mismatching.