Underwater acoustic target recognition method based on the multi-scale sparse simple recurrent unit model

Abstract

Aiming at the significant decline of target recognition performance in a noisy environment in practical applications, a multi-scale sparse simple recurrent unit (SRU) model is proposed based on a supervised SRU. The model utilizes the internal feedback mechanism of the SRU to model the underwater acoustic target time series (time-domain waveform). Then, it utilizes SRU blocks stacked with different layers to learn the multi-scale feature representations of time series and fuses feature representations. Meanwhile, skip connections are added between the model input and multi-feature layer (feature fusion layer) to accelerate model convergence. A comparative experiment of three types of measured underwater acoustic target radiated noise data shows that compared with the multi-layer classification CNN model, the multi-scale sparse SRU model maintains a higher recognition accuracy when the noise conditions of the training sample and test sample do not match. Therefore, the proposed model is a noise-robust network model.