Underwater Time-domain Signal Recognition Network with Improved Channel Attention Mechanism

‍ ‍In order to improve the performance of time-domain underwater acoustic signal recognition network， a convolutional neural network for time-domain signal recognition based on improved channel attention mechanism was proposed. The network extracted features from the original signal and the time-domain reconstruction sequence， and randomly discarded the input data points during the training process to prevent overfitting of the network training. Meanwhile， multi-scale convolution modules constructed by multiple convolution layers or residual blocks were used to extract signal features under different frequency components. According to the characteristics of time domain signals， this paper introduced the energy information of sample feature channels， the amplitude information of sample feature channels and the correlation between sample feature channels and the whole sample into the channel attention mechanism to solve the weights of feature channels and enhanced the effective components of features. Finally， the weight regularization term of the classifier was added to the loss function to highlight the effective features extracted by the network. Experimental results on ShipsEar and DeepShip databases showed that the proposed convolutional neural network can effectively identify the target signals in time-domain when the training data and test data have similar distribution.