A Deep Learning Model for Detecting the Arrival Time of Weak Underwater Signals in Fluvial Acoustic Tomography Systems

The fluvial acoustic tomography (FAT) system relies on the arrival time of the system signal to calculate the parameters of the region. The traditional method uses the matching filter method to calculate the peak position of the received acoustic signal after cross-correlation calculation within a certain time as the signal arrival time point, but this method is difficult to be effectively applied to the complex underwater environment, especially in the case of extremely low SNR. To solve this problem, a two-channel deep learning model (DCA-Net) is proposed to detect the arrival time of acoustic chromatographic signals. Firstly, an interactive module is designed to transmit the auxiliary information from the cross-correlation subnetwork to the original signal subnet to improve the feature information extraction capability of the network. In addition, an attention module is designed to enable the network to selectively focus on the important features of the received acoustic signals. Under the background of white Gaussian noise and real river environment noise, we use the received signals of the acoustic tomography system collected in the field to synthesize low SNR data of -10, -15, and -20 different decibels as datasets. The experimental results show that the proposed network model is superior to the traditional matching filtering method and some other deep neural networks in three low SNR datasets.