Underwater acoustic target recognition based on ship-radiated noise is difficult owing to the complex marine environment and the interference by multiple targets. As an important technology for target recognition, deep-learning has high accuracy but poor interpretability. In this study, an attention-based neural network (ABNN) is proposed for target recognition in the pressure spectrogram with multi-source interference using an attention module to inspect the inner workings of the neural network. From data obtained during a September 2020 sea trial, the ABNN exhibited a gradual focus on the frequency-domain feature of the target ship and suppressed environmental noises and marine vessel interference, which led to high accuracy in the target detection and recognition.

The you-only-look-once (YOLO) model identifies objects in complex images by framing detection as a regression problem with spatially separated boundaries and class probabilities. Object detection from complex images is somewhat similar to underwater source detection from acoustic data, e.g., time-frequency distributions. Herein, YOLO is modified for joint source detection and azimuth estimation in a multi-interfering underwater acoustic environment. The modified you-only-look-once (M-YOLO) input is a frequency-beam domain (FBD) sample containing the target and multi-interfering spectra at different azimuths, generated from the received data of a towed horizontal line array. M-YOLO processes the whole FBD sample using a single-regression neural network and directly outputs the target-existence probability and spectrum azimuth. Model performance is assessed on both simulated and at-sea data. Simulation results reveal the strong robustness of M-YOLO toward different signal-to-noise ratios and mismatched ocean environments. As tested on the data collected in an actual multi-interfering environment, M-YOLO achieved near-100% target detection and a root mean square error of 0.54° in azimuth estimation.

A modified convolutional neural network (CNN) is proposed to enhance the reliability of source ranging based on acoustic field data received by a vertical array. Compared to the traditional method, the output layer is modified by outputting Gauss regression sequences, expressed using a Gaussian probability distribution form centered on the actual distance. The processed results of deep-sea experimental data confirmed that the ranging performance of the CNN with a Gauss regression output was better than that using single regression and classification outputs. The mean relative error between the predicted distance and the actual value was ~2.77%, and the positioning accuracy with 10% and 5% error was 99.56% and 90.14%, respectively.