Micro-Motion Parameter Estimation of Underwater Acoustic Target Using Convolutional Neural Networks

As a unique attribute of objects, micro-motion characteristics have been widely utilized in target feature extraction and recognition. The estimation of micro-motion parameters in complex and dynamic underwater environments possesses significant research value. To address the inherent limitations of traditional linear time-frequency analysis methods constrained by the Heisenberg uncertainty principle in short-pulse echo processing, we propose a micro-Doppler feature extraction framework based on a convolutional neural network (CNN), achieving super-resolution parameter estimation through an end-to-end approach. By developing a skip-connection-based feature learning algorithm, the network model takes raw echo signals as input and employs large-scale convolutional kernel groups to hierarchically extract micro-motion features from time-domain waveforms. Numerical simulations demonstrate that the proposed CNN model achieves a mean absolute percentage error (MAPE) of 0.31 % in micro-motion frequency inversion at a signal-to-noise ratio (SNR) of 10 dB. Compared with conventional linear time-frequency analysis methods, the CNN model maintains prediction accuracy even when the micro-Doppler shift is smaller than the frequency resolution, with an error of 2.55% at a micro-motion amplitude of 0.01 m. These results validate the method's effectiveness in detecting weak underwater acoustic targets and offer a novel approach for dynamic feature recognition of cooperative underwater targets.