Time–Frequency Cross-Information Fusion Neural Network for Ship Shaft-Rate Magnetic Field Detection

The evident line-spectrum characteristics of the shaft-rate magnetic field generated by moving ships in the frequency domain offer a solution for underwater target detection. Nevertheless, in shallow water, intricate magnetic noise significantly diminishes the signal-to-noise ratio (SNR) of the shaft-rate magnetic field. The lack of measured data makes the use of deep-learning approaches for detection difficult. This article proposes a deep learning-based method for detecting shaft-rate magnetic field signals from colored noise samples with a very low SNR. Initially, we employ a time-harmonic magnetic dipole to simulate shaft-rate magnetic field signals. Subsequently, we implement a Gaussian-weighted approach to incorporate simulated shallow-sea magnetic background noise. This study concurrently establishes a reliable shaft-rate magnetic field dataset for the deep-learning system by integrating it with a subset of recorded geomagnetic field signals from the same location. We propose a time–frequency information cross-fusion shaft-rate magnetic field detection network (TFSRM-Net) to analyze the transmission properties of the shaft-rate magnetic field. This enhances the resilience and transferability of shaft-rate magnetic field signal detection at low SNR. TFSRM-Net achieves 97.92% accuracy on the measured signals in a field experiment, illustrating its effectiveness in both simulated and empirical data.