This paper proposes a joint time–frequency analysis method that combines Rao detector with dynamic sliding thresholds to enhance the detection performance of electric source axial frequency magnetic field signals. For each signal-to-noise ratio (SNR) point, 1000 Monte Carlo simulations were independently conducted, with SNR ranging from 15 dB to −30 dB. The results show that the proposed method maintains high detection rates even at extremely low SNRs, achieving about 90% detection probability at −13 dB, significantly outperforming traditional energy detectors (with a threshold of 2 dB). Under conditions where the detection probability is ≥90% and the false alarm probability is 10⁻³, the SNR threshold for the Rao detector is reduced by 15 dB compared to energy detectors, greatly improving detection performance. Even at lower SNRs (−30 dB), the Rao detector still maintains a certain detection rate, while the detection rate of energy detectors rapidly drops to zero. Further analysis of the impact of different frequencies (1–5 Hz) and CPA distances (45–80 cm) on performance verifies the algorithm’s robustness and practicality in complex non-Gaussian noise environments. This method provides an effective technical solution for low SNR detection of ship axial frequency magnetic fields and has good potential for practical application.

This manuscript introduces MAGInet, a novel deep learning framework designed for the magnetic multi-parameter inversion of complex structures. The architecture of MAGInet integrates a classifier and several solvers, where the classifier performs a preliminary categorization of magnetic field signals and the solvers execute a detailed regression analysis to predict multiple parameters of the structures. The operational sequence of MAGInet is as follows: magnetic field signals are initially fed into the classifier for classification, and the outcomes are subsequently channeled into the appropriate solver for multi-parameter forecasting. The underlying principle of MAGInet’s multi-parameter inversion is to establish a learned mapping between the magnetic field signals and the multi-magnetic parameters of complex structures, leveraging extensive training datasets. In this study, simulation experiments are conducted, with training datasets generated via finite element magnetic field modeling, which are then utilized to fine-tune the MAGInet model. The results of these experiments demonstrate that the accuracy of multi-parameter prediction for complex structures can reach up to 97.5% under zero-error conditions. Furthermore, the efficacy of the MAGInet inversion approach is contrasted with the traditional, deep, fully connected neural network methodologies. Comparative analyses reveal that MAGInet significantly outperforms traditional deep neural networks in terms of accuracy for predicting the magnetic multi-parameters of complex structures, showcasing superior performance.