Unsupervised/Semi-Supervised Magnetic Anomaly Detection Method Based on Deep Support Vector Data Description

Existing magnetic anomaly detection (MAD) methods are widely categorized into target-, noise-, and machine learning-based methods. This article first analyzes the commonalities and characteristics of these methods, unifying them into noise- and target-based frameworks. Focusing on the MAD problem under static sensing systems, and considering that the noise-based methods have better stability in real-world detection but suffer from poor performance at low signal-to-noise ratios (SNRs), this article proposes a novel MAD method based on deep support vector data description (Deep SVDD). The proposed method characterizes long-term magnetic background noise patterns in the region of interest. A deep neural network encoder is employed to extract time–frequency features of the signals, yielding a compact low-dimensional latent representation. The latent space is constrained by a prior distribution derived from a pretrained model, and the probability of noise signal features is maximized in the form of maximum likelihood estimation. To effectively avoid overfitting caused by hypersphere collapse, the Kullback–Leibler (KL) divergence is incorporated into the loss function. Statistical tests and visualizations confirm the method’s effectiveness and alignment with theoretical foundations, while comparative experiments demonstrate that the proposed method achieves significant performance improvements, especially at low SNRs over existing noise-based methods. Furthermore, to prevent the performance collapse in simulation-to-reality transfer that occurs in deep learning (DL) methods driven by semi-realistic data due to inaccurate prior information, this article also proposes a novel semi-supervised learning MAD method driven by sparse prior information about the magnetic anomaly. Experiments demonstrate that the proposed method exhibits superior stability, particularly showing enhanced robustness against $1/f^{\alpha }$ noise compared to supervised learning methods. Moreover, the controlled prior information integration mode enables the proposed method to achieve effective tradeoffs between sensitivity and stability in practical deployments. This confirms the method’s reduced dependence on simulation-derived prior information, making it particularly suitable for complex real-world detection scenarios. Finally, the method’s practical performance has been validated using both terrestrial and marine field data.