Magnetic Anomaly Detection Based on Attention-Bi-LSTM Network

The performance of conventional methods for detecting magnetic anomalies has been limited by a low signal-to-noise ratio (SNR) and the presence of complex noise environments, particularly in the context of Gaussian-colored noise. In addition, the data obtained from fluxgate magnetometers is in the form of time series, with traditional approaches often neglecting time-domain features in favor of frequency-domain features. Despite significant advancements in time series models in recent years, their potential application to magnetic anomaly detection has been largely disregarded. In response to these challenges, this paper introduces a magnetic anomaly detection approach based on the Attention-Bi-directional Long Short-Term Memory (ATT-Bi) network. To address these issues, the proposed method employs a preprocessing method involving wavelet decomposition and filtering to extract low-frequency features, thereby enhancing time-domain features and concurrently improving the SNR. Thereafter, an ATT-Bi network is utilized to extract time-domain features and capture the data correlation between time sequences for the detection of magnetic anomaly signals. The performance of the ATT-Bi network is evaluated through simulation and field testing, with comparisons made against other methods. The simulation results demonstrate that, in low SNR and Gaussian-colored noise environments, the ATT-Bi network achieves the highest detection accuracy. Moreover, the field test results consistently exhibit a detection accuracy of over 95% for ATT-Bi, highlighting the superior performance of this detection method and confirming the viability of prioritizing time-domain features.