Improving Stability and Generalization of Magnetic Anomaly Detection Using Deep Convolutional Siamese Neural Networks

A Siamese neural network architecture is introduced to enhance the stability and generalization of deep neural networks for magnetic anomaly detection (MAD). Grounded in signal disparity contrastive learning, this study addresses the statistical disparity in signals from various regions and times. Within the proposed architecture, two identical 1-D convolutional neural networks with shared parameters are used as feature extractors for obtaining the embedding of paired input signals in the target space. Decision networks are then formulated to measure the discrepancies between these embeddings, shedding light on the differences between the original signals. A base signal family is crafted for detection using multiple noisy signals that are spatially and temporally aligned with the evaluated signal. The difference between the measured signal and those in the base family is computed. A voting mechanism subsequently determines if the assessed signal is a magnetic anomaly. Numerous semi-realistic datasets are employed for network training. The results indicate that the proposed network surpasses several existing networks in robustness with regard to detection area, time, and signal parameter variations and also has excellent detection capability and temperature in the face of measured magnetic anomaly signals. Notably, with changes in test parameters, the network only requires the background noise signal as the base, maintaining high detection performance without retraining.