Magnetic Dipole Target Signal Detection via Convolutional Neural Network

Abstract

In this letter, two convolutional neural network (CNN) models for detection of magnetic anomaly target signals are proposed. One is a 1-D CNN combined with signal feature (SF-CNN1D), and the other is a 2-D CNN based on time–frequency diagrams (TF-CNN2D). To train the models, simulated signals are added to the measured background noise to generate the positive sample set, while the negative sample set is the pure measured noise. Simulation results show that both of the models have satisfying train and test accuracies. A field experiment is conducted to verify the generalization ability of the two CNN models to real data. It is demonstrated that the two CNN models have good detection performances and are some better than the conventional support vector machine (SVM) approach by several percent in colored Gaussian noise scenarios.