Multisource gravity anomaly data fusion algorithm based on feature preservation

Gravity-aided inertial navigation systems can realize the autonomy, long endurance, and high accuracy required for underwater vehicles. As prior information for navigation, the accuracy of gravity anomaly reference maps is essential for navigation performance. To construct a high-accuracy and rich-featured gravity anomaly map, it is necessary to effectively fuse multi-source gravity anomaly data. In current fusion algorithms, the accuracy of the least square collocation method and the wavelet transform method is limited by the difficulty of accurately selecting algorithm parameters. To address this issue, a deep learning network based on federal feature preservation is proposed in this paper for constructing gravity anomaly maps using a unified fusion strategy. The intensity and gradient features are incorporated into the network training process, driving the fusion results to preserve the parts of the source data with richer features. By measuring the richness of information, an adaptive loss function is constructed and completes the training of the network parameters. Finally, the experiments are conducted using an actual shipborne measurement dataset from the Gulf of Mexico, and high fusion accuracy is demonstrated by the proposed method. In addition, the capabilities of different fusion algorithms in improving gravity-aided navigation positioning accuracy are compared.