Aeromagnetic Compensation Based on Deep Transfer Learning Combined With Physical Constraints

With the rapid development of deep learning technology, it has also been widely applied in the field of aeromagnetic compensation. However, when obtaining data on aviation platforms, data quality is unstable, and geomagnetic gradient interference exists. Currently, data-driven neural networks lack the integration of physical information constraints for practical scenarios, and their hardware computing power requirements are constantly increasing. This has led to a decrease in the accuracy of compensation results and the adaptability of methods, limiting their application scenarios. In response to such issues, this article proposes an aeromagnetic compensation method based on deep transfer learning combined with physical information constraints. The method first preprocesses the data to remove invalid data and geomagnetic gradient interference. Afterward, a neural network physics-informed neural network (PINN) that integrates physical information is used to compensate and train the magnetic measurement data, thereby improving the accuracy of the training results, and then use knowledge distillation to compress the trained neural network, in order to reduce the complexity of neural network computation and save limited computing resources on the platform. Finally, the compressed neural network-accelerated PINN (AC-PINN) will be migrated to run on a Raspberry Pi suitable for uncrewed aerial vehicles (UAVs). The experimental results show that the proposed model algorithm can achieve reliable magnetic data compensation calculation on UAV aeromagnetic measurement systems with limited computing power from four aspects: floating-point operations (FLOPs), standard deviation ( ${\mathrm {STD}}_{\mathrm {c}}$ ), improvement ratio (IR), and running time (TIME).