In this letter, a deep-learning-based neural network for magnetic dipole modeling (MDMnet) is introduced in the framework of dc magnetic cleanliness for space missions. The developed method targets modeling the static magnetic signature of a spacecraft unit that is obtained during the unit-level characterization stage of the extensive prelaunch electromagnetic compatibility test campaign. By employing synthetic magnetic flux density data generated by virtual dipole sources, the MDMnet can be trained to accurately estimate the magnetic parameters of real equipment based on its near magnetic flux density measurements. The target of the deep learning algorithm is, on the one hand, to effectively minimize the prediction errors (loss function) throughout the training process and, on the other hand, to enable the generalization of the model predictions, i.e., exhibit accurate model estimations with unseen magnetic induction data. Extensive simulations toward the stabilization of the MDMnet hyperparameters are outlined, and indicative model inferences employing artificial magnetic flux density data are carried out. Finally, the MDMnet can achieve a predictive accuracy of 0.8 mm with respect to the dipole localization and 1% with respect to the magnetic induction magnitude, verifying the potency of the developed method.

中文翻译：

---

使用多个磁偶极子对航天器设备的磁特征进行建模的深度学习方法


在这封信中，在太空任务的直流磁洁净度框架中引入了一种基于深度学习的磁偶极子建模神经网络（MDMnet）。所开发的方法的目标是对航天器单元的静磁特征进行建模，该静磁特征是在广泛的发射前电磁兼容性测试活动的单元级表征阶段获得的。通过采用虚拟偶极源生成的合成磁通密度数据，可以训练 MDMnet 根据其近磁通密度测量值准确估计真实设备的磁参数。深度学习算法的目标一方面是有效地最小化整个训练过程中的预测误差（损失函数），另一方面是使模型预测能够泛化，即表现出准确的模型估计具有看不见的磁感应数据。概述了针对 MDMnet 超参数稳定性的广泛模拟，并利用人工磁通密度数据进行了指示性模型推论。最后，MDMnet 在偶极子定位方面可以达到 0.8 mm 的预测精度，在磁感应强度方面可以达到 1% 的预测精度，验证了所开发方法的有效性。