The data quality of an aeromagnetic survey is determined by the compensation for the magnetic interference of the aircraft. In existing neural-network-based magnetic compensation, the dimensionality of the input parameters of the magnetic interference network is high, and the determination of the input depends on personal experience and preliminary experiments, increasing the probability of overfitting and compensation instability. In addition, the interference features vary with time, while the existing models do not have temporal memory, resulting in poor generalization of the interference prediction. In addressing this issue, a novel compensation model architecture based on autoencoder bidirectional long short-term memory is proposed to adaptively extract features and reduce the parameter dimensionality, and thus enhance self-adaptability. In addition, the temporal dependencies among historical magnetic data along positive and negative directions are learned to effectively reduce the prediction error. Furthermore, two loss functions are defined to mitigate the degradation of prediction performance due to the dataset shift between calibration and measurement flights and to reduce the sensitivity of the interference prediction to geomagnetic noise. To verify the proposed method, we developed a dedicated compensator and built a flight test platform. The results show that the improvement ratio of magnetic interference on the verification flight for our method reaches 28.36, which is significantly higher than the ratios for the existing methods and the state-of-the-art commercial compensator, resulting in better interference mitigation. From the calibration flight to the verification flight, the improvement ratio of our method decreases by only 2%, resulting in a satisfactory generalization.
Wang et al. (Wed,) studied this question.