Intelligent aeromagnetic compensation approach for enhancing adaptability and generalization.

Area of Science:

• Geophysics
• Aeromagnetic Surveying
• Data Quality Control

Background:

• Aeromagnetic survey data quality relies on compensating for aircraft magnetic interference.
• Existing neural network methods suffer from high input dimensionality, reliance on experience, and lack of temporal memory, leading to instability and poor generalization.
• Time-varying interference features and dataset shifts degrade prediction performance.

Purpose of the Study:

• To propose a novel compensation model architecture for adaptive feature extraction and dimensionality reduction in aeromagnetic surveys.
• To enhance self-adaptability and improve the generalization of magnetic interference prediction.
• To mitigate performance degradation caused by dataset shifts and geomagnetic noise.

Main Methods:

• Developed a novel compensation model architecture using autoencoder bidirectional long short-term memory (AE-BiLSTM).
• Implemented adaptive feature extraction and parameter dimensionality reduction.
• Incorporated temporal dependency learning and defined two specialized loss functions to address dataset shift and geomagnetic noise.

Main Results:

• The proposed AE-BiLSTM method achieved an improvement ratio of 28.36 for magnetic interference mitigation on verification flights.
• Demonstrated significantly higher performance compared to existing methods and commercial compensators.
• Showcased excellent generalization, with only a 2% decrease in improvement ratio from calibration to verification flights.

Conclusions:

• The novel AE-BiLSTM model effectively addresses limitations of existing methods for aeromagnetic interference compensation.
• The proposed approach enhances self-adaptability, prediction accuracy, and generalization capabilities.
• This method offers a significant advancement in improving the data quality of aeromagnetic surveys.