An Encoder-Guided Neural Network Approach for Suppressing Wheel-Induced Magnetic Interference in Unmanned Ground Vehicle Magnetometers

Geomagnetic surveys conducted by unmanned ground vehicles (UGVs) offer significant advantages in various applications, but their data quality is severely degraded by wheel-induced magnetic interference. This interference, whose fundamental frequency is synchronous with the vehicle's speed, becomes deeply intertwined with the low-frequency geomagnetic signals of interest, posing a significant challenge for traditional denoising and compensation methods like the Tolles-Lawson (T-L) model. To address this challenge, we propose a guidance-based extraction method, postulating that a signal representing the instantaneous state of the wheel can be used to explicitly guide the noise cancellation process. We materialize this idea by introducing the refinement U-Mamba network, a tailored deep learning architecture that leverages a dual-channel input to process magnetometer and wheel encoder data in parallel. The network is built upon a synergistic integration of a U-Net structure and the Mamba state-space model. This architecture is further enhanced by our proposed guidance-gated refinement Mamba-Transformer (GGRT) and gated attention delta (GAD) blocks, which enable precise and guidance-informed noise suppression. To validate its practical utility, we evaluated our method by integrating it as a preprocessing step for the standard T-L compensation pipeline in real-world operational tests. The model demonstrates sufficient real-time capabilities, achieving processing rates for both online ground station and onboard embedded-platform deployment. The results show that our approach achieves a performance gain of approximately 41% over the next best baseline model. This significant improvement underscores the superiority of the guidance-based method over conventional unguided denoising techniques for this specific and critical real-world application.