Noise Suppression Method for Magnetic Resonance Sounding Data Based on Causal Modular Parallel Interactive Attention Network

Magnetic resonance sounding (MRS) requires detecting weak nuclear spin responses under low signal-to-noise conditions, which are often susceptible to complex noise including non-stationary noise and interference near the Larmor frequency. Due to the time-varying and complex nature of such noise, more adaptive and robust high-precision signal extraction strategies are urgently needed. To overcome the limitations of traditional approaches which require multiple algorithms for different noise types and rely heavily on field data in deep learning methods, we propose the Causal Modular Parallel Interactive Attention Network (CM-PIAN). Trained solely on simulated data, the framework demonstrates strong generalization ability through a causal intra-modular parallel architecture, enabling direct processing of field data without additional adjustments. CM-PIAN employs a time-sequential causal prediction strategy based on a sliding-window approach, achieving a balance between overall reconstruction accuracy and local detail preservation. This design improves data efficiency and training stability when training data are limited, leading to better generalization to field measurements. The network employs a modular architecture with parallel multi-scale feature extraction, offering flexible adaptability to various noise conditions. Notably, CM-PIAN maintains high parameter estimation accuracy under extremely noisy conditions where conventional methods fail, and achieves superior denoising performance on field data. The proposed method is expected to alleviate denoising challenges caused by limited measured data in groundwater exploration, mineral surveys, and other geophysical applications.