Lightweight Deep Learning for Identification and Suppression of Subway Interference in Geomagnetic Observations
Abstract
Geomagnetic observatory networks provide continuous data for space weather monitoring, field modeling, and solid Earth geophysics. Rapid urbanization, however, has introduced pervasive electromagnetic interference from subway traction systems, degrading data quality at a growing number of urban stations. Existing denoising methods are computationally expensive, generalize poorly, and do not reliably separate subway interference from natural geomagnetic variations that overlap the interference band. We propose a lightweight, identification‐driven framework in which TinyCNN detects interference with 102K parameters and ResidualCNN suppresses it with 413K parameters. The detector was trained on data from 22 stations across China's geomagnetic network (2014–2025), while the denoising model was trained on the two persistently subway‐affected stations using VMD‐guided pseudo‐labels. TinyCNN achieves an F 1‐score of 0.9997 with inference time of about 1.5 ms per hourly segment on a single central processing unit core. ResidualCNN achieves >99.99% energy suppression in the 0.004–0.032 Hz interference band identified via adaptive Otsu thresholding, while preserving the Sq diurnal variation with median amplitude errors of about 2% and phase errors of about 2 min across all stations, with long‐window low‐frequency spectral fidelity above 0.998. Cross‐station validation on 589 completely withheld Taiyuan station‐days shows stable suppression and low‐frequency fidelity across low‐, moderate‐, and strong‐interference conditions, with overall performance comparable to the weak‐to‐moderate interference training station. Across 1,384 station‐days, the denoised first‐order differences remain below 0.1 nT, consistent with observatory noise‐level requirements. The framework can be applied to new stations in similar urban environments without station‐specific retraining, supporting automated quality control across observatory networks.